Upload folder using huggingface_hub

icml26/80c20b7b-ead6-454a-849e-56702a6c828f/appendix_chunks.jsonl

@@ -0,0 +1,57 @@
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0083", "section": "A. Data Sources", "page_start": 14, "page_end": 14, "type": "Text", "text": "Our work is compiled based on following public medical image repositories:", "source": "marker_v2", "marker_block_id": "/page/13/Text/2"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0084", "section": "A. Data Sources", "page_start": 14, "page_end": 14, "type": "TableGroup", "text": "Table 4. Summary of public medical datasets utilized in the construction of MieDB-100k. The columns #Train and #Benchmark denote the number of samples allocated to our training and benchmark splits respectively from each source dataset. DatasetName #Train #Benchmark Modality AbdomenUS (Vitale et al., 2020) 569 62 Ultrasound Bbbc010 (Ljosa et al., 2012) 70 20 Microscopy Bkai-Igh (Ngoc Lan et al., 2021) 700 81 Endoscopy Brats-gli (de Verdier et al., 2024) 1529 80 MRI BriFiSeg (Mathieu et al., 2022) 1005 40 Microscopy BUSI (Al-Dhabyani et al., 2020) 452 80 Ultrasound CellNuclei (Caicedo et al., 2019) 469 51 Microscopy ChaseDB1 (Carballal et al., 2018) 19 7 Fundus Chest-ct-segmentation (Polo, 2025) 278 19 CT Chest-xray-masks-and-labels (Pandey, 2025) 666 32 Xray CHUAC (Angiographics) 17 5 Fundus COVID-19 Radiography Dataset (Chowdhury et al., 2020) 2010 95 Xray COVID-19-CT-SCAN-Lesion (Morozov et al., 2020) 255 15 CT CovidQU (Tahir et al., 2021) 5684 122 Xray CT MAR (Haneda et al., 2025) 1595 82 CT CT-Low-Dose-Reconstruction (AAPM, 2016) 867 51 CT CystoidFluid (Ahmed et al., 2022) 703 59 OCT Dca1 (Cervantes-Sanchez et al., 2019) 93 28 Fundus Deepbacs (Spahn et al., 2022) 17 10 Microscopy Drive (Staal et al., 2004) 18 20 Fundus DynamicNuclear (Van Valen et al., 2016) 50 17 Microscopy FHPsAOP (Lu et al., 2022) 2800 80 Ultrasound IDRiD (Porwal et al., 2018) 47 27 Fundus ISIC2016 (Gutman et al., 2016) 810 80 Dermoscopy ISIC2018 (Codella et al., 2019) 9973 115 Dermoscopy KMAR-50K (Wang & Shi, 2025) 651 47 MRI Kvasir (Jha et al., 2019) 4429 139 Endoscopy Lgg-mri-segmentation (Buda et al., 2019) 1669 55 MRI MoNuSAC (Verma et al., 2021) 0 21 Microscopy MR-ART (Narai et al. ´ , 2022) 820 18 MRI MSD (Antonelli et al., 2022) 797 3912 MRI NuSeT (Yang et al., 2020) 2383 40 Microscopy Paried MRI CT (Younus Akon, 2025) 1974 72 CT, MRI Pandental (Abdi et al., 2015) 81 24 Xray Pasta-GEN (Lei et al., 2025) 32299 731 CT PolypGen (Ali et al., 2024) 984 75 Endoscopy PROMISE12 (Litjens et al., 2014) 1031 80 MRI QaTa-COV19 (Aysen et al., 2024) 3573 85 Xray Refuge (Fang et al., 2022) 80 80 Fundus RoboTool (Garcia-Peraza-Herrera et al., 2021) 350 76 Surgical Photo ThyroidXL (Duong et al., 2025) 7029 138 Ultrasound Tnbcnuclei (Naylor et al., 2018) 35 10 Microscopy TotalSegmentator (Wasserthal et al., 2023) 5206 154 CT, MRI UltrasoundNerve (Montoya, 2026) 1651 50 Ultrasound USforKidney (Song et al., 2022) 4351 50 Ultrasound UWSkinCancer (Vision & Lab, 2024) 143 44 Dermoscopy VinDr-Multiphase (Dao et al., 2022) 3486 44 CT WBC (Zheng et al., 2018) 280 40 Microscopy YeaZ (Dietler et al., 2020) 358 51 Microscopy YGA low dose ct (pazhoulab, 2024) 4387 44 CT", "source": "marker_v2", "marker_block_id": "/page/13/TableGroup/545"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0085", "section": "A. Data Sources", "page_start": 14, "page_end": 14, "type": "Text", "text": "We also appreciate MedSegBench(Kus¸ & Aydin, 2024) and MedSegDB (Zhang et al., 2025) for collecting and pre-processing some of these datasets.", "source": "marker_v2", "marker_block_id": "/page/13/Text/5"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0086", "section": "B. Construction Details", "page_start": 15, "page_end": 15, "type": "FigureGroup", "text": "Figure 6. Construction details of three perspective. We manually curate the benchmark split to uphold high clinical standards. The remaining training data is validated through sampling-based quality checks, establishing a high-quality data proportion exceeding 95%.", "source": "marker_v2", "marker_block_id": "/page/14/FigureGroup/437"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0087", "section": "C. Implementation Details of OmniGen2-MIE", "page_start": 15, "page_end": 15, "type": "TableGroup", "text": "Hyper-Parameter Value Finetuning method Full-Parameter Finetuning snr type lognorm do shift True dynamic time shift True Steps 20, 000 #GPUs 8 Per-device batch size 8 Gradient accumulation 1 Global batch size (effective) 64 Learning rate 1 × 10−4 LR scheduler timm constant with warmup Warm-up t 500 Precision BF16 Random seed 2233 Table 5. Training hyper-parameters used for finetuning OmniGen2-MIE on our dataset.", "source": "marker_v2", "marker_block_id": "/page/14/TableGroup/438"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0088", "section": "D.1.1. MATHEMATICS", "page_start": 16, "page_end": 16, "type": "Text", "text": "To recover the segmentation mask from the visualized output, we model the edited image O as a linear interpolation between the original background image B (a.k.a. the input image I) and a known overlay color C (red, green or blue). This relationship is governed by the per-pixel alpha channel α ∈ [0, 1], according to the standard alpha blending equation:", "source": "marker_v2", "marker_block_id": "/page/15/Text/4"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0089", "section": "D.1.1. MATHEMATICS", "page_start": 16, "page_end": 16, "type": "Equation", "text": "\\mathbf{O} = (1 - \\alpha)\\mathbf{B} + \\alpha\\mathbf{C} \\tag{1}", "source": "marker_v2", "marker_block_id": "/page/15/Equation/5"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0090", "section": "D.1.1. MATHEMATICS", "page_start": 16, "page_end": 16, "type": "Text", "text": "By rearranging the terms as O − B = α(C − B), the scalar value α can be interpreted as the projection of the observed color shift onto the vector representing the maximum possible color change. To account for potential noise in the RGB space, we solve for α at each pixel using the least-squares solution:", "source": "marker_v2", "marker_block_id": "/page/15/Text/6"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0091", "section": "D.1.1. MATHEMATICS", "page_start": 16, "page_end": 16, "type": "Equation", "text": "\\alpha = \\frac{(\\mathbf{O} - \\mathbf{B}) \\cdot (\\mathbf{C} - \\mathbf{B})}{|\\mathbf{C} - \\mathbf{B}|^2} (2)", "source": "marker_v2", "marker_block_id": "/page/15/Equation/7"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0092", "section": "D.1.1. MATHEMATICS", "page_start": 16, "page_end": 16, "type": "Text", "text": "The continuous alpha map is subsequently binarized to produce the final segmentation mask M. This is achieved by applying a global threshold τ , such that:", "source": "marker_v2", "marker_block_id": "/page/15/Text/8"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0093", "section": "D.1.1. MATHEMATICS", "page_start": 16, "page_end": 16, "type": "Equation", "text": "M_{i,j} = \\begin{cases} 1 & \\text{if } \\alpha_{i,j} > \\tau \\\\ 0 & \\text{otherwise} \\end{cases} (3)", "source": "marker_v2", "marker_block_id": "/page/15/Equation/9"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0094", "section": "D.1.1. MATHEMATICS", "page_start": 16, "page_end": 16, "type": "Text", "text": "In our implementation, a threshold of τ = 0.5 is utilized to effectively separate the predicted regions from the background.", "source": "marker_v2", "marker_block_id": "/page/15/Text/10"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0095", "section": "D.1.2. CASE OF MASK RECONSTRUCTION", "page_start": 16, "page_end": 16, "type": "PictureGroup", "text": "Figure 7. Case of perception mask reconstruction.", "source": "marker_v2", "marker_block_id": "/page/15/PictureGroup/455"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0096", "section": "Scoring Rubric for Modification Tasks", "page_start": 16, "page_end": 16, "type": "Text", "text": "You are a helpful assistant in evaluating medical image editing result.", "source": "marker_v2", "marker_block_id": "/page/15/Text/17"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0097", "section": "Scoring Rubric for Modification Tasks", "page_start": 16, "page_end": 16, "type": "Text", "text": "You will be provided with an edit instruction and a collage image where the leftmost is origin image, center is edited image and rightmost is the reference ground truth image.", "source": "marker_v2", "marker_block_id": "/page/15/Text/18"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0098", "section": "Scoring Rubric for Modification Tasks", "page_start": 16, "page_end": 16, "type": "Text", "text": "You should score how well an edited image matches the intended edit while preserving clinical realism and image integrity based on following scoring rubrics:", "source": "marker_v2", "marker_block_id": "/page/15/Text/19"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0099", "section": "1) Edit Goal Fulfillment (Edit Correctness): Assesses whether the intended lesion change is achieved.", "page_start": 16, "page_end": 16, "type": "Text", "text": "Scoring reference:", "source": "marker_v2", "marker_block_id": "/page/15/Text/21"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0100", "section": "1) Edit Goal Fulfillment (Edit Correctness): Assesses whether the intended lesion change is achieved.", "page_start": 16, "page_end": 16, "type": "ListGroup", "text": "5: Lesion added/removed exactly as intended; no residuals or unintended remnants. 4: Mostly correct; slight residual signal after removal or slight under/over-addition. 3: Partial success; lesion still partially present (removal) or incomplete/incorrect lesion (addition).", "source": "marker_v2", "marker_block_id": "/page/15/ListGroup/456"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0101", "section": "1) Edit Goal Fulfillment (Edit Correctness): Assesses whether the intended lesion change is achieved.", "page_start": 17, "page_end": 17, "type": "ListGroup", "text": "2: Wrong area or wrong type of change; target lesion largely unchanged. 1: No effective edit or opposite edit performed. 2) Edit Area Morphology (Shape, Margins, Internal Structure): Evaluates whether edit area matches expected morphology and/or reference.", "source": "marker_v2", "marker_block_id": "/page/16/ListGroup/406"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0102", "section": "Scoring reference", "page_start": 17, "page_end": 17, "type": "ListGroup", "text": "5: Shape, border characteristics, and internal texture are highly consistent. 4: Minor border/shape irregularities; still plausible. 3: Morphology is generic/unconvincing; borders/texture inconsistent. 2: Clearly artificial morphology (blocky, repeated patterns, unnatural contours). 1: Morphology nonsensical or misleading (e.g., appears like different pathology). 3) Intensity / Signal / Attenuation Consistency: Checks whether edited region match modality-specific intensities.", "source": "marker_v2", "marker_block_id": "/page/16/ListGroup/407"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0103", "section": "Scoring reference", "page_start": 17, "page_end": 17, "type": "ListGroup", "text": "5: Intensities match local tissue statistics; no intensity discontinuities. 4: Slight intensity mismatch detectable with careful viewing. 3: Obvious mismatch (too bright/dark), inconsistent with modality or anatomy. 2: Strong intensity discontinuity; clearly edited. 1: Severe intensity errors that invalidate the image (e.g., saturation/clipping, inverted contrast). 4) Boundary Blending & Transition Naturalness: Rates edge blending and transitions between edited and unedited regions.", "source": "marker_v2", "marker_block_id": "/page/16/ListGroup/408"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0104", "section": "Scoring reference", "page_start": 17, "page_end": 17, "type": "ListGroup", "text": "5: Seamless blending; no halos, ringing, cut-paste edges. 4: Minor halo/edge artifacts only on close inspection. 3: Visible seams; boundary looks edited. 2: Strong cutout appearance or blur patches. 1: Boundary artifacts dominate the image. 5) Background / Non-target Preservation: Measures unintended changes outside the lesion edit region. Scoring reference: 5: Non-target anatomy and background unchanged (within expected noise). 4: Small unintended changes but not clinically meaningful. 3: Noticeable unintended alterations in nearby structures. 2: Large unintended modifications to anatomy or overall image. 1: Global corruption or major anatomical distortions. 6) Anatomical Plausibility & Clinical Coherence: Assesses whether result respects anatomy and pathology logic (e.g., lesion doesn't cross impossible boundaries). Scoring reference: 5: Fully plausible; consistent with organ boundaries and expected presentation. 4: Mostly plausible; minor oddity but acceptable. 3: Questionable plausibility (e.g., lesion overlaps structures unnaturally). 2: Clearly implausible anatomy/pathology relationship. 1: Clinically nonsensical or misleading. 7) Artifact Introduction (Noise, Texture, Aliasing, Compression, Repetition): Evaluates new artifacts introduced by editing. Scoring reference:", "source": "marker_v2", "marker_block_id": "/page/16/ListGroup/409"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0105", "section": "Scoring reference", "page_start": 18, "page_end": 18, "type": "ListGroup", "text": "5: No new artifacts; noise texture consistent with original. 4: Minor artifacts (subtle smoothing/grain mismatch). 3: Artifacts visible and distracting. 2: Strong artifacts (banding, checkerboard, repeated texture). 1: Severe artifacts preventing clinical use. 8) Image Quality & Acquisition Consistency: Checks consistency with scanner characteristics (resolution, blur, point spread, slice thickness cues, motion). Scoring reference: 5: Matches acquisition characteristics; sharpness/noise consistent. 4: Slight mismatch in sharpness or noise level. 3: Clear mismatch (over-smoothed or over-sharpened region). 2: Strong mismatch; edited region appears from different source. 1: Completely inconsistent with acquisition; unusable.", "source": "marker_v2", "marker_block_id": "/page/17/ListGroup/336"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0106", "section": "Return Format: Return a JSON dictionary with two fields", "page_start": 18, "page_end": 18, "type": "ListGroup", "text": "\"conclusion\": A brief conclusion to the edited image. \"score list\": The scores of the eight aspects in a JSON list.", "source": "marker_v2", "marker_block_id": "/page/17/ListGroup/337"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0107", "section": "Input collage VLM Output", "page_start": 18, "page_end": 18, "type": "Text", "text": "conclusion : The edited image successfully removes the renal cyst and replaces it with plausible parenchyma closely matching the reference, with only mild residual smoothing/texture mismatch in the filled region and minimal impact on surrounding anatomy.", "source": "marker_v2", "marker_block_id": "/page/17/Text/28"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0108", "section": "Input collage VLM Output", "page_start": 18, "page_end": 18, "type": "Text", "text": "score_list : [5, 4, 4, 4, 5, 5, 4, 4]", "source": "marker_v2", "marker_block_id": "/page/17/Text/29"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0109", "section": "Input collage VLM Output", "page_start": 18, "page_end": 18, "type": "Text", "text": "conclusion : The edited image does not introduce a clinically plausible gallstone. Instead, it contains multiple obvious synthetic circular overlays and line artifacts across the abdomen, with major non-target corruption and mismatch to CT texture; it does not match the reference which shows no such artifacts.", "source": "marker_v2", "marker_block_id": "/page/17/Text/33"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0110", "section": "Input collage VLM Output", "page_start": 18, "page_end": 18, "type": "Text", "text": "score_list : [1, 1, 1, 1, 1, 1, 1, 1]", "source": "marker_v2", "marker_block_id": "/page/17/Text/34"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0111", "section": "Input collage VLM Output", "page_start": 18, "page_end": 18, "type": "Caption", "text": "Figure 8. Cases of VLM rubric scoring.", "source": "marker_v2", "marker_block_id": "/page/17/Caption/35"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0112", "section": "E.1. Modality-Wise Performance Analysis", "page_start": 19, "page_end": 19, "type": "FigureGroup", "text": "Figure 9. Modality-wise performance analysis result within perception perspective. Left: DICE score; right: PSNR score.", "source": "marker_v2", "marker_block_id": "/page/18/FigureGroup/183"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0113", "section": "E.1. Modality-Wise Performance Analysis", "page_start": 19, "page_end": 19, "type": "Text", "text": "To investigate the impact of modality deviation, we conduct a modality-wise analysis of the benchmarking results within the Perception perspective. Specifically, we report the DICE and PSNR scores of six representative models across all medical imaging modalities included in MieDB-100k. As illustrated in Fig. 9, the experimental results are consistent with our hypotheses. For the baseline models, performance is unevenly distributed across the various modalities: They achieve relatively strong results on modalities that resemble natural images, such as Endoscopy, Dermoscopy, and Surgical Photo. However, on non-optical modalities (e.g., CT, MRI, Ultrasound), their performance degrades drastically. In contrast, the model trained on our dataset exhibits balanced and superior performance across all imaging types. Collectively, these results demonstrate that a diverse dataset like MieDB-100k is essential for successfully adapting multi-modal generative models to the medical domain.", "source": "marker_v2", "marker_block_id": "/page/18/Text/5"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0114", "section": "E.2. Multi-Round Generation", "page_start": 19, "page_end": 19, "type": "TableGroup", "text": "Table 6. Multi-round generation result. Best values are marked in Bold Perception Trasnformation DICE P-A P-ACC B-PSNR B-SSIM PSNR SSIM Pass@1 Pass@3 Pass@1 Pass@3 Pass@1 Pass@3 Pass@1 Pass@3 Pass@1 Pass@3 Pass@1 Pass@3 Open-Source SDXL-turbo 0.002 0.003 0.000 0.000 16.6 17.0 0.467 0.484 15.9 16.1 0.397 0.427 Bagel 0.263 0.383 0.069 0.137 13.9 16.1 0.620 0.703 12.7 15.2 0.442 0.548 OmniGen2 0.248 0.357 0.065 0.125 11.9 14.4 0.541 0.628 8.3 16.0 0.280 0.551 Step1X-Edit 0.332 0.369 0.126 0.143 15.5 16.4 0.727 0.748 16.6 17.1 0.539 0.558 FLUX.1-Kontext-dev 0.347 0.41 0.126 0.174 15.4 16.5 0.701 0.761 17.9 19.5 0.543 0.602 Qwen-Image-Edit 0.387 0.493 0.153 0.249 15.4 17.4 0.722 0.795 18.9 20.3 0.606 0.652 OmniGen2-MIE (Ours) 0.831 0.856 0.737 0.789 28.1 28.8 0.917 0.921 22.6 23.3 0.685 0.711", "source": "marker_v2", "marker_block_id": "/page/18/TableGroup/184"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0115", "section": "E.2. Multi-Round Generation", "page_start": 19, "page_end": 19, "type": "Text", "text": "To mitigate the inherent variance of the generative process, we report Pass@3 scores for the open-source models on Perception tasks. Specifically, we generate three independent outputs for each editing task and select the highest-performing sample to represent the task's score. These results are then averaged across all tasks to provide a robust assessment of overall performance.", "source": "marker_v2", "marker_block_id": "/page/18/Text/9"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0116", "section": "E.2. Multi-Round Generation", "page_start": 19, "page_end": 19, "type": "Text", "text": "The results of the multi-round generation tests are summarized in Table 6. While multi-round generation improves the absolute scores for baseline models, it does not alter the underlying fact that these models lack essential medical knowledge. Furthermore, the significant fluctuations across rounds expose the high-variance nature of these baselines, undermining their reliability under clinical applications. In contrast, our model exhibits remarkable stability across all three trials. This", "source": "marker_v2", "marker_block_id": "/page/18/Text/10"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0117", "section": "E.2. Multi-Round Generation", "page_start": 20, "page_end": 20, "type": "Text", "text": "consistency suggests that model trained on MieDB-100k has developed a deterministic understanding of medical concepts rather than relying on fortuitous generation.", "source": "marker_v2", "marker_block_id": "/page/19/Text/1"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0118", "section": "E.3. Out-Of-Distribution Image Edit", "page_start": 20, "page_end": 20, "type": "Text", "text": "While Section 4.5 demonstrates that the model trained on MieDB-100k generalizes effectively to OOD editing targets, we further evaluate its robustness by performing edits on 'in-the-wild' medical images sourced from the internet (Fig. 10) .", "source": "marker_v2", "marker_block_id": "/page/19/Text/3"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0119", "section": "E.3. Out-Of-Distribution Image Edit", "page_start": 20, "page_end": 20, "type": "FigureGroup", "text": "Figure 10. Examples of Out-Of-Distribution Editing.", "source": "marker_v2", "marker_block_id": "/page/19/FigureGroup/403"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0120", "section": "E.3. Out-Of-Distribution Image Edit", "page_start": 20, "page_end": 20, "type": "Text", "text": "The results indicate that our model is capable of readily adapting to medical images outside of datasets. This suggests that the diversity of MieDB-100k has successfully decoupled the model from specific data distribution, allowing it to internalize generalizable edit operations that are applicable to real-world clinical scenarios.", "source": "marker_v2", "marker_block_id": "/page/19/Text/6"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0121", "section": "F.1. Examples of Healthy Tissue Inpainting", "page_start": 20, "page_end": 20, "type": "FigureGroup", "text": "Figure 11. Examples of Inpainting. We train different inpainting models on each medical modalities. H: the Healthy image; L: the Lesion-bearing image.", "source": "marker_v2", "marker_block_id": "/page/19/FigureGroup/404"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0122", "section": "F.2. Extended Examples of Qualitative result", "page_start": 21, "page_end": 21, "type": "Text", "text": "Input Image Bagel FLUX.1 KontextOmniGen2OmniGen2-MIE Reference Nano Banana ProGPT Qwen Image Edit -Image [Edit Instruction] Illustrate all nuclei instances by painting GREEN masks directly onto the microscopy image, with zero modification to the background. Edit [Edit Instruction] Use a RED mask to highlight the polyp(s) by drawing it directly onto the given image, preserving the background exactly as is. Edit [Edit Instruction] Overlay unique RED masks on each organ in the ultrasound image by painting them directly, keeping the original background intact. Edit [Edit Instruction] Visually segment Caenorhabditis elegans by applying BLUE masks directly onto the microscope image without affecting the background. [Edit Instruction] Mark the region corresponding to BLADDER TUMOR in the Enhanced CT image using a solid RED mask applied directly, while ensuring the background stays unmodified. [Edit Instruction] Encode the lesion location with a GREEN mask drawn directly on the image, maintaining original background appearance. [Edit Instruction] Cover the infected zones with a RED mask by painting it directly onto the chest X-ray, ensuring background fidelity. Edit Edit Edit Edit", "source": "marker_v2", "marker_block_id": "/page/20/Text/1"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0123", "section": "F.2. Extended Examples of Qualitative result", "page_start": 22, "page_end": 22, "type": "PictureGroup", "text": "[Edit Instruction] Delete all visible manifestations of COVID pneumonia from the X-ray image and replace them with normal lung appearance, maintaining vascular and bronchial structures.", "source": "marker_v2", "marker_block_id": "/page/21/PictureGroup/571"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0124", "section": "F.2. Extended Examples of Qualitative result", "page_start": 22, "page_end": 22, "type": "PictureGroup", "text": "[Edit Instruction] Eliminate the Tumor in the Stomach while ensuring the reconstructed zone blends naturally with adjacent healthy tissue in terms of Hounsfield units and spatial patterns.", "source": "marker_v2", "marker_block_id": "/page/21/PictureGroup/572"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0125", "section": "F.2. Extended Examples of Qualitative result", "page_start": 22, "page_end": 22, "type": "PictureGroup", "text": "[Edit Instruction] Simulate a focal mucosal lesion in the white masked region that could represent a common colonic or gastric polyp, based on realistic endoscopic criteria.", "source": "marker_v2", "marker_block_id": "/page/21/PictureGroup/573"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0126", "section": "F.2. Extended Examples of Qualitative result", "page_start": 22, "page_end": 22, "type": "PictureGroup", "text": "[Edit Instruction] Reconstruct the affected skin area by removing the lesion and generating realistic, symmetryconsistent healthy tissue based on surrounding context.", "source": "marker_v2", "marker_block_id": "/page/21/PictureGroup/574"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0127", "section": "F.2. Extended Examples of Qualitative result", "page_start": 22, "page_end": 22, "type": "PictureGroup", "text": "[Edit Instruction] Convert the white masked area into a realistic representation of healthy skin, avoiding artificial smoothness or color mismatches.", "source": "marker_v2", "marker_block_id": "/page/21/PictureGroup/575"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0128", "section": "F.2. Extended Examples of Qualitative result", "page_start": 22, "page_end": 22, "type": "PictureGroup", "text": "[Edit Instruction] Synthetically insert a thyroid nodule into the given ultrasound scan, ensuring it exhibits clinically plausible features such as shape, margin, echogenicity, and vascularity.", "source": "marker_v2", "marker_block_id": "/page/21/PictureGroup/576"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0129", "section": "F.2. Extended Examples of Qualitative result", "page_start": 22, "page_end": 22, "type": "PictureGroup", "text": "[Edit Instruction] Add a convincing brain tumor to the FLAIR MRI while maintaining correct contrast dynamics—such as suppressed CSF and bright pathological signal.", "source": "marker_v2", "marker_block_id": "/page/21/PictureGroup/577"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0130", "section": "F.2. Extended Examples of Qualitative result", "page_start": 23, "page_end": 23, "type": "PictureGroup", "text": "[Edit Instruction] Improve diagnostic clarity of the CT by removing metal degradation while ensuring zero change to medically relevant content outside the artifact zone.", "source": "marker_v2", "marker_block_id": "/page/22/PictureGroup/531"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0131", "section": "F.2. Extended Examples of Qualitative result", "page_start": 23, "page_end": 23, "type": "PictureGroup", "text": "[Edit Instruction] Filter out motion artifacts from the PD MRI scan without smoothing, warping, or otherwise altering true tissue signals.", "source": "marker_v2", "marker_block_id": "/page/22/PictureGroup/532"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0132", "section": "F.2. Extended Examples of Qualitative result", "page_start": 23, "page_end": 23, "type": "PictureGroup", "text": "[Edit Instruction] Render the CT with Abdominal window parameters to mimic how it would appear on a PACS viewer configured for that tissue type.", "source": "marker_v2", "marker_block_id": "/page/22/PictureGroup/533"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0133", "section": "F.2. Extended Examples of Qualitative result", "page_start": 23, "page_end": 23, "type": "PictureGroup", "text": "[Edit Instruction] Synthesize a Arterial phase-equivalent CT image from the provided Non-Contrast phase input, ensuring realistic vascular and tissue enhancement patterns.", "source": "marker_v2", "marker_block_id": "/page/22/PictureGroup/534"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0134", "section": "F.2. Extended Examples of Qualitative result", "page_start": 23, "page_end": 23, "type": "PictureGroup", "text": "[Edit Instruction] Apply post-processing to remove metal artifacts from the CT image without smoothing, interpolating, or altering real anatomical details unnecessarily.", "source": "marker_v2", "marker_block_id": "/page/22/PictureGroup/535"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0135", "section": "F.2. Extended Examples of Qualitative result", "page_start": 23, "page_end": 23, "type": "PictureGroup", "text": "[Edit Instruction] Convert the given T2-MRI image to resemble a CT image, maintaining all underlying medical information intact.", "source": "marker_v2", "marker_block_id": "/page/22/PictureGroup/536"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0136", "section": "F.2. Extended Examples of Qualitative result", "page_start": 23, "page_end": 23, "type": "PictureGroup", "text": "[Edit Instruction] Process the low-dose CT using a SOFT kernel to suppress graininess while preserving low-contrast diagnostic details.", "source": "marker_v2", "marker_block_id": "/page/22/PictureGroup/537"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0137", "section": "G. Failure Cases", "page_start": 24, "page_end": 24, "type": "FigureGroup", "text": "Figure 12. Failure Cases.", "source": "marker_v2", "marker_block_id": "/page/23/FigureGroup/437"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0138", "section": "G. Failure Cases", "page_start": 24, "page_end": 24, "type": "Text", "text": "Fig 12 illustrates several representative failure cases of OmniGen2-MIE. The most frequent failure modes include: (1) semantic confusion between targeted anatomical features and morphologically similar background tissues; (2) intensity inconsistency, where the brightness of the edited region deviates from the surrounding context in a physically implausible manner; and (3) background inconsistency, especially after holistic transformations. These limitations underscore the need for more sophisticated multimodal architectures capable of preserving fine-grained details, as well as even more comprehensive training datasets to satisfy the requirements of rigorous clinical applications.", "source": "marker_v2", "marker_block_id": "/page/23/Text/4"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0139", "section": "H. Limitations", "page_start": 24, "page_end": 24, "type": "Text", "text": "Despite MieDB-100k provides a large-scale and diverse dataset for medical image editing, the primary limitation lies in the inherent difficulty of capturing ALL possible medical imaging modalities, and the relative scarcity of data for rare clinical cases. Continuous efforts to enrich these underrepresented categories will be vital for enhancing the dataset's diversity and effectivity. Furthermore, while our work establishes a foundation for unified understanding and generation, it focuses exclusively on editing tasks. Integrating medical VQA and text-to-image datasets represents a natural progression of this research direction, resulting in a more comprehensive resource for the development of holistic medical models.", "source": "marker_v2", "marker_block_id": "/page/23/Text/6"}

icml26/80c20b7b-ead6-454a-849e-56702a6c828f/appendix_text_v3.txt

@@ -0,0 +1,170 @@
+[p. 14 | section: A. Data Sources | type: Text]
+Our work is compiled based on following public medical image repositories:
+
+[p. 14 | section: A. Data Sources | type: TableGroup]
+Table 4. Summary of public medical datasets utilized in the construction of MieDB-100k. The columns #Train and #Benchmark denote the number of samples allocated to our training and benchmark splits respectively from each source dataset. DatasetName #Train #Benchmark Modality AbdomenUS (Vitale et al., 2020) 569 62 Ultrasound Bbbc010 (Ljosa et al., 2012) 70 20 Microscopy Bkai-Igh (Ngoc Lan et al., 2021) 700 81 Endoscopy Brats-gli (de Verdier et al., 2024) 1529 80 MRI BriFiSeg (Mathieu et al., 2022) 1005 40 Microscopy BUSI (Al-Dhabyani et al., 2020) 452 80 Ultrasound CellNuclei (Caicedo et al., 2019) 469 51 Microscopy ChaseDB1 (Carballal et al., 2018) 19 7 Fundus Chest-ct-segmentation (Polo, 2025) 278 19 CT Chest-xray-masks-and-labels (Pandey, 2025) 666 32 Xray CHUAC (Angiographics) 17 5 Fundus COVID-19 Radiography Dataset (Chowdhury et al., 2020) 2010 95 Xray COVID-19-CT-SCAN-Lesion (Morozov et al., 2020) 255 15 CT CovidQU (Tahir et al., 2021) 5684 122 Xray CT MAR (Haneda et al., 2025) 1595 82 CT CT-Low-Dose-Reconstruction (AAPM, 2016) 867 51 CT CystoidFluid (Ahmed et al., 2022) 703 59 OCT Dca1 (Cervantes-Sanchez et al., 2019) 93 28 Fundus Deepbacs (Spahn et al., 2022) 17 10 Microscopy Drive (Staal et al., 2004) 18 20 Fundus DynamicNuclear (Van Valen et al., 2016) 50 17 Microscopy FHPsAOP (Lu et al., 2022) 2800 80 Ultrasound IDRiD (Porwal et al., 2018) 47 27 Fundus ISIC2016 (Gutman et al., 2016) 810 80 Dermoscopy ISIC2018 (Codella et al., 2019) 9973 115 Dermoscopy KMAR-50K (Wang & Shi, 2025) 651 47 MRI Kvasir (Jha et al., 2019) 4429 139 Endoscopy Lgg-mri-segmentation (Buda et al., 2019) 1669 55 MRI MoNuSAC (Verma et al., 2021) 0 21 Microscopy MR-ART (Narai et al. ´ , 2022) 820 18 MRI MSD (Antonelli et al., 2022) 797 3912 MRI NuSeT (Yang et al., 2020) 2383 40 Microscopy Paried MRI CT (Younus Akon, 2025) 1974 72 CT, MRI Pandental (Abdi et al., 2015) 81 24 Xray Pasta-GEN (Lei et al., 2025) 32299 731 CT PolypGen (Ali et al., 2024) 984 75 Endoscopy PROMISE12 (Litjens et al., 2014) 1031 80 MRI QaTa-COV19 (Aysen et al., 2024) 3573 85 Xray Refuge (Fang et al., 2022) 80 80 Fundus RoboTool (Garcia-Peraza-Herrera et al., 2021) 350 76 Surgical Photo ThyroidXL (Duong et al., 2025) 7029 138 Ultrasound Tnbcnuclei (Naylor et al., 2018) 35 10 Microscopy TotalSegmentator (Wasserthal et al., 2023) 5206 154 CT, MRI UltrasoundNerve (Montoya, 2026) 1651 50 Ultrasound USforKidney (Song et al., 2022) 4351 50 Ultrasound UWSkinCancer (Vision & Lab, 2024) 143 44 Dermoscopy VinDr-Multiphase (Dao et al., 2022) 3486 44 CT WBC (Zheng et al., 2018) 280 40 Microscopy YeaZ (Dietler et al., 2020) 358 51 Microscopy YGA low dose ct (pazhoulab, 2024) 4387 44 CT
+
+[p. 14 | section: A. Data Sources | type: Text]
+We also appreciate MedSegBench(Kus¸ & Aydin, 2024) and MedSegDB (Zhang et al., 2025) for collecting and pre-processing some of these datasets.
+
+[p. 15 | section: B. Construction Details | type: FigureGroup]
+Figure 6. Construction details of three perspective. We manually curate the benchmark split to uphold high clinical standards. The remaining training data is validated through sampling-based quality checks, establishing a high-quality data proportion exceeding 95%.
+
+[p. 15 | section: C. Implementation Details of OmniGen2-MIE | type: TableGroup]
+Hyper-Parameter Value Finetuning method Full-Parameter Finetuning snr type lognorm do shift True dynamic time shift True Steps 20, 000 #GPUs 8 Per-device batch size 8 Gradient accumulation 1 Global batch size (effective) 64 Learning rate 1 × 10−4 LR scheduler timm constant with warmup Warm-up t 500 Precision BF16 Random seed 2233 Table 5. Training hyper-parameters used for finetuning OmniGen2-MIE on our dataset.
+
+[p. 16 | section: D.1.1. MATHEMATICS | type: Text]
+To recover the segmentation mask from the visualized output, we model the edited image O as a linear interpolation between the original background image B (a.k.a. the input image I) and a known overlay color C (red, green or blue). This relationship is governed by the per-pixel alpha channel α ∈ [0, 1], according to the standard alpha blending equation:
+
+[p. 16 | section: D.1.1. MATHEMATICS | type: Equation]
+\mathbf{O} = (1 - \alpha)\mathbf{B} + \alpha\mathbf{C} \tag{1}
+
+[p. 16 | section: D.1.1. MATHEMATICS | type: Text]
+By rearranging the terms as O − B = α(C − B), the scalar value α can be interpreted as the projection of the observed color shift onto the vector representing the maximum possible color change. To account for potential noise in the RGB space, we solve for α at each pixel using the least-squares solution:
+
+[p. 16 | section: D.1.1. MATHEMATICS | type: Equation]
+\alpha = \frac{(\mathbf{O} - \mathbf{B}) \cdot (\mathbf{C} - \mathbf{B})}{|\mathbf{C} - \mathbf{B}|^2} (2)
+
+[p. 16 | section: D.1.1. MATHEMATICS | type: Text]
+The continuous alpha map is subsequently binarized to produce the final segmentation mask M. This is achieved by applying a global threshold τ , such that:
+
+[p. 16 | section: D.1.1. MATHEMATICS | type: Equation]
+M_{i,j} = \begin{cases} 1 & \text{if } \alpha_{i,j} > \tau \\ 0 & \text{otherwise} \end{cases} (3)
+
+[p. 16 | section: D.1.1. MATHEMATICS | type: Text]
+In our implementation, a threshold of τ = 0.5 is utilized to effectively separate the predicted regions from the background.
+
+[p. 16 | section: D.1.2. CASE OF MASK RECONSTRUCTION | type: PictureGroup]
+Figure 7. Case of perception mask reconstruction.
+
+[p. 16 | section: Scoring Rubric for Modification Tasks | type: Text]
+You are a helpful assistant in evaluating medical image editing result.
+
+[p. 16 | section: Scoring Rubric for Modification Tasks | type: Text]
+You will be provided with an edit instruction and a collage image where the leftmost is origin image, center is edited image and rightmost is the reference ground truth image.
+
+[p. 16 | section: Scoring Rubric for Modification Tasks | type: Text]
+You should score how well an edited image matches the intended edit while preserving clinical realism and image integrity based on following scoring rubrics:
+
+[p. 16 | section: 1) Edit Goal Fulfillment (Edit Correctness): Assesses whether the intended lesion change is achieved. | type: Text]
+Scoring reference:
+
+[p. 16 | section: 1) Edit Goal Fulfillment (Edit Correctness): Assesses whether the intended lesion change is achieved. | type: ListGroup]
+5: Lesion added/removed exactly as intended; no residuals or unintended remnants. 4: Mostly correct; slight residual signal after removal or slight under/over-addition. 3: Partial success; lesion still partially present (removal) or incomplete/incorrect lesion (addition).
+
+[p. 17 | section: 1) Edit Goal Fulfillment (Edit Correctness): Assesses whether the intended lesion change is achieved. | type: ListGroup]
+2: Wrong area or wrong type of change; target lesion largely unchanged. 1: No effective edit or opposite edit performed. 2) Edit Area Morphology (Shape, Margins, Internal Structure): Evaluates whether edit area matches expected morphology and/or reference.
+
+[p. 17 | section: Scoring reference | type: ListGroup]
+5: Shape, border characteristics, and internal texture are highly consistent. 4: Minor border/shape irregularities; still plausible. 3: Morphology is generic/unconvincing; borders/texture inconsistent. 2: Clearly artificial morphology (blocky, repeated patterns, unnatural contours). 1: Morphology nonsensical or misleading (e.g., appears like different pathology). 3) Intensity / Signal / Attenuation Consistency: Checks whether edited region match modality-specific intensities.
+
+[p. 17 | section: Scoring reference | type: ListGroup]
+5: Intensities match local tissue statistics; no intensity discontinuities. 4: Slight intensity mismatch detectable with careful viewing. 3: Obvious mismatch (too bright/dark), inconsistent with modality or anatomy. 2: Strong intensity discontinuity; clearly edited. 1: Severe intensity errors that invalidate the image (e.g., saturation/clipping, inverted contrast). 4) Boundary Blending & Transition Naturalness: Rates edge blending and transitions between edited and unedited regions.
+
+[p. 17 | section: Scoring reference | type: ListGroup]
+5: Seamless blending; no halos, ringing, cut-paste edges. 4: Minor halo/edge artifacts only on close inspection. 3: Visible seams; boundary looks edited. 2: Strong cutout appearance or blur patches. 1: Boundary artifacts dominate the image. 5) Background / Non-target Preservation: Measures unintended changes outside the lesion edit region. Scoring reference: 5: Non-target anatomy and background unchanged (within expected noise). 4: Small unintended changes but not clinically meaningful. 3: Noticeable unintended alterations in nearby structures. 2: Large unintended modifications to anatomy or overall image. 1: Global corruption or major anatomical distortions. 6) Anatomical Plausibility & Clinical Coherence: Assesses whether result respects anatomy and pathology logic (e.g., lesion doesn't cross impossible boundaries). Scoring reference: 5: Fully plausible; consistent with organ boundaries and expected presentation. 4: Mostly plausible; minor oddity but acceptable. 3: Questionable plausibility (e.g., lesion overlaps structures unnaturally). 2: Clearly implausible anatomy/pathology relationship. 1: Clinically nonsensical or misleading. 7) Artifact Introduction (Noise, Texture, Aliasing, Compression, Repetition): Evaluates new artifacts introduced by editing. Scoring reference:
+
+[p. 18 | section: Scoring reference | type: ListGroup]
+5: No new artifacts; noise texture consistent with original. 4: Minor artifacts (subtle smoothing/grain mismatch). 3: Artifacts visible and distracting. 2: Strong artifacts (banding, checkerboard, repeated texture). 1: Severe artifacts preventing clinical use. 8) Image Quality & Acquisition Consistency: Checks consistency with scanner characteristics (resolution, blur, point spread, slice thickness cues, motion). Scoring reference: 5: Matches acquisition characteristics; sharpness/noise consistent. 4: Slight mismatch in sharpness or noise level. 3: Clear mismatch (over-smoothed or over-sharpened region). 2: Strong mismatch; edited region appears from different source. 1: Completely inconsistent with acquisition; unusable.
+
+[p. 18 | section: Return Format: Return a JSON dictionary with two fields | type: ListGroup]
+"conclusion": A brief conclusion to the edited image. "score list": The scores of the eight aspects in a JSON list.
+
+[p. 18 | section: Input collage VLM Output | type: Text]
+conclusion : The edited image successfully removes the renal cyst and replaces it with plausible parenchyma closely matching the reference, with only mild residual smoothing/texture mismatch in the filled region and minimal impact on surrounding anatomy.
+
+[p. 18 | section: Input collage VLM Output | type: Text]
+score_list : [5, 4, 4, 4, 5, 5, 4, 4]
+
+[p. 18 | section: Input collage VLM Output | type: Text]
+conclusion : The edited image does not introduce a clinically plausible gallstone. Instead, it contains multiple obvious synthetic circular overlays and line artifacts across the abdomen, with major non-target corruption and mismatch to CT texture; it does not match the reference which shows no such artifacts.
+
+[p. 18 | section: Input collage VLM Output | type: Text]
+score_list : [1, 1, 1, 1, 1, 1, 1, 1]
+
+[p. 18 | section: Input collage VLM Output | type: Caption]
+Figure 8. Cases of VLM rubric scoring.
+
+[p. 19 | section: E.1. Modality-Wise Performance Analysis | type: FigureGroup]
+Figure 9. Modality-wise performance analysis result within perception perspective. Left: DICE score; right: PSNR score.
+
+[p. 19 | section: E.1. Modality-Wise Performance Analysis | type: Text]
+To investigate the impact of modality deviation, we conduct a modality-wise analysis of the benchmarking results within the Perception perspective. Specifically, we report the DICE and PSNR scores of six representative models across all medical imaging modalities included in MieDB-100k. As illustrated in Fig. 9, the experimental results are consistent with our hypotheses. For the baseline models, performance is unevenly distributed across the various modalities: They achieve relatively strong results on modalities that resemble natural images, such as Endoscopy, Dermoscopy, and Surgical Photo. However, on non-optical modalities (e.g., CT, MRI, Ultrasound), their performance degrades drastically. In contrast, the model trained on our dataset exhibits balanced and superior performance across all imaging types. Collectively, these results demonstrate that a diverse dataset like MieDB-100k is essential for successfully adapting multi-modal generative models to the medical domain.
+
+[p. 19 | section: E.2. Multi-Round Generation | type: TableGroup]
+Table 6. Multi-round generation result. Best values are marked in Bold Perception Trasnformation DICE P-A P-ACC B-PSNR B-SSIM PSNR SSIM Pass@1 Pass@3 Pass@1 Pass@3 Pass@1 Pass@3 Pass@1 Pass@3 Pass@1 Pass@3 Pass@1 Pass@3 Open-Source SDXL-turbo 0.002 0.003 0.000 0.000 16.6 17.0 0.467 0.484 15.9 16.1 0.397 0.427 Bagel 0.263 0.383 0.069 0.137 13.9 16.1 0.620 0.703 12.7 15.2 0.442 0.548 OmniGen2 0.248 0.357 0.065 0.125 11.9 14.4 0.541 0.628 8.3 16.0 0.280 0.551 Step1X-Edit 0.332 0.369 0.126 0.143 15.5 16.4 0.727 0.748 16.6 17.1 0.539 0.558 FLUX.1-Kontext-dev 0.347 0.41 0.126 0.174 15.4 16.5 0.701 0.761 17.9 19.5 0.543 0.602 Qwen-Image-Edit 0.387 0.493 0.153 0.249 15.4 17.4 0.722 0.795 18.9 20.3 0.606 0.652 OmniGen2-MIE (Ours) 0.831 0.856 0.737 0.789 28.1 28.8 0.917 0.921 22.6 23.3 0.685 0.711
+
+[p. 19 | section: E.2. Multi-Round Generation | type: Text]
+To mitigate the inherent variance of the generative process, we report Pass@3 scores for the open-source models on Perception tasks. Specifically, we generate three independent outputs for each editing task and select the highest-performing sample to represent the task's score. These results are then averaged across all tasks to provide a robust assessment of overall performance.
+
+[p. 19 | section: E.2. Multi-Round Generation | type: Text]
+The results of the multi-round generation tests are summarized in Table 6. While multi-round generation improves the absolute scores for baseline models, it does not alter the underlying fact that these models lack essential medical knowledge. Furthermore, the significant fluctuations across rounds expose the high-variance nature of these baselines, undermining their reliability under clinical applications. In contrast, our model exhibits remarkable stability across all three trials. This
+
+[p. 20 | section: E.2. Multi-Round Generation | type: Text]
+consistency suggests that model trained on MieDB-100k has developed a deterministic understanding of medical concepts rather than relying on fortuitous generation.
+
+[p. 20 | section: E.3. Out-Of-Distribution Image Edit | type: Text]
+While Section 4.5 demonstrates that the model trained on MieDB-100k generalizes effectively to OOD editing targets, we further evaluate its robustness by performing edits on 'in-the-wild' medical images sourced from the internet (Fig. 10) .
+
+[p. 20 | section: E.3. Out-Of-Distribution Image Edit | type: FigureGroup]
+Figure 10. Examples of Out-Of-Distribution Editing.
+
+[p. 20 | section: E.3. Out-Of-Distribution Image Edit | type: Text]
+The results indicate that our model is capable of readily adapting to medical images outside of datasets. This suggests that the diversity of MieDB-100k has successfully decoupled the model from specific data distribution, allowing it to internalize generalizable edit operations that are applicable to real-world clinical scenarios.
+
+[p. 20 | section: F.1. Examples of Healthy Tissue Inpainting | type: FigureGroup]
+Figure 11. Examples of Inpainting. We train different inpainting models on each medical modalities. H: the Healthy image; L: the Lesion-bearing image.
+
+[p. 21 | section: F.2. Extended Examples of Qualitative result | type: Text]
+Input Image Bagel FLUX.1 KontextOmniGen2OmniGen2-MIE Reference Nano Banana ProGPT Qwen Image Edit -Image [Edit Instruction] Illustrate all nuclei instances by painting GREEN masks directly onto the microscopy image, with zero modification to the background. Edit [Edit Instruction] Use a RED mask to highlight the polyp(s) by drawing it directly onto the given image, preserving the background exactly as is. Edit [Edit Instruction] Overlay unique RED masks on each organ in the ultrasound image by painting them directly, keeping the original background intact. Edit [Edit Instruction] Visually segment Caenorhabditis elegans by applying BLUE masks directly onto the microscope image without affecting the background. [Edit Instruction] Mark the region corresponding to BLADDER TUMOR in the Enhanced CT image using a solid RED mask applied directly, while ensuring the background stays unmodified. [Edit Instruction] Encode the lesion location with a GREEN mask drawn directly on the image, maintaining original background appearance. [Edit Instruction] Cover the infected zones with a RED mask by painting it directly onto the chest X-ray, ensuring background fidelity. Edit Edit Edit Edit
+
+[p. 22 | section: F.2. Extended Examples of Qualitative result | type: PictureGroup]
+[Edit Instruction] Delete all visible manifestations of COVID pneumonia from the X-ray image and replace them with normal lung appearance, maintaining vascular and bronchial structures.
+
+[p. 22 | section: F.2. Extended Examples of Qualitative result | type: PictureGroup]
+[Edit Instruction] Eliminate the Tumor in the Stomach while ensuring the reconstructed zone blends naturally with adjacent healthy tissue in terms of Hounsfield units and spatial patterns.
+
+[p. 22 | section: F.2. Extended Examples of Qualitative result | type: PictureGroup]
+[Edit Instruction] Simulate a focal mucosal lesion in the white masked region that could represent a common colonic or gastric polyp, based on realistic endoscopic criteria.
+
+[p. 22 | section: F.2. Extended Examples of Qualitative result | type: PictureGroup]
+[Edit Instruction] Reconstruct the affected skin area by removing the lesion and generating realistic, symmetryconsistent healthy tissue based on surrounding context.
+
+[p. 22 | section: F.2. Extended Examples of Qualitative result | type: PictureGroup]
+[Edit Instruction] Convert the white masked area into a realistic representation of healthy skin, avoiding artificial smoothness or color mismatches.
+
+[p. 22 | section: F.2. Extended Examples of Qualitative result | type: PictureGroup]
+[Edit Instruction] Synthetically insert a thyroid nodule into the given ultrasound scan, ensuring it exhibits clinically plausible features such as shape, margin, echogenicity, and vascularity.
+
+[p. 22 | section: F.2. Extended Examples of Qualitative result | type: PictureGroup]
+[Edit Instruction] Add a convincing brain tumor to the FLAIR MRI while maintaining correct contrast dynamics—such as suppressed CSF and bright pathological signal.
+
+[p. 23 | section: F.2. Extended Examples of Qualitative result | type: PictureGroup]
+[Edit Instruction] Improve diagnostic clarity of the CT by removing metal degradation while ensuring zero change to medically relevant content outside the artifact zone.
+
+[p. 23 | section: F.2. Extended Examples of Qualitative result | type: PictureGroup]
+[Edit Instruction] Filter out motion artifacts from the PD MRI scan without smoothing, warping, or otherwise altering true tissue signals.
+
+[p. 23 | section: F.2. Extended Examples of Qualitative result | type: PictureGroup]
+[Edit Instruction] Render the CT with Abdominal window parameters to mimic how it would appear on a PACS viewer configured for that tissue type.
+
+[p. 23 | section: F.2. Extended Examples of Qualitative result | type: PictureGroup]
+[Edit Instruction] Synthesize a Arterial phase-equivalent CT image from the provided Non-Contrast phase input, ensuring realistic vascular and tissue enhancement patterns.
+
+[p. 23 | section: F.2. Extended Examples of Qualitative result | type: PictureGroup]
+[Edit Instruction] Apply post-processing to remove metal artifacts from the CT image without smoothing, interpolating, or altering real anatomical details unnecessarily.
+
+[p. 23 | section: F.2. Extended Examples of Qualitative result | type: PictureGroup]
+[Edit Instruction] Convert the given T2-MRI image to resemble a CT image, maintaining all underlying medical information intact.
+
+[p. 23 | section: F.2. Extended Examples of Qualitative result | type: PictureGroup]
+[Edit Instruction] Process the low-dose CT using a SOFT kernel to suppress graininess while preserving low-contrast diagnostic details.
+
+[p. 24 | section: G. Failure Cases | type: FigureGroup]
+Figure 12. Failure Cases.
+
+[p. 24 | section: G. Failure Cases | type: Text]
+Fig 12 illustrates several representative failure cases of OmniGen2-MIE. The most frequent failure modes include: (1) semantic confusion between targeted anatomical features and morphologically similar background tissues; (2) intensity inconsistency, where the brightness of the edited region deviates from the surrounding context in a physically implausible manner; and (3) background inconsistency, especially after holistic transformations. These limitations underscore the need for more sophisticated multimodal architectures capable of preserving fine-grained details, as well as even more comprehensive training datasets to satisfy the requirements of rigorous clinical applications.
+
+[p. 24 | section: H. Limitations | type: Text]
+Despite MieDB-100k provides a large-scale and diverse dataset for medical image editing, the primary limitation lies in the inherent difficulty of capturing ALL possible medical imaging modalities, and the relative scarcity of data for rare clinical cases. Continuous efforts to enrich these underrepresented categories will be vital for enhancing the dataset's diversity and effectivity. Furthermore, while our work establishes a foundation for unified understanding and generation, it focuses exclusively on editing tasks. Integrating medical VQA and text-to-image datasets represents a natural progression of this research direction, resulting in a more comprehensive resource for the development of holistic medical models.

icml26/80c20b7b-ead6-454a-849e-56702a6c828f/assets.json

@@ -0,0 +1,310 @@
+[
+  {
+    "filename": "_page_14_Figure_2.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/80c20b7b-ead6-454a-849e-56702a6c828f/marker_markdown/80c20b7b-ead6-454a-849e-56702a6c828f/_page_14_Figure_2.jpeg",
+    "bytes": 100469,
+    "width": 1307,
+    "height": 527,
+    "aspect_ratio": 2.480075901328273,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_14_Figure_2.jpeg"
+  },
+  {
+    "filename": "_page_15_Picture_12.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/80c20b7b-ead6-454a-849e-56702a6c828f/marker_markdown/80c20b7b-ead6-454a-849e-56702a6c828f/_page_15_Picture_12.jpeg",
+    "bytes": 18690,
+    "width": 640,
+    "height": 265,
+    "aspect_ratio": 2.4150943396226414,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_15_Picture_12.jpeg"
+  },
+  {
+    "filename": "_page_17_Picture_26.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/80c20b7b-ead6-454a-849e-56702a6c828f/marker_markdown/80c20b7b-ead6-454a-849e-56702a6c828f/_page_17_Picture_26.jpeg",
+    "bytes": 16053,
+    "width": 481,
+    "height": 184,
+    "aspect_ratio": 2.614130434782609,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_17_Picture_26.jpeg"
+  },
+  {
+    "filename": "_page_17_Picture_31.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/80c20b7b-ead6-454a-849e-56702a6c828f/marker_markdown/80c20b7b-ead6-454a-849e-56702a6c828f/_page_17_Picture_31.jpeg",
+    "bytes": 29798,
+    "width": 669,
+    "height": 208,
+    "aspect_ratio": 3.2163461538461537,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_17_Picture_31.jpeg"
+  },
+  {
+    "filename": "_page_18_Figure_3.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/80c20b7b-ead6-454a-849e-56702a6c828f/marker_markdown/80c20b7b-ead6-454a-849e-56702a6c828f/_page_18_Figure_3.jpeg",
+    "bytes": 58829,
+    "width": 1105,
+    "height": 549,
+    "aspect_ratio": 2.012750455373406,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_18_Figure_3.jpeg"
+  },
+  {
+    "filename": "_page_19_Figure_4.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/80c20b7b-ead6-454a-849e-56702a6c828f/marker_markdown/80c20b7b-ead6-454a-849e-56702a6c828f/_page_19_Figure_4.jpeg",
+    "bytes": 102839,
+    "width": 1329,
+    "height": 479,
+    "aspect_ratio": 2.7745302713987474,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_19_Figure_4.jpeg"
+  },
+  {
+    "filename": "_page_19_Figure_9.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/80c20b7b-ead6-454a-849e-56702a6c828f/marker_markdown/80c20b7b-ead6-454a-849e-56702a6c828f/_page_19_Figure_9.jpeg",
+    "bytes": 61735,
+    "width": 826,
+    "height": 490,
+    "aspect_ratio": 1.6857142857142857,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_19_Figure_9.jpeg"
+  },
+  {
+    "filename": "_page_1_Figure_1.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/80c20b7b-ead6-454a-849e-56702a6c828f/marker_markdown/80c20b7b-ead6-454a-849e-56702a6c828f/_page_1_Figure_1.jpeg",
+    "bytes": 163874,
+    "width": 1315,
+    "height": 571,
+    "aspect_ratio": 2.3029772329246936,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_1_Figure_1.jpeg"
+  },
+  {
+    "filename": "_page_21_Picture_10.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/80c20b7b-ead6-454a-849e-56702a6c828f/marker_markdown/80c20b7b-ead6-454a-849e-56702a6c828f/_page_21_Picture_10.jpeg",
+    "bytes": 17305,
+    "width": 1274,
+    "height": 122,
+    "aspect_ratio": 10.442622950819672,
+    "keep": false,
+    "reject_reason": "extreme_aspect_ratio",
+    "model_path": null
+  },
+  {
+    "filename": "_page_21_Picture_12.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/80c20b7b-ead6-454a-849e-56702a6c828f/marker_markdown/80c20b7b-ead6-454a-849e-56702a6c828f/_page_21_Picture_12.jpeg",
+    "bytes": 27466,
+    "width": 1268,
+    "height": 136,
+    "aspect_ratio": 9.323529411764707,
+    "keep": false,
+    "reject_reason": "extreme_aspect_ratio",
+    "model_path": null
+  },
+  {
+    "filename": "_page_21_Picture_14.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/80c20b7b-ead6-454a-849e-56702a6c828f/marker_markdown/80c20b7b-ead6-454a-849e-56702a6c828f/_page_21_Picture_14.jpeg",
+    "bytes": 28054,
+    "width": 1243,
+    "height": 138,
+    "aspect_ratio": 9.007246376811594,
+    "keep": false,
+    "reject_reason": "extreme_aspect_ratio",
+    "model_path": null
+  },
+  {
+    "filename": "_page_21_Picture_2.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/80c20b7b-ead6-454a-849e-56702a6c828f/marker_markdown/80c20b7b-ead6-454a-849e-56702a6c828f/_page_21_Picture_2.jpeg",
+    "bytes": 29366,
+    "width": 1295,
+    "height": 157,
+    "aspect_ratio": 8.248407643312103,
+    "keep": false,
+    "reject_reason": "extreme_aspect_ratio",
+    "model_path": null
+  },
+  {
+    "filename": "_page_21_Picture_4.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/80c20b7b-ead6-454a-849e-56702a6c828f/marker_markdown/80c20b7b-ead6-454a-849e-56702a6c828f/_page_21_Picture_4.jpeg",
+    "bytes": 32719,
+    "width": 1243,
+    "height": 126,
+    "aspect_ratio": 9.865079365079366,
+    "keep": false,
+    "reject_reason": "extreme_aspect_ratio",
+    "model_path": null
+  },
+  {
+    "filename": "_page_21_Picture_6.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/80c20b7b-ead6-454a-849e-56702a6c828f/marker_markdown/80c20b7b-ead6-454a-849e-56702a6c828f/_page_21_Picture_6.jpeg",
+    "bytes": 33194,
+    "width": 1275,
+    "height": 142,
+    "aspect_ratio": 8.97887323943662,
+    "keep": false,
+    "reject_reason": "extreme_aspect_ratio",
+    "model_path": null
+  },
+  {
+    "filename": "_page_21_Picture_8.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/80c20b7b-ead6-454a-849e-56702a6c828f/marker_markdown/80c20b7b-ead6-454a-849e-56702a6c828f/_page_21_Picture_8.jpeg",
+    "bytes": 18811,
+    "width": 1258,
+    "height": 128,
+    "aspect_ratio": 9.828125,
+    "keep": false,
+    "reject_reason": "extreme_aspect_ratio",
+    "model_path": null
+  },
+  {
+    "filename": "_page_22_Picture_11.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/80c20b7b-ead6-454a-849e-56702a6c828f/marker_markdown/80c20b7b-ead6-454a-849e-56702a6c828f/_page_22_Picture_11.jpeg",
+    "bytes": 26731,
+    "width": 1255,
+    "height": 134,
+    "aspect_ratio": 9.365671641791044,
+    "keep": false,
+    "reject_reason": "extreme_aspect_ratio",
+    "model_path": null
+  },
+  {
+    "filename": "_page_22_Picture_13.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/80c20b7b-ead6-454a-849e-56702a6c828f/marker_markdown/80c20b7b-ead6-454a-849e-56702a6c828f/_page_22_Picture_13.jpeg",
+    "bytes": 33082,
+    "width": 1262,
+    "height": 136,
+    "aspect_ratio": 9.279411764705882,
+    "keep": false,
+    "reject_reason": "extreme_aspect_ratio",
+    "model_path": null
+  },
+  {
+    "filename": "_page_22_Picture_15.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/80c20b7b-ead6-454a-849e-56702a6c828f/marker_markdown/80c20b7b-ead6-454a-849e-56702a6c828f/_page_22_Picture_15.jpeg",
+    "bytes": 27396,
+    "width": 1243,
+    "height": 142,
+    "aspect_ratio": 8.753521126760564,
+    "keep": false,
+    "reject_reason": "extreme_aspect_ratio",
+    "model_path": null
+  },
+  {
+    "filename": "_page_22_Picture_2.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/80c20b7b-ead6-454a-849e-56702a6c828f/marker_markdown/80c20b7b-ead6-454a-849e-56702a6c828f/_page_22_Picture_2.jpeg",
+    "bytes": 21965,
+    "width": 1262,
+    "height": 101,
+    "aspect_ratio": 12.495049504950495,
+    "keep": false,
+    "reject_reason": "extreme_aspect_ratio",
+    "model_path": null
+  },
+  {
+    "filename": "_page_22_Picture_3.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/80c20b7b-ead6-454a-849e-56702a6c828f/marker_markdown/80c20b7b-ead6-454a-849e-56702a6c828f/_page_22_Picture_3.jpeg",
+    "bytes": 32717,
+    "width": 1249,
+    "height": 151,
+    "aspect_ratio": 8.271523178807946,
+    "keep": false,
+    "reject_reason": "extreme_aspect_ratio",
+    "model_path": null
+  },
+  {
+    "filename": "_page_22_Picture_5.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/80c20b7b-ead6-454a-849e-56702a6c828f/marker_markdown/80c20b7b-ead6-454a-849e-56702a6c828f/_page_22_Picture_5.jpeg",
+    "bytes": 24968,
+    "width": 1245,
+    "height": 124,
+    "aspect_ratio": 10.040322580645162,
+    "keep": false,
+    "reject_reason": "extreme_aspect_ratio",
+    "model_path": null
+  },
+  {
+    "filename": "_page_22_Picture_7.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/80c20b7b-ead6-454a-849e-56702a6c828f/marker_markdown/80c20b7b-ead6-454a-849e-56702a6c828f/_page_22_Picture_7.jpeg",
+    "bytes": 32967,
+    "width": 1253,
+    "height": 134,
+    "aspect_ratio": 9.350746268656716,
+    "keep": false,
+    "reject_reason": "extreme_aspect_ratio",
+    "model_path": null
+  },
+  {
+    "filename": "_page_22_Picture_9.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/80c20b7b-ead6-454a-849e-56702a6c828f/marker_markdown/80c20b7b-ead6-454a-849e-56702a6c828f/_page_22_Picture_9.jpeg",
+    "bytes": 32269,
+    "width": 1256,
+    "height": 132,
+    "aspect_ratio": 9.515151515151516,
+    "keep": false,
+    "reject_reason": "extreme_aspect_ratio",
+    "model_path": null
+  },
+  {
+    "filename": "_page_23_Figure_2.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/80c20b7b-ead6-454a-849e-56702a6c828f/marker_markdown/80c20b7b-ead6-454a-849e-56702a6c828f/_page_23_Figure_2.jpeg",
+    "bytes": 111959,
+    "width": 1053,
+    "height": 565,
+    "aspect_ratio": 1.8637168141592921,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_23_Figure_2.jpeg"
+  },
+  {
+    "filename": "_page_2_Figure_3.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/80c20b7b-ead6-454a-849e-56702a6c828f/marker_markdown/80c20b7b-ead6-454a-849e-56702a6c828f/_page_2_Figure_3.jpeg",
+    "bytes": 30757,
+    "width": 652,
+    "height": 347,
+    "aspect_ratio": 1.8789625360230549,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_2_Figure_3.jpeg"
+  },
+  {
+    "filename": "_page_3_Figure_1.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/80c20b7b-ead6-454a-849e-56702a6c828f/marker_markdown/80c20b7b-ead6-454a-849e-56702a6c828f/_page_3_Figure_1.jpeg",
+    "bytes": 161268,
+    "width": 1253,
+    "height": 801,
+    "aspect_ratio": 1.564294631710362,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_3_Figure_1.jpeg"
+  },
+  {
+    "filename": "_page_6_Figure_1.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/80c20b7b-ead6-454a-849e-56702a6c828f/marker_markdown/80c20b7b-ead6-454a-849e-56702a6c828f/_page_6_Figure_1.jpeg",
+    "bytes": 290786,
+    "width": 1230,
+    "height": 1254,
+    "aspect_ratio": 0.9808612440191388,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_6_Figure_1.jpeg"
+  },
+  {
+    "filename": "_page_7_Figure_10.jpeg",
+    "path": "data/processed_papers/icml26_20260429_1952_duequeue/marker_raw/80c20b7b-ead6-454a-849e-56702a6c828f/marker_markdown/80c20b7b-ead6-454a-849e-56702a6c828f/_page_7_Figure_10.jpeg",
+    "bytes": 62586,
+    "width": 674,
+    "height": 478,
+    "aspect_ratio": 1.4100418410041842,
+    "keep": true,
+    "reject_reason": null,
+    "model_path": "assets/_page_7_Figure_10.jpeg"
+  }
+]

icml26/80c20b7b-ead6-454a-849e-56702a6c828f/dataset_meta.json

@@ -0,0 +1,63 @@
+{
+  "paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f",
+  "pipeline": "Paper2Markdown-V3",
+  "ok": true,
+  "page_count": 24,
+  "chunk_count": 140,
+  "main_body_chunk_count": 70,
+  "appendix_chunk_count": 57,
+  "reference_chunk_count": 13,
+  "model_text_chars": 34560,
+  "raw_markdown_chars": 87319,
+  "sanitized_chars": 77147,
+  "page_provenance": {
+    "min_page": 1,
+    "max_page": 24,
+    "invalid_count": 0
+  },
+  "marker_block_type_counts": {
+    "Caption": 3,
+    "Equation": 3,
+    "Figure": 1,
+    "FigureGroup": 9,
+    "Footnote": 1,
+    "ListGroup": 19,
+    "PageFooter": 24,
+    "PageHeader": 31,
+    "Picture": 3,
+    "PictureGroup": 15,
+    "SectionHeader": 55,
+    "TableGroup": 6,
+    "Text": 367
+  },
+  "asset_count_raw": 28,
+  "asset_count_model_kept": 13,
+  "asset_count_rejected": 15,
+  "asset_reject_reasons": {
+    "extreme_aspect_ratio": 15,
+    "kept": 13
+  },
+  "artifact_leak_audit": {
+    "ok": true,
+    "hits": {
+      "Anonymous Authors": [],
+      "ACKNOWLEDGMENT": [],
+      "OpenReview": [],
+      "\"accept_label\"": [],
+      "\"decision\"": [],
+      "\"decision_tier\"": [],
+      "\"source_status\"": [],
+      "Meta-review": [],
+      "Official Review": [],
+      "official_reviews": [],
+      "meta_reviews": [],
+      "suggested_verdict_score": []
+    },
+    "artifact_count": 2
+  },
+  "default_model_input": "model_text_v3.txt",
+  "appendix_input": "appendix_text_v3.txt",
+  "reference_input": "reference_text_v3.txt",
+  "source": "koala_icml26_due_queue",
+  "run_name": "icml26_20260429_1952_duequeue"
+}

icml26/80c20b7b-ead6-454a-849e-56702a6c828f/main_body_chunks.jsonl

@@ -0,0 +1,70 @@
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0000", "section": "Abstract", "page_start": 1, "page_end": 1, "type": "Text", "text": "The scarcity of high-quality data remains a primary bottleneck in adapting multimodal generative models for medical image editing. Existing medical image editing datasets often suffer from limited diversity, neglect of medical image understanding and inability to balance quality with scalability. To address these gaps, we propose MieDB-100k, a large-scale, high-quality and diverse dataset for text-guided medical image editing. It categorizes editing tasks into perspectives of Perception, Modification and Transformation, considering both understanding and generation abilities. We construct MieDB-100k via a data curation pipeline leveraging both modality-specific expert models and rule-based data synthetic methods, followed by rigorous manual inspection to ensure clinical fidelity. Extensive experiments demonstrate that model trained with MieDB-100k consistently outperform both open-source and proprietary models while exhibiting strong generalization ability. We anticipate that this dataset will serve as a cornerstone for future advancements in specialized medical image editing. Dataset and code are publicly available at ub.com/Raiiyf/MieDB-100k", "source": "marker_v2", "marker_block_id": "/page/0/Text/4"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0001", "section": "1. Introduction", "page_start": 1, "page_end": 1, "type": "Text", "text": "Multimodal generative models (Wu et al., 2025a ;b; Liu et al., 2025c) have developed rapidly in recent years. In natural image domains, generative models are not only gradually unifying text-guided generation and editing tasks, but also progressively expanding their capabilities to encompass image modification and image understanding (Deng et al., 2025; Tong et al., 2025) . However, in medical image domains, their performance remains conspicuously limited, especially in the area of unified editing tasks (Liu et al.,", "source": "marker_v2", "marker_block_id": "/page/0/Text/6"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0002", "section": "1. Introduction", "page_start": 1, "page_end": 1, "type": "Text", "text": "2025b; Yang et al., 2025) . We attribute this performance degradation primarily to a fundamental scarcity of specialized medical image-editing data.", "source": "marker_v2", "marker_block_id": "/page/0/Text/9"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0003", "section": "1. Introduction", "page_start": 1, "page_end": 1, "type": "Text", "text": "While a few contemporary studies have proposed benchmarks or datasets for medical image editing, they remain insufficient in three key aspects: (1) limited diversity in medical image modalities. Unlike general computer vision, clinical imaging encompasses diverse modalities with distinct physical and structural foundations. However, existing research and datasets are restricted to a narrow range of imaging modalities (Chen & Feng, 2025; Liu et al., 2025b) , typically the widely available modalities such as Chest Xrays and CTs, which cannot adequately train or evaluate a model's ability across diverse clinical settings.", "source": "marker_v2", "marker_block_id": "/page/0/Text/10"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0004", "section": "1. Introduction", "page_start": 1, "page_end": 1, "type": "ListGroup", "text": "(2) Neglect of medical image understanding. Almost all medical image editing works only focus on conceptual modification and stylistic transformation tasks, but ignore visual perception tasks ( e.g. organ/lesion detection), which has been considered to be beneficial to the generation of image editing models (Huang et al., 2025; Deng et al., 2025) . Additionally, this clinical grounding ensures interpretability and corrects 'right-for-the-wrong-reason' edits, which is vital for safety-critical medical applications. Finally, neglecting understanding tasks also hinders the development of unified medical models that bridge understanding and generation. (3) Failure to ensure both data quality and scalability. Collection of medical image editing data is hindered by the difficulty of generating ground-truth counterfactuals. Some existing studies (Chen & Feng, 2025; Yang et al., 2025) distills general-purpose generative models for quick data scaling. However, these models are not tailored for medical use and hence produce results that lack clinical reliability and explainability. Conversely, previous work (Liu et al., 2025b) relies on extensive human involvement to manually collect real medical image pairs, which is notoriously difficult to scale up. Moreover, real-world longitudinal data often exhibits spatial misalignment and background inconsistency, as obtaining perfectly calibrated scan pairs is rare in practical medical settings.", "source": "marker_v2", "marker_block_id": "/page/0/ListGroup/495"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0005", "section": "1. Introduction", "page_start": 1, "page_end": 1, "type": "Text", "text": "In this paper, we address the aforementioned limitations in previous research by introducing MieDB-100k, a largescale, high-quality and diverse dataset for text-guided medical image editing. MieDB-100k includes 112, 228 editing", "source": "marker_v2", "marker_block_id": "/page/0/Text/13"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0006", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "FigureGroup", "text": "Figure 1. MieDB-100k overview. It categorizes medical image editing tasks into three perspectives, covering diverse medical modalities.", "source": "marker_v2", "marker_block_id": "/page/1/FigureGroup/324"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0007", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "Text", "text": "data, covering 69 distinct editing targets and 10 diverse medical image modalities. We categorize editing tasks into three types: Perception , Modification and Transformation , which consider both model's intrinsic understanding and generation abilities on medical images. To enhance the data fidelity while preserving the scalability, we propose a data curation pipeline leveraging both modality-specific expert models and rule-based data synthetic methods. Additionally, for some complex tasks such as lesion modification, we introduce individuals with medical knowledge to perform manual quality checks on the data to ensure data quality. Finally, we introduced task-specific evaluation metrics to facilitate a comprehensive assessment of the editing models' performance.", "source": "marker_v2", "marker_block_id": "/page/1/Text/3"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0008", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "Text", "text": "We evaluate existing open-source and closed-source multimodal generative models on MieDB-100k and argue that most of them cannot perform well in medical image editing. To further validate the reliability and utility of MieDB-100k, we finetune the OmniGen2 baseline on our dataset. Experimental results demonstrate that MieDB-100k facilitates a substantial performance leap in medical image editing tasks, surpassing or matching SOTA models including Nano Banana Pro. It also exhibits strong generalization ability driven by the synergy of understanding and generation tasks. We anticipate that this dataset will serve as a cornerstone for future advancements in specialized medical image editing.", "source": "marker_v2", "marker_block_id": "/page/1/Text/4"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0009", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "Text", "text": "Our contributions can be summarized as follows:", "source": "marker_v2", "marker_block_id": "/page/1/Text/5"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0010", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "ListGroup", "text": "(1) We propose a credible and scalable data curation pipeline to construct MieDB-100k , a large-scale, high-quality and highly diverse dataset for medical image editing with 69 targets and 10 medical image modalities. (2) We first unify the medical image understanding and generation into the paradigm of edit, and find that joint", "source": "marker_v2", "marker_block_id": "/page/1/ListGroup/325"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0011", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "Text", "text": "training yields performance gains for specific tasks.", "source": "marker_v2", "marker_block_id": "/page/1/Text/8"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0012", "section": "1. Introduction", "page_start": 2, "page_end": 2, "type": "Text", "text": "(3) We evaluate popular open-source and closed-source multimodal generative models on MieDB-100k , and observe that training with our data can significantly strengthens the model's capacity for medical image editing.", "source": "marker_v2", "marker_block_id": "/page/1/Text/9"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0013", "section": "2.1. Data Research for Medical Image Editing", "page_start": 2, "page_end": 2, "type": "Text", "text": "As an emerging area, multimodal medical generative modeling is currently supported by relatively few publicly available datasets for training and benchmarking (Tab. 1). In these works, the primary challenge lies in the construction of high-quality image-edit pairs. MedEBench (Liu et al., 2025b), an early benchmarking effort, curated pairs by manually collecting related images from medical documents. While this ensures clinical validity, the approach lacks scalability. Furthermore, the resulting image pairs often exhibit background inconsistencies, as achieving strict spatial calibration in real-world clinical settings is virtually impossible. Conversely, Med-banana-50K (Chen & Feng, 2025) proposed a fully autonomous pipeline where data construction and quality control were managed by Gemini. However, applying general-purpose models to specialized medical scenarios may introduce factual errors or inconsistent edits, raising concerns about data fidelity. Finally, MedGEN-Bench (Yang et al., 2025) introduced image-edit pairs using a mix of rule-based and model-based methods; however, the lack of specific architectural details hinders a thorough evaluation of their data quality. Moreover, existing benchmarks only focus on content generation evaluation, overlooking the critical aspect of medical image understanding.", "source": "marker_v2", "marker_block_id": "/page/1/Text/12"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0014", "section": "2.1. Data Research for Medical Image Editing", "page_start": 3, "page_end": 3, "type": "TableGroup", "text": "Table 1. Comparison of contemporary medical image editing benchmarks and datasets. In 'Perspective' column, P stands for Perception, M stands for Modification, and T stands for Transformation. Benchmark Size Modalities Targets Perspectives Source Human Inspection MedE-Bench (Liu et al., 2025b) \\sim 1 k 4 13 M Real ✓ Med-banana-50K (Chen & Feng, 2025) \\sim 50k 3 23 M Synthetic X MedGEN-Bench (Yang et al., 2025) \\sim 6k 6 16 M, T Real & Synthetic ✓ MieDB-100k (Ours) \\sim 100k 10 69 P, M, T Real & Synthetic ✓", "source": "marker_v2", "marker_block_id": "/page/2/TableGroup/388"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0015", "section": "2.1. Data Research for Medical Image Editing", "page_start": 3, "page_end": 3, "type": "FigureGroup", "text": "Figure 2. Modality distribution (a) and prompt word cloud (b).", "source": "marker_v2", "marker_block_id": "/page/2/FigureGroup/389"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0016", "section": "2.2. Multimodal Generative Model", "page_start": 3, "page_end": 3, "type": "Text", "text": "129130", "source": "marker_v2", "marker_block_id": "/page/2/Text/28"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0017", "section": "2.2. Multimodal Generative Model", "page_start": 3, "page_end": 3, "type": "Text", "text": "131132", "source": "marker_v2", "marker_block_id": "/page/2/Text/29"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0018", "section": "2.2. Multimodal Generative Model", "page_start": 3, "page_end": 3, "type": "Text", "text": "Multimodal generative models (Liu et al., 2025c; Brooks et al., 2023) accept both images and natural language instructions as input, performing edits by translating semantic commands into precise visual manipulations. Recent studies (Wu et al., 2025a) often leverage vision-language model encoder and large-scale vision-language pretraining to align the semantic instruction with image modification. For instance, OmniGen2 (Wu et al., 2025b) utilizes Qwen2.5-VL (Bai et al., 2025b) to extract latent representations for semantic alignment, supported by a large-scale, multi-task training strategy. Furthermore, many recent studies (Deng et al., 2025) integrate image understanding and editing within a unified architecture. Exploiting these synergies is essential for creating robust models that are capable of performing both multimodal understanding and visual generation. On the commercial front, SOTA proprietary models like Gemini-3-Pro-Image (Nano Banana Pro) (DeepMind, 2025a) exhibit sophisticated image manipulation abilities, further realizing the real-world potential of multi-modal generative models. Despite these advancements, current models still struggle with the complexities of medical imaging(Liu et al., 2025b; Yang et al., 2025), highlighting the urgent need for comprehensive datasets to accelerate their adaptation to clinical domains.", "source": "marker_v2", "marker_block_id": "/page/2/Text/6"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0019", "section": "3. MieDB-100k", "page_start": 3, "page_end": 3, "type": "Text", "text": "This section introduces MieDB-100k, a high-quality, rigorous, and highly diverse dataset for medical image editing with more than 69 associated medical targets. It contains", "source": "marker_v2", "marker_block_id": "/page/2/Text/8"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0020", "section": "3. MieDB-100k", "page_start": 3, "page_end": 3, "type": "Text", "text": "112, 228 image-editing triplets. Figure 2(a) summarizes the distribution of samples across 10 imaging modalities.", "source": "marker_v2", "marker_block_id": "/page/2/Text/9"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0021", "section": "3.1. Data Definition", "page_start": 3, "page_end": 3, "type": "Text", "text": "Each entry in MieDB-100k is a triplet (I, P, O), where I is the input medical image, P is the textual prompt that describes edit operation, and O is the target image.", "source": "marker_v2", "marker_block_id": "/page/2/Text/11"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0022", "section": "3.2. Three Perspectives of MieDB-100k", "page_start": 3, "page_end": 3, "type": "Text", "text": "MieDB-100k is constructed under a novel categorization of three perspectives, considering both understanding and generation capabilities: (1) Perception tasks, which focus on model's intrinsic medical knowledge via pixel-wise identification of prompted clinical targets in the input image; (2) Modification tasks, which require the model to locate and alter specific medical features; and (3) Transformation tasks, involving medical image restoration, enhancement, and other low-level transformation. To ensure the rigor of the data triplets while maintaining scalability, we designed and implemented a specialized data construction pipeline for MieDB-100k (Fig 3), and we list all source datasets used for construction in App. A.", "source": "marker_v2", "marker_block_id": "/page/2/Text/13"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0023", "section": "3.2.1. Perception", "page_start": 3, "page_end": 3, "type": "Text", "text": "Perception tasks focus on medical image understanding, and we we formulate it as an editing task by instructing model to generate masks over regions of interest (ROIs), such as specific organs or lesions, through textual prompts. Notably, to align with image editing paradigm, the model is prompted to overlay the localization mask directly onto the source image rather than generating a standalone binary mask. This task serves two primary functions: First, since the mask-painting task only requires minimal pixel manipulation (typically modifying a single channel within a specific region), it serves as a direct assessment of the medical knowledge embedded in the generative model, isolating its perceptual accuracy from complex synthesis capabilities. Second, it introduces a promising application for multimodal generative models in the medical domain: assisted interpretation in multimodal manner. By allowing users to highlight specific targets in medical image through natural language prompts, this approach can assist", "source": "marker_v2", "marker_block_id": "/page/2/Text/15"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0024", "section": "3.2.1. Perception", "page_start": 4, "page_end": 4, "type": "Text", "text": "patients in understanding their diagnostic images, aid medical students in their education, and reduce screening time for senior clinicians.", "source": "marker_v2", "marker_block_id": "/page/3/Text/2"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0025", "section": "3.2.1. Perception", "page_start": 4, "page_end": 4, "type": "Text", "text": "The rule-based construction process for the Perception task's data triplets is illustrated in Fig. 3. Specifically, for a segmentation dataset, the original image serves as the input I. The output image O is synthesized by overlaying the ground-truth segmentation label, which is rendered in a randomly selected color (red, green, or blue), onto the input image with an alpha-blending transparency of 0.6. The ROIs of perception can be classified into three types: anatomical structure (organ, organism and so on), lesion area and holistic segmentation (segment all visible and clinically significant structures). We specifies the perception target and visualizing color scheme in the textual prompt P. Since this part of the data is constructed following a definite rule, it can be readily scaled up to a diverse set of medical knowledge assessments and to the associated training dataset by leveraging the extensive body of existing medical segmentation research. Finally, to ensure a high-quality final benchmark, we manually filtered the initial data pool to remove trivial, redundant, or incorrectly labeled samples.", "source": "marker_v2", "marker_block_id": "/page/3/Text/3"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0026", "section": "3.2.2. MODIFICATION", "page_start": 4, "page_end": 4, "type": "Text", "text": "167168", "source": "marker_v2", "marker_block_id": "/page/3/Text/11"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0027", "section": "3.2.2. MODIFICATION", "page_start": 4, "page_end": 4, "type": "Text", "text": "171172", "source": "marker_v2", "marker_block_id": "/page/3/Text/14"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0028", "section": "3.2.2. MODIFICATION", "page_start": 4, "page_end": 4, "type": "Text", "text": "177178179", "source": "marker_v2", "marker_block_id": "/page/3/Text/18"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0029", "section": "3.2.2. MODIFICATION", "page_start": 4, "page_end": 4, "type": "Text", "text": "The perspective of Modification is specifically designed for semantically modifying medical contents, so as to address the diverse requirements of editing beyond just locate them. However, constructing modification data triplets is", "source": "marker_v2", "marker_block_id": "/page/3/Text/5"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0030", "section": "3.2.2. MODIFICATION", "page_start": 4, "page_end": 4, "type": "Text", "text": "challenging because counterfactual image pairs cannot be captured simultaneously in the real world. While one could theoretically leverage general-purpose generative models (e.g., Nano Banana Pro or Qwen-Image-Edit) to produce these edits, such models are not specialized for the medical domain, and therefore are prone to severe hallucinations, which is unacceptable in a healthcare context. To construct rigorous edit triplets and preserve scalability, we propose a four-stage process (Fig. 3) designed to bridge the gap between task complexity and model competence so as to fully utilize these automatic tools.", "source": "marker_v2", "marker_block_id": "/page/3/Text/6"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0031", "section": "3.2.2. MODIFICATION", "page_start": 4, "page_end": 4, "type": "Text", "text": "Stage I: We develop a suite of modality-specific expert models for healthy tissue inpainting, built upon the FLUX.1-Fill-dev model. This strategy is based on the observation that generating healthy anatomical structures is more stable and predictable than generating lesions, as the former exhibits more tractable patterns and textures. For each modality, we curate a training dataset consisting exclusively of non-pathological samples from existing medical image repositories. Through parameter-efficient finetuning, these models learn to inpaint masked areas with high clinical accuracy. We further apply background restoration and edge blending to correct any unintended modifications made by the FLUX model outside the mask, ensuring the edited region blends seamlessly into the original image.", "source": "marker_v2", "marker_block_id": "/page/3/Text/7"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0032", "section": "3.2.2. MODIFICATION", "page_start": 4, "page_end": 4, "type": "Text", "text": "Stage II: We leverage these expert models to modify lesion-bearing images (L) into their counterfactual 'healthy' results (H). Specifically, we fill the lesion area", "source": "marker_v2", "marker_block_id": "/page/3/Text/8"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0033", "section": "3.2.2. MODIFICATION", "page_start": 5, "page_end": 5, "type": "Text", "text": "in L using white pixels based on its ground-truth segmentation label. This masked image and its corresponding binary mask are then processed by the modality-matched expert model to synthesize H, where healthy tissue replaces lesion. Compared to distilling general-purpose generative models, our modality-specific approach not only restricts the high-variance generative process to a localized region to guarantee background consistency during the edit, but also ensures that tasks remain within the model's learned distribution, thereby significantly reducing hallucinations. Furthermore, unlike manual data collection from the internet, our approach provides superior scalability and efficiency.", "source": "marker_v2", "marker_block_id": "/page/4/Text/1"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0034", "section": "3.2.2. MODIFICATION", "page_start": 5, "page_end": 5, "type": "Text", "text": "Stage III: We implement a rejection sampling mechanism for the generated 'healthy' images (H) to further enhance the data quality within the Modification tasks. For modalities that resemble natural images (e.g., endoscopy and dermoscopy), we prompt the Qwen3-VL-32B-Instruct model (Bai et al., 2025a) to filter out H that still contain lesions, exhibit artifacts, or are of low quality. For other modalities, we train separate nnUNet models (Isensee et al., 2021) for lesion segmentation and discard H where lesions remain detectable.", "source": "marker_v2", "marker_block_id": "/page/4/Text/2"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0035", "section": "3.2.2. MODIFICATION", "page_start": 5, "page_end": 5, "type": "Text", "text": "Stage IV: Triplet combination. Using these high quality 'lesion-healthy' counterfactual pairs, we generate diverse Modification task data by swapping L and H from niche of input and output and varying the textual prompts P.", "source": "marker_v2", "marker_block_id": "/page/4/Text/3"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0036", "section": "3.2.3. TRANSFORMATION", "page_start": 5, "page_end": 5, "type": "Text", "text": "Transformation tasks include a wide array of low-level medical image processing operations. Unlike the localized edits found in Perception and Modification categories, tasks in this category typically require a holistic transformation of the entire input image.", "source": "marker_v2", "marker_block_id": "/page/4/Text/5"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0037", "section": "3.2.3. TRANSFORMATION", "page_start": 5, "page_end": 5, "type": "Text", "text": "The rule-based construction pipeline of Transformation tasks is shown in Fig. 3. From public repositories, we compile medical image pairs (I and O) representing 17 distinct transformation targets under four typical low-level vision categories. We then design specialized textual prompts P for each task to unify diverse medical image processing functions into a consistent image editing framework.", "source": "marker_v2", "marker_block_id": "/page/4/Text/6"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0038", "section": "3.2.4. POST PROCESSING", "page_start": 5, "page_end": 5, "type": "Text", "text": "Prompt rephrasing. To enhance linguistic diversity, we utilize the Qwen-Max model to rephrase the prompts P for each data triplet. We also illustrate the linguistic diversity of our prompts via a word cloud in Figure 2( b).", "source": "marker_v2", "marker_block_id": "/page/4/Text/8"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0039", "section": "3.2.4. POST PROCESSING", "page_start": 5, "page_end": 5, "type": "Text", "text": "Benchmark curation. The training and test split of source dataset are strictly followed during the construction of MieDB-100k to preclude any data leakage. Furthermore, we recruit three people with clinical background to manually evaluate and curate 3, 485 of the most representative", "source": "marker_v2", "marker_block_id": "/page/4/Text/9"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0040", "section": "3.2.4. POST PROCESSING", "page_start": 5, "page_end": 5, "type": "Text", "text": "samples characterized by high clinical fidelity from raw data test split to serve as the benchmark of MieDB-100k, and we keep their original image size to minimize information loss.", "source": "marker_v2", "marker_block_id": "/page/4/Text/10"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0041", "section": "3.2.4. POST PROCESSING", "page_start": 5, "page_end": 5, "type": "Text", "text": "Train split construction. For train split, we establish three resolution bins (128, 256, and 512) and resize images to their nearest corresponding value. To check the fidelity of training split, we randomly select 6, 000 triplets for clinician evaluation, and over 95% are viewed as high quality.", "source": "marker_v2", "marker_block_id": "/page/4/Text/11"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0042", "section": "3.3. MieDB-100k Evaluation", "page_start": 5, "page_end": 5, "type": "Text", "text": "We evaluate MieDB-100k through two distinct approaches: (1) verifiable metrics for the Perception and Transformation tasks, amenable to reward design in prevailing reinforcement learning algorithms (Shao et al., 2024; Liu et al., 2025a) ; and (2) more subjective evaluations for the Modification tasks, reflecting their greater complexity.", "source": "marker_v2", "marker_block_id": "/page/4/Text/13"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0043", "section": "3.3.1. VERIFIABLE EVALUATION", "page_start": 5, "page_end": 5, "type": "Text", "text": "Localization Accuracy Metric. We use the DICE Score for evaluating the spatial overlapping performance in Perception tasks. Notably, reconstructing a binary mask from the colored regions of an edited image is mathematically feasible when the background image and overlay color are known, and we detail this process in App. D.1. This procedure is applied to both model's output O M and the ground truth images O to derive the mask of model's perceptual region and the ground truth region for DICE calculation.", "source": "marker_v2", "marker_block_id": "/page/4/Text/15"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0044", "section": "3.3.1. VERIFIABLE EVALUATION", "page_start": 5, "page_end": 5, "type": "Text", "text": "To differentiate between models that accurately identify specific medical targets and those that merely generate coarsegrained masks, we further propose Perception Accuracy. Under this metric, a result is considered successful only if the DICE score exceeds a threshold of τ = 0.8. This metric allows us to analyze whether a model possesses the specialized medical knowledge required for image understanding.", "source": "marker_v2", "marker_block_id": "/page/4/Text/16"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0045", "section": "3.3.1. VERIFIABLE EVALUATION", "page_start": 5, "page_end": 5, "type": "Text", "text": "Image Similarity Metrics. We utilize PSNR and SSIM (Wang et al., 2004) to evaluate the similarity between the ground-truth and edited images at both the pixel and structural levels. For evaluations within the Perception perspective, we mask out the pixels corresponding to the groundtruth segmentation in both images. This allows us to specifically assess the model's ability to preserve the background while performing the requested edit.", "source": "marker_v2", "marker_block_id": "/page/4/Text/17"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0046", "section": "3.3.2. EVALUATION FOR MODIFICATION TASKS", "page_start": 5, "page_end": 5, "type": "Text", "text": "Vision-Language Model Rubric Scoring. Automating reliable assessments in the Modification tasks is inherently challenging, as edits are defined semantically and cannot be evaluated via deterministic rules. Existing benchmarks often leverage Vision-Language Models (VLMs) for this purpose, and we standardize the process and mitigate potential critic hallucinations by implementing a rubric-based", "source": "marker_v2", "marker_block_id": "/page/4/Text/19"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0047", "section": "3.3.2. EVALUATION FOR MODIFICATION TASKS", "page_start": 6, "page_end": 6, "type": "TableGroup", "text": "Table 2. Overall result on MieDB-100k benchmark. P-ACC means Perception Accuracy; B-PSNR and B-SSIM mean only calculate PSNR and SSIM on background pixels respectively; Rubric-S stands for the Rubric Score from VLM and Pref-Rank stands for human preference ranking. Best values are marked in red while second bests are in blue. Perception Modification Trasnformation Model Name Size DICE P-ACC B-PSNR B-SSIM Rubric-S Pref-Rank PSNR SSIM Open-Source SDXL-turbo (Sauer et al., 2024) 3.5B 0.002 0.000 16.6 0.467 8.4 7.7 15.9 0.397 Bagel (Deng et al., 2025) 7B 0.263 0.069 13.9 0.620 34.4 6.2 12.7 0.442 OmniGen2 (Wu et al., 2025b) 7B 0.248 0.065 11.9 0.541 29.1 7.1 8.3 0.280 Step1X-Edit (Liu et al., 2025c) 21B 0.332 0.126 15.5 0.727 35.6 4.5 16.6 0.539 Qwen-Image-Edit (Wu et al., 2025a) 27B 0.387 0.153 15.4 0.722 32.2 5.5 18.9 0.606 FLUX.1-Kontext-dev (Labs et al., 2025) 12B 0.341 0.126 15.4 0.701 37.8 6.2 17.9 0.543 OmniGen2-MIE (Ours) 7B 0.831 0.737 28.1 0.917 65.9 1.4 22.6 0.685 Closed-Source GPT-Image-1 (OpenAI, 2025) 0.467 0.221 16.3 0.510 42.8 4.8 14.4 0.451 Nano Banana Pro (DeepMind, 2025a) 0.426 0.202 12.8 0.413 63.4 2.0 20.0 0.610 Imagen4 (DeepMind, 2025b) 0.142 0.000 8.9 0.210 19.7 7.4 7.9 0.174", "source": "marker_v2", "marker_block_id": "/page/5/TableGroup/625"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0048", "section": "3.3.2. EVALUATION FOR MODIFICATION TASKS", "page_start": 6, "page_end": 6, "type": "Text", "text": "scoring system. Specifically, we provide the VLM with the input image I, edit instruction P, reference output O, and the model's generated result OM. Guided by the rubric, the VLM then performs a holistic evaluation of OM.", "source": "marker_v2", "marker_block_id": "/page/5/Text/9"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0049", "section": "3.3.2. EVALUATION FOR MODIFICATION TASKS", "page_start": 6, "page_end": 6, "type": "Text", "text": "We design a comprehensive scoring rubric (App. D.2) that assesses both the fulfillment of the editing intent and the model's ability to preserve the background. We utilize GPT-5.2 as an automated evaluator for this process, and map the final score to [0, 100].", "source": "marker_v2", "marker_block_id": "/page/5/Text/10"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0050", "section": "3.3.2. EVALUATION FOR MODIFICATION TASKS", "page_start": 6, "page_end": 6, "type": "Text", "text": "Human Preference Ranking. For each test case, we present the original triplet (I, P, O) and the outputs of all tested models simultaneously to evaluators, who are then asked to rank the various model-generated results according to their preference. By forcing this comparative ordering of all models, we are able to move beyond absolute quality scores and capture the relative strengths and weaknesses of current generative frameworks in a clinical setting. Specifically, we recruit 3 evaluators with clinical backgrounds to assess and rank the images edited by the benchmarked models, and compute the average ranking.", "source": "marker_v2", "marker_block_id": "/page/5/Text/11"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0051", "section": "4.1. Baselines", "page_start": 6, "page_end": 6, "type": "Text", "text": "We evaluate nine models on MieDB-100k, comprising six open-source models: Qwen-Image-Edit-2511 (Wu et al., 2025a) , Bagel (Deng et al., 2025) , OmniGen2 (Wu et al., 2025b) , Step1X-Edit-v1p2 (Liu et al., 2025c) , FLUX.1- Kontext-dev (Labs et al., 2025) and SDXL-turbo (Sauer et al., 2024) , plus three closed-source models: Nano Banana Pro (DeepMind, 2025a) , GPT-Image-1 (OpenAI, 2025) , and Imagen4 (DeepMind, 2025b) . We implement open-source models following their official inference settings.", "source": "marker_v2", "marker_block_id": "/page/5/Text/14"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0052", "section": "4.1. Baselines", "page_start": 6, "page_end": 6, "type": "Text", "text": "To validate the effectiveness of MieDB-100k, we finetune", "source": "marker_v2", "marker_block_id": "/page/5/Text/15"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0053", "section": "4.1. Baselines", "page_start": 6, "page_end": 6, "type": "Text", "text": "the OmniGen2 baseline on the training split and subject it to the same evaluation protocol as the other models. Specifically, we train the Diffusion Transformer (DiT) component for 20,000 iterations, employing a global batch size of 64 and a learning rate of 1e-4.", "source": "marker_v2", "marker_block_id": "/page/5/Text/16"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0054", "section": "4.2. Quantitative Results", "page_start": 6, "page_end": 6, "type": "Text", "text": "We report the benchmarking results of MieDB-100k in Tab. 2. First, the extremely low perception accuracy indicates that all tested models except ours fail to accurately comprehend and localize the specified anatomical targets under our evaluation protocol. Consequently, in Modification tasks, most of them are unable to generate clinically meaningful edits. Although a few models, such as Nano Banana Pro, achieve competitive results, we are indeed observing the 'right-for-the-wrong-reason' phenomenon, a risk that must be strictly avoided in clinical settings. Since the poor performance in Perception tasks expose their intrinsic lack of necessary medical knowledge, their edits cannot be justified. Notably in Transformation tasks, Nano Banana Pro also presents competitive results in certain cases. This may be attributed to the similarity between tasks like denoising or artifact removal and general-purpose low-level vision tasks, for which the model already possesses some capability (Zuo et al., 2025) . Alternatively, it is possible that similar medical image processing tasks were included in its training set. Regardless, its absolute performance remains insufficient for practical clinical deployment. In summary, the benchmark result demonstrates that current multimodal generative model cannot meet the requirement of medical imaging editing.", "source": "marker_v2", "marker_block_id": "/page/5/Text/18"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0055", "section": "4.2. Quantitative Results", "page_start": 6, "page_end": 6, "type": "Text", "text": "Conversely, after training on MieDB-100k, a standard baseline model can achieve superior medical editing capabilities. As shown in Tab. 2, the OmniGen2-MIE model delivers the best performance across all three edit-", "source": "marker_v2", "marker_block_id": "/page/5/Text/19"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0056", "section": "4.2. Quantitative Results", "page_start": 7, "page_end": 7, "type": "FigureGroup", "text": "[Edit Instruction] Process the T1 MRI to mitigate motion effects without changing the appearance of brain tissue, tumors, or other clinical features.", "source": "marker_v2", "marker_block_id": "/page/6/FigureGroup/606"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0057", "section": "4.2. Quantitative Results", "page_start": 7, "page_end": 7, "type": "Caption", "text": "Figure 4. Qualitative editing result comparison.", "source": "marker_v2", "marker_block_id": "/page/6/Caption/3"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0058", "section": "4.2. Quantitative Results", "page_start": 7, "page_end": 7, "type": "Text", "text": "ing perspectives. The most significant improvements are observed in the Perception perspective, which demonstrate that MieDB-100k can effectively inject essential medical knowledge, thereby enhancing the interpretability of downstream editing tasks. Furthermore, in the Modification and Transformation tasks, where general-purpose editing abilities transfer more readily, our enhanced model still yields superior editing results compared to Nano Banana Pro, the SOTA multi-modal generative models. These findings highlight the pivotal role of our dataset in domain adaptation and", "source": "marker_v2", "marker_block_id": "/page/6/Text/4"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0059", "section": "4.2. Quantitative Results", "page_start": 7, "page_end": 7, "type": "Text", "text": "establish a foundation for the development of understandinggeneration unified medical models.", "source": "marker_v2", "marker_block_id": "/page/6/Text/5"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0060", "section": "4.3. Qualitative Results", "page_start": 7, "page_end": 7, "type": "Text", "text": "Fig. 4 presents qualitative editing results for several baseline models across the diverse modalities and tasks in MieDB-100k. These results demonstrate that the finetuned model exhibits an enhanced capability in both understanding and generation, allowing it to navigate the inherent complexities", "source": "marker_v2", "marker_block_id": "/page/6/Text/7"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0061", "section": "4.3. Qualitative Results", "page_start": 8, "page_end": 8, "type": "TableGroup", "text": "Table 3. Ablation study result on MieDB-100k. P stands for Perception, M stands for Modification, and T stands for Transformation. Best values are marked in red, second bests are in blue. Perce ption Modification Trasnformation Training Data DICE ACC RubricScore PSNR SSIM Baseline (No train) 0.248 0.065 29.1 8.3 0.280 P-only 0.833 0.740 37.8 19.7 0.631 M-only 0.001 0.000 57.5 19.8 0.631 T-only 0.034 0.000 15.0 23.7 0.702 MieDB-100k 0.831 0.737 65.9 22.6 0.685", "source": "marker_v2", "marker_block_id": "/page/7/TableGroup/316"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0062", "section": "4.3. Qualitative Results", "page_start": 8, "page_end": 8, "type": "Text", "text": "of medical image editing. Moreover, despite being explicitly prompted, even sophisticated closed-source models such as Nano Banana Pro fail to maintain background consistency in certain tasks. While their instruction-following proficiency stems from large-scale pre-training on natural image pairs, these capabilities tend to degrade when the distribution of medical modalities deviates significantly from the natural images seen during pre-training. To further study the impact of modality deviation, we conduct a modality-wise analysis in App. E.1, and the results prove our judgment. This observation underscores the necessity of a highly diverse dataset like MieDB-100k to equip models with the capacity to handle a vast range of medical imaging modalities.", "source": "marker_v2", "marker_block_id": "/page/7/Text/5"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0063", "section": "4.4. Ablation Study", "page_start": 8, "page_end": 8, "type": "Text", "text": "To investigate the contribution of each task category, we conduct an ablation study by training models on individual perspective of MieDB-100k. We again utilize OmniGen2 as baseline model, following the training recipe described above while varying only the training data. As shown in Tab 3, each specialized model significantly outperforms the original baseline in its respective domain, validating the high information density and clinical relevance of our data. For the model trained on the full dataset, it achieves comparable or even better performance on all three perspectives, showing the effectiveness of the joint training. More importantly, we observe significant performance improvement in the Modification perspective, demonstrating visual understanding ability has the potential to enhance visual generation ability. In summary, the ablation study shows that MieDB-100k can provide a synergistic training signal, enabling the development of a versatile model capable of handling diverse medical editing tasks simultaneously.", "source": "marker_v2", "marker_block_id": "/page/7/Text/7"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0064", "section": "4.5. Generalization Test", "page_start": 8, "page_end": 8, "type": "Text", "text": "To further investigate the cross-task synergy and the resulting generalization capabilities, we conduct an out-ofdistribution (OOD) editing experiment. Specifically, we target 'bone metastasis', a medical target included in Perception tasks but strictly excluded from the Modification training data. We then prompt the OmniGen2-MIE model", "source": "marker_v2", "marker_block_id": "/page/7/Text/9"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0065", "section": "4.5. Generalization Test", "page_start": 8, "page_end": 8, "type": "FigureGroup", "text": "Figure 5. Generalization test assessment. (a) and (b): Edit samples output by different models on bone metastasis addition (a) and removal (b) tasks. Red bounding boxes are added post-hoc to highlight the edited regions for visualization; (c): Quantitative assessments following the recipe of Modification task evaluation.", "source": "marker_v2", "marker_block_id": "/page/7/FigureGroup/317"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0066", "section": "4.5. Generalization Test", "page_start": 8, "page_end": 8, "type": "Text", "text": "to perform metastasis addition and removal in CT scans.", "source": "marker_v2", "marker_block_id": "/page/7/Text/12"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0067", "section": "4.5. Generalization Test", "page_start": 8, "page_end": 8, "type": "Text", "text": "As shown in Fig. 5, OmniGen2-MIE significantly outperforms OmniGen2 on this unseen task, demonstrating that our unified training on MieDB-100k can enhance the model's generalization capabilities across editing tasks. We also observe that Nano Banana Pro achieves the best OOD editing performance, marginally surpassing OmniGen2-MIE. We attribute this performance to the utilization of massive-scale general and medical editing data, which further underscores the necessity of scaling up medical editing data.", "source": "marker_v2", "marker_block_id": "/page/7/Text/13"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0068", "section": "5. Conclusion", "page_start": 8, "page_end": 8, "type": "Text", "text": "In this paper, we introduce MieDB-100k, a large-scale and diverse dataset for text-guided medical image editing. By unifying Perception, Modification, and Transformation tasks into the paradigm of editing, our dataset bridges the gap between medical image understanding and generation. We develop a robust curation pipeline, integrating modalityspecific expert models with rule-based synthesis, and enforce rigorous manual quality control to ensure clinical fidelity across all data. Extensive benchmarking demonstrates that model trained on MieDB-100k consistently outperform both SOTA open-source and proprietary multimodal models while exhibiting exceptional generalization to unseen clinical tasks. Our work thus provides the data foundation to support the development and evaluation of multimodal generative models for clinical applications.", "source": "marker_v2", "marker_block_id": "/page/7/Text/15"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0069", "section": "Impact Statement", "page_start": 9, "page_end": 9, "type": "Text", "text": "This paper presents work whose goal is to advance the field of Machine Learning. There are many potential societal consequences of our work, none which we feel must be specifically highlighted here.", "source": "marker_v2", "marker_block_id": "/page/8/Text/2"}

icml26/80c20b7b-ead6-454a-849e-56702a6c828f/marker_meta.json

@@ -0,0 +1,2520 @@
+{
+  "table_of_contents": [
+    {
+      "title": "MieDB-100k: A Comprehensive Dataset for Medical Image Editing",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          92.65999603271484,
+          88.9453125
+        ],
+        [
+          504.2238464355469,
+          88.9453125
+        ],
+        [
+          504.2238464355469,
+          104.2581787109375
+        ],
+        [
+          92.65999603271484,
+          104.2581787109375
+        ]
+      ]
+    },
+    {
+      "title": "Anonymous Authors1",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          242.00900268554688,
+          141.15234375
+        ],
+        [
+          333.4921875,
+          141.15234375
+        ],
+        [
+          333.4921875,
+          152.96612548828125
+        ],
+        [
+          242.00900268554688,
+          152.96612548828125
+        ]
+      ]
+    },
+    {
+      "title": "Abstract",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          148.81640625,
+          175.18359375
+        ],
+        [
+          194.68328857421875,
+          175.18359375
+        ],
+        [
+          194.68328857421875,
+          188.015380859375
+        ],
+        [
+          148.81640625,
+          188.015380859375
+        ]
+      ]
+    },
+    {
+      "title": "1. Introduction",
+      "heading_level": null,
+      "page_id": 0,
+      "polygon": [
+        [
+          54.685546875,
+          521.68359375
+        ],
+        [
+          132.2760772705078,
+          521.68359375
+        ],
+        [
+          132.2760772705078,
+          534.0033569335938
+        ],
+        [
+          54.685546875,
+          534.0033569335938
+        ]
+      ]
+    },
+    {
+      "title": "2. Related Work",
+      "heading_level": null,
+      "page_id": 1,
+      "polygon": [
+        [
+          304.8046875,
+          380.91796875
+        ],
+        [
+          391.5,
+          380.91796875
+        ],
+        [
+          391.5,
+          390.75
+        ],
+        [
+          304.8046875,
+          390.75
+        ]
+      ]
+    },
+    {
+      "title": "2.1. Data Research for Medical Image Editing",
+      "heading_level": null,
+      "page_id": 1,
+      "polygon": [
+        [
+          305.40234375,
+          402.0
+        ],
+        [
+          503.25,
+          402.0
+        ],
+        [
+          503.25,
+          412.62890625
+        ],
+        [
+          305.40234375,
+          412.62890625
+        ]
+      ]
+    },
+    {
+      "title": "2.2. Multimodal Generative Model",
+      "heading_level": null,
+      "page_id": 2,
+      "polygon": [
+        [
+          52.59375,
+          340.5
+        ],
+        [
+          203.25,
+          340.5
+        ],
+        [
+          203.25,
+          349.5
+        ],
+        [
+          52.59375,
+          349.5
+        ]
+      ]
+    },
+    {
+      "title": "3. MieDB-100k",
+      "heading_level": null,
+      "page_id": 2,
+      "polygon": [
+        [
+          52.892578125,
+          663.0
+        ],
+        [
+          132.75,
+          663.0
+        ],
+        [
+          132.75,
+          672.50390625
+        ],
+        [
+          52.892578125,
+          672.50390625
+        ]
+      ]
+    },
+    {
+      "title": "3.1. Data Definition",
+      "heading_level": null,
+      "page_id": 2,
+      "polygon": [
+        [
+          306.0,
+          220.5
+        ],
+        [
+          391.5,
+          220.5
+        ],
+        [
+          391.5,
+          229.5
+        ],
+        [
+          306.0,
+          229.5
+        ]
+      ]
+    },
+    {
+      "title": "3.2. Three Perspectives of MieDB-100k",
+      "heading_level": null,
+      "page_id": 2,
+      "polygon": [
+        [
+          306.0,
+          288.0
+        ],
+        [
+          473.25,
+          288.0
+        ],
+        [
+          473.25,
+          298.546875
+        ],
+        [
+          306.0,
+          298.546875
+        ]
+      ]
+    },
+    {
+      "title": "3.2.1. Perception",
+      "heading_level": null,
+      "page_id": 2,
+      "polygon": [
+        [
+          306.0,
+          472.18359375
+        ],
+        [
+          390.0,
+          472.18359375
+        ],
+        [
+          390.0,
+          483.75
+        ],
+        [
+          306.0,
+          483.75
+        ]
+      ]
+    },
+    {
+      "title": "3.2.2. MODIFICATION",
+      "heading_level": null,
+      "page_id": 3,
+      "polygon": [
+        [
+          53.19140625,
+          651.62109375
+        ],
+        [
+          147.75,
+          651.62109375
+        ],
+        [
+          147.75,
+          660.75
+        ],
+        [
+          53.19140625,
+          660.75
+        ]
+      ]
+    },
+    {
+      "title": "3.2.3. TRANSFORMATION",
+      "heading_level": null,
+      "page_id": 4,
+      "polygon": [
+        [
+          54.087890625,
+          403.99542236328125
+        ],
+        [
+          163.81727600097656,
+          403.99542236328125
+        ],
+        [
+          163.81727600097656,
+          413.9580383300781
+        ],
+        [
+          54.087890625,
+          413.9580383300781
+        ]
+      ]
+    },
+    {
+      "title": "3.2.4. POST PROCESSING",
+      "heading_level": null,
+      "page_id": 4,
+      "polygon": [
+        [
+          54.38671875,
+          583.55859375
+        ],
+        [
+          162.5739288330078,
+          583.55859375
+        ],
+        [
+          162.5739288330078,
+          593.7870788574219
+        ],
+        [
+          54.38671875,
+          593.7870788574219
+        ]
+      ]
+    },
+    {
+      "title": "3.3. MieDB-100k Evaluation",
+      "heading_level": null,
+      "page_id": 4,
+      "polygon": [
+        [
+          304.8046875,
+          183.3046875
+        ],
+        [
+          428.00738525390625,
+          183.3046875
+        ],
+        [
+          428.00738525390625,
+          194.1361083984375
+        ],
+        [
+          304.8046875,
+          194.1361083984375
+        ]
+      ]
+    },
+    {
+      "title": "3.3.1. VERIFIABLE EVALUATION",
+      "heading_level": null,
+      "page_id": 4,
+      "polygon": [
+        [
+          306.0,
+          286.38543701171875
+        ],
+        [
+          445.5876159667969,
+          286.38543701171875
+        ],
+        [
+          445.5876159667969,
+          296.3480529785156
+        ],
+        [
+          306.0,
+          296.3480529785156
+        ]
+      ]
+    },
+    {
+      "title": "3.3.2. EVALUATION FOR MODIFICATION TASKS",
+      "heading_level": null,
+      "page_id": 4,
+      "polygon": [
+        [
+          306.59765625,
+          614.49609375
+        ],
+        [
+          508.0213317871094,
+          614.49609375
+        ],
+        [
+          508.0213317871094,
+          625.6180725097656
+        ],
+        [
+          306.59765625,
+          625.6180725097656
+        ]
+      ]
+    },
+    {
+      "title": "4. Experiments",
+      "heading_level": null,
+      "page_id": 5,
+      "polygon": [
+        [
+          53.490234375,
+          546.0371704101562
+        ],
+        [
+          132.47930908203125,
+          546.0371704101562
+        ],
+        [
+          132.47930908203125,
+          557.9923706054688
+        ],
+        [
+          53.490234375,
+          557.9923706054688
+        ]
+      ]
+    },
+    {
+      "title": "4.1. Baselines",
+      "heading_level": null,
+      "page_id": 5,
+      "polygon": [
+        [
+          53.490234375,
+          564.99609375
+        ],
+        [
+          112.17700958251953,
+          564.99609375
+        ],
+        [
+          112.17700958251953,
+          576.9021148681641
+        ],
+        [
+          53.490234375,
+          576.9021148681641
+        ]
+      ]
+    },
+    {
+      "title": "4.2. Quantitative Results",
+      "heading_level": null,
+      "page_id": 5,
+      "polygon": [
+        [
+          306.0,
+          351.52734375
+        ],
+        [
+          411.8480529785156,
+          351.52734375
+        ],
+        [
+          411.8480529785156,
+          362.8150939941406
+        ],
+        [
+          306.0,
+          362.8150939941406
+        ]
+      ]
+    },
+    {
+      "title": "4.3. Qualitative Results",
+      "heading_level": null,
+      "page_id": 6,
+      "polygon": [
+        [
+          307.1953125,
+          634.8195190429688
+        ],
+        [
+          405.7608947753906,
+          634.8195190429688
+        ],
+        [
+          405.7608947753906,
+          644.7821197509766
+        ],
+        [
+          307.1953125,
+          644.7821197509766
+        ]
+      ]
+    },
+    {
+      "title": "4.4. Ablation Study",
+      "heading_level": null,
+      "page_id": 7,
+      "polygon": [
+        [
+          54.0,
+          380.91796875
+        ],
+        [
+          138.0,
+          380.91796875
+        ],
+        [
+          138.0,
+          390.0
+        ],
+        [
+          54.0,
+          390.0
+        ]
+      ]
+    },
+    {
+      "title": "4.5. Generalization Test",
+      "heading_level": null,
+      "page_id": 7,
+      "polygon": [
+        [
+          51.697265625,
+          626.87109375
+        ],
+        [
+          156.0,
+          626.87109375
+        ],
+        [
+          156.0,
+          637.5
+        ],
+        [
+          51.697265625,
+          637.5
+        ]
+      ]
+    },
+    {
+      "title": "5. Conclusion",
+      "heading_level": null,
+      "page_id": 7,
+      "polygon": [
+        [
+          306.0,
+          518.58984375
+        ],
+        [
+          377.25,
+          518.58984375
+        ],
+        [
+          377.25,
+          528.0
+        ],
+        [
+          306.0,
+          528.0
+        ]
+      ]
+    },
+    {
+      "title": "Impact Statement",
+      "heading_level": null,
+      "page_id": 8,
+      "polygon": [
+        [
+          51.99609375,
+          67.84710693359375
+        ],
+        [
+          146.74185180664062,
+          67.84710693359375
+        ],
+        [
+          146.74185180664062,
+          79.80230712890625
+        ],
+        [
+          51.99609375,
+          79.80230712890625
+        ]
+      ]
+    },
+    {
+      "title": "References",
+      "heading_level": null,
+      "page_id": 8,
+      "polygon": [
+        [
+          52.294921875,
+          151.651123046875
+        ],
+        [
+          110.9838638305664,
+          151.651123046875
+        ],
+        [
+          110.9838638305664,
+          163.6063232421875
+        ],
+        [
+          52.294921875,
+          163.6063232421875
+        ]
+      ]
+    },
+    {
+      "title": "A. Data Sources",
+      "heading_level": null,
+      "page_id": 13,
+      "polygon": [
+        [
+          53.7890625,
+          67.84710693359375
+        ],
+        [
+          137.2374725341797,
+          67.84710693359375
+        ],
+        [
+          137.2374725341797,
+          79.80230712890625
+        ],
+        [
+          53.7890625,
+          79.80230712890625
+        ]
+      ]
+    },
+    {
+      "title": "B. Construction Details",
+      "heading_level": null,
+      "page_id": 14,
+      "polygon": [
+        [
+          53.490234375,
+          67.2890625
+        ],
+        [
+          174.65725708007812,
+          67.2890625
+        ],
+        [
+          174.65725708007812,
+          79.80230712890625
+        ],
+        [
+          53.490234375,
+          79.80230712890625
+        ]
+      ]
+    },
+    {
+      "title": "C. Implementation Details of OmniGen2-MIE",
+      "heading_level": null,
+      "page_id": 14,
+      "polygon": [
+        [
+          54.38671875,
+          330.2578125
+        ],
+        [
+          290.2162170410156,
+          330.2578125
+        ],
+        [
+          290.2162170410156,
+          342.370361328125
+        ],
+        [
+          54.38671875,
+          342.370361328125
+        ]
+      ]
+    },
+    {
+      "title": "D. Evaluation Details",
+      "heading_level": null,
+      "page_id": 15,
+      "polygon": [
+        [
+          52.892578125,
+          67.67578125
+        ],
+        [
+          163.68238830566406,
+          67.67578125
+        ],
+        [
+          163.68238830566406,
+          79.80230712890625
+        ],
+        [
+          52.892578125,
+          79.80230712890625
+        ]
+      ]
+    },
+    {
+      "title": "D.1. Mask Reconstruction via Alpha De-blending",
+      "heading_level": null,
+      "page_id": 15,
+      "polygon": [
+        [
+          53.19140625,
+          88.74945068359375
+        ],
+        [
+          264.1863098144531,
+          88.74945068359375
+        ],
+        [
+          264.1863098144531,
+          99.0
+        ],
+        [
+          53.19140625,
+          99.0
+        ]
+      ]
+    },
+    {
+      "title": "D.1.1. MATHEMATICS",
+      "heading_level": null,
+      "page_id": 15,
+      "polygon": [
+        [
+          53.19140625,
+          106.734375
+        ],
+        [
+          149.16822814941406,
+          106.734375
+        ],
+        [
+          149.16822814941406,
+          117.45599365234375
+        ],
+        [
+          53.19140625,
+          117.45599365234375
+        ]
+      ]
+    },
+    {
+      "title": "D.1.2. CASE OF MASK RECONSTRUCTION",
+      "heading_level": null,
+      "page_id": 15,
+      "polygon": [
+        [
+          54.38671875,
+          345.7265625
+        ],
+        [
+          230.9055633544922,
+          345.7265625
+        ],
+        [
+          230.9055633544922,
+          355.7240905761719
+        ],
+        [
+          54.38671875,
+          355.7240905761719
+        ]
+      ]
+    },
+    {
+      "title": "D.2. VLM Automatic Scoring",
+      "heading_level": null,
+      "page_id": 15,
+      "polygon": [
+        [
+          54.087890625,
+          499.1495361328125
+        ],
+        [
+          180.09201049804688,
+          499.1495361328125
+        ],
+        [
+          180.09201049804688,
+          509.11212158203125
+        ],
+        [
+          54.087890625,
+          509.11212158203125
+        ]
+      ]
+    },
+    {
+      "title": "D.2.1. VLM SCORING RUBRIC",
+      "heading_level": null,
+      "page_id": 15,
+      "polygon": [
+        [
+          54.087890625,
+          517.8934936523438
+        ],
+        [
+          188.560546875,
+          517.8934936523438
+        ],
+        [
+          188.560546875,
+          527.8561096191406
+        ],
+        [
+          54.087890625,
+          527.8561096191406
+        ]
+      ]
+    },
+    {
+      "title": "Scoring Rubric for Modification Tasks",
+      "heading_level": null,
+      "page_id": 15,
+      "polygon": [
+        [
+          71.12109375,
+          535.6611633300781
+        ],
+        [
+          269.59100341796875,
+          535.6611633300781
+        ],
+        [
+          269.59100341796875,
+          547.6163635253906
+        ],
+        [
+          71.12109375,
+          547.6163635253906
+        ]
+      ]
+    },
+    {
+      "title": "1) Edit Goal Fulfillment (Edit Correctness): Assesses whether the intended lesion change is achieved.",
+      "heading_level": null,
+      "page_id": 15,
+      "polygon": [
+        [
+          73.71700286865234,
+          631.2895202636719
+        ],
+        [
+          492.46875,
+          631.2895202636719
+        ],
+        [
+          492.46875,
+          641.56640625
+        ],
+        [
+          73.71700286865234,
+          641.56640625
+        ]
+      ]
+    },
+    {
+      "title": "Scoring reference:",
+      "heading_level": null,
+      "page_id": 16,
+      "polygon": [
+        [
+          86.0625,
+          144.6328125
+        ],
+        [
+          160.30181884765625,
+          144.6328125
+        ],
+        [
+          160.30181884765625,
+          155.3360595703125
+        ],
+        [
+          86.0625,
+          155.3360595703125
+        ]
+      ]
+    },
+    {
+      "title": "Scoring reference:",
+      "heading_level": null,
+      "page_id": 16,
+      "polygon": [
+        [
+          85.46484375,
+          262.93243408203125
+        ],
+        [
+          160.30181884765625,
+          262.93243408203125
+        ],
+        [
+          160.30181884765625,
+          272.89501953125
+        ],
+        [
+          85.46484375,
+          272.89501953125
+        ]
+      ]
+    },
+    {
+      "title": "Scoring reference:",
+      "heading_level": null,
+      "page_id": 16,
+      "polygon": [
+        [
+          84.568359375,
+          380.4914855957031
+        ],
+        [
+          160.30181884765625,
+          380.4914855957031
+        ],
+        [
+          160.30181884765625,
+          390.4541015625
+        ],
+        [
+          84.568359375,
+          390.4541015625
+        ]
+      ]
+    },
+    {
+      "title": "Return Format: Return a JSON dictionary with two fields:",
+      "heading_level": null,
+      "page_id": 17,
+      "polygon": [
+        [
+          72.9140625,
+          254.84765625
+        ],
+        [
+          312.34112548828125,
+          254.84765625
+        ],
+        [
+          312.34112548828125,
+          264.92510986328125
+        ],
+        [
+          72.9140625,
+          264.92510986328125
+        ]
+      ]
+    },
+    {
+      "title": "D.2.2. CASE",
+      "heading_level": null,
+      "page_id": 17,
+      "polygon": [
+        [
+          54.087890625,
+          350.35748291015625
+        ],
+        [
+          108.9789810180664,
+          350.35748291015625
+        ],
+        [
+          108.9789810180664,
+          360.3200988769531
+        ],
+        [
+          54.087890625,
+          360.3200988769531
+        ]
+      ]
+    },
+    {
+      "title": "",
+      "heading_level": null,
+      "page_id": 17,
+      "polygon": [
+        [
+          58.271484375,
+          376.6640625
+        ],
+        [
+          101.900390625,
+          376.6640625
+        ],
+        [
+          101.900390625,
+          384.3984375
+        ],
+        [
+          58.271484375,
+          384.3984375
+        ]
+      ]
+    },
+    {
+      "title": "Input collage VLM Output",
+      "heading_level": null,
+      "page_id": 17,
+      "polygon": [
+        [
+          60.417598724365234,
+          374.79254150390625
+        ],
+        [
+          308.98828125,
+          374.79254150390625
+        ],
+        [
+          308.98828125,
+          400.25390625
+        ],
+        [
+          60.417598724365234,
+          400.25390625
+        ]
+      ]
+    },
+    {
+      "title": "",
+      "heading_level": null,
+      "page_id": 17,
+      "polygon": [
+        [
+          61.259765625,
+          453.62109375
+        ],
+        [
+          118.037109375,
+          453.62109375
+        ],
+        [
+          118.037109375,
+          462.12890625
+        ],
+        [
+          61.259765625,
+          462.12890625
+        ]
+      ]
+    },
+    {
+      "title": "",
+      "heading_level": null,
+      "page_id": 17,
+      "polygon": [
+        [
+          261.17578125,
+          453.62109375
+        ],
+        [
+          303.01171875,
+          453.62109375
+        ],
+        [
+          303.01171875,
+          462.12890625
+        ],
+        [
+          261.17578125,
+          462.12890625
+        ]
+      ]
+    },
+    {
+      "title": "E. Supplementary Experiments",
+      "heading_level": null,
+      "page_id": 18,
+      "polygon": [
+        [
+          53.19140625,
+          67.5
+        ],
+        [
+          216.75,
+          67.5
+        ],
+        [
+          216.75,
+          79.27734375
+        ],
+        [
+          53.19140625,
+          79.27734375
+        ]
+      ]
+    },
+    {
+      "title": "E.1. Modality-Wise Performance Analysis",
+      "heading_level": null,
+      "page_id": 18,
+      "polygon": [
+        [
+          52.59375,
+          87.75
+        ],
+        [
+          234.0,
+          87.75
+        ],
+        [
+          234.0,
+          98.25
+        ],
+        [
+          52.59375,
+          98.25
+        ]
+      ]
+    },
+    {
+      "title": "E.2. Multi-Round Generation",
+      "heading_level": null,
+      "page_id": 18,
+      "polygon": [
+        [
+          54.0,
+          456.71484375
+        ],
+        [
+          181.5,
+          456.71484375
+        ],
+        [
+          181.5,
+          465.75
+        ],
+        [
+          54.0,
+          466.5
+        ]
+      ]
+    },
+    {
+      "title": "E.3. Out-Of-Distribution Image Edit",
+      "heading_level": null,
+      "page_id": 19,
+      "polygon": [
+        [
+          52.59375,
+          105.9609375
+        ],
+        [
+          210.49790954589844,
+          105.9609375
+        ],
+        [
+          210.49790954589844,
+          116.4281005859375
+        ],
+        [
+          52.59375,
+          116.4281005859375
+        ]
+      ]
+    },
+    {
+      "title": "F. Data Gallery and More Qualitative Result",
+      "heading_level": null,
+      "page_id": 19,
+      "polygon": [
+        [
+          53.7890625,
+          405.0331726074219
+        ],
+        [
+          281.79974365234375,
+          405.0331726074219
+        ],
+        [
+          281.79974365234375,
+          416.9883728027344
+        ],
+        [
+          53.7890625,
+          416.9883728027344
+        ]
+      ]
+    },
+    {
+      "title": "F.1. Examples of Healthy Tissue Inpainting",
+      "heading_level": null,
+      "page_id": 19,
+      "polygon": [
+        [
+          52.59375,
+          425.77734375
+        ],
+        [
+          238.07435607910156,
+          425.77734375
+        ],
+        [
+          238.07435607910156,
+          435.8981018066406
+        ],
+        [
+          52.59375,
+          435.8981018066406
+        ]
+      ]
+    },
+    {
+      "title": "F.2. Extended Examples of Qualitative result",
+      "heading_level": null,
+      "page_id": 19,
+      "polygon": [
+        [
+          53.7890625,
+          666.3875122070312
+        ],
+        [
+          244.63975524902344,
+          666.3875122070312
+        ],
+        [
+          244.63975524902344,
+          676.3501129150391
+        ],
+        [
+          53.7890625,
+          676.3501129150391
+        ]
+      ]
+    },
+    {
+      "title": "G. Failure Cases",
+      "heading_level": null,
+      "page_id": 23,
+      "polygon": [
+        [
+          53.19140625,
+          67.67578125
+        ],
+        [
+          139.6046142578125,
+          67.67578125
+        ],
+        [
+          139.6046142578125,
+          79.80230712890625
+        ],
+        [
+          53.19140625,
+          79.80230712890625
+        ]
+      ]
+    },
+    {
+      "title": "H. Limitations",
+      "heading_level": null,
+      "page_id": 23,
+      "polygon": [
+        [
+          54.087890625,
+          410.30859375
+        ],
+        [
+          129.83721923828125,
+          410.30859375
+        ],
+        [
+          129.83721923828125,
+          424.27935791015625
+        ],
+        [
+          54.087890625,
+          424.27935791015625
+        ]
+      ]
+    }
+  ],
+  "page_stats": [
+    {
+      "page_id": 0,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          337
+        ],
+        [
+          "Line",
+          142
+        ],
+        [
+          "Text",
+          7
+        ],
+        [
+          "SectionHeader",
+          4
+        ],
+        [
+          "ListItem",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Footnote",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "ListGroup",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 1,
+      "text_extraction_method": "surya",
+      "block_counts": [
+        [
+          "Line",
+          183
+        ],
+        [
+          "Span",
+          92
+        ],
+        [
+          "Text",
+          41
+        ],
+        [
+          "ListItem",
+          2
+        ],
+        [
+          "SectionHeader",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Figure",
+          1
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "FigureGroup",
+          1
+        ],
+        [
+          "ListGroup",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 2,
+      "text_extraction_method": "surya",
+      "block_counts": [
+        [
+          "Line",
+          203
+        ],
+        [
+          "Span",
+          123
+        ],
+        [
+          "Text",
+          51
+        ],
+        [
+          "TableCell",
+          35
+        ],
+        [
+          "SectionHeader",
+          5
+        ],
+        [
+          "Caption",
+          2
+        ],
+        [
+          "Reference",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Table",
+          1
+        ],
+        [
+          "Figure",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "TableGroup",
+          1
+        ],
+        [
+          "FigureGroup",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 3,
+      "text_extraction_method": "surya",
+      "block_counts": [
+        [
+          "Line",
+          201
+        ],
+        [
+          "Span",
+          82
+        ],
+        [
+          "Text",
+          46
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Figure",
+          1
+        ],
+        [
+          "SectionHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 4,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          379
+        ],
+        [
+          "Line",
+          135
+        ],
+        [
+          "Text",
+          27
+        ],
+        [
+          "SectionHeader",
+          5
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 5,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          473
+        ],
+        [
+          "Line",
+          127
+        ],
+        [
+          "TableCell",
+          121
+        ],
+        [
+          "Text",
+          12
+        ],
+        [
+          "PageHeader",
+          7
+        ],
+        [
+          "SectionHeader",
+          3
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "Table",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "TableGroup",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 6,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          501
+        ],
+        [
+          "Line",
+          86
+        ],
+        [
+          "Text",
+          13
+        ],
+        [
+          "Caption",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Figure",
+          1
+        ],
+        [
+          "SectionHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "FigureGroup",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 7,
+      "text_extraction_method": "surya",
+      "block_counts": [
+        [
+          "Line",
+          130
+        ],
+        [
+          "Span",
+          125
+        ],
+        [
+          "Text",
+          50
+        ],
+        [
+          "TableCell",
+          40
+        ],
+        [
+          "Caption",
+          3
+        ],
+        [
+          "SectionHeader",
+          3
+        ],
+        [
+          "Reference",
+          3
+        ],
+        [
+          "PageHeader",
+          2
+        ],
+        [
+          "Table",
+          1
+        ],
+        [
+          "Figure",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "TableGroup",
+          1
+        ],
+        [
+          "FigureGroup",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 8,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          356
+        ],
+        [
+          "Line",
+          150
+        ],
+        [
+          "ListItem",
+          18
+        ],
+        [
+          "Reference",
+          17
+        ],
+        [
+          "SectionHeader",
+          2
+        ],
+        [
+          "Text",
+          2
+        ],
+        [
+          "ListGroup",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 9,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          329
+        ],
+        [
+          "Line",
+          141
+        ],
+        [
+          "ListItem",
+          16
+        ],
+        [
+          "Reference",
+          15
+        ],
+        [
+          "Text",
+          12
+        ],
+        [
+          "ListGroup",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 10,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          346
+        ],
+        [
+          "Line",
+          139
+        ],
+        [
+          "ListItem",
+          22
+        ],
+        [
+          "Reference",
+          21
+        ],
+        [
+          "Text",
+          10
+        ],
+        [
+          "ListGroup",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 11,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          357
+        ],
+        [
+          "Line",
+          152
+        ],
+        [
+          "ListItem",
+          19
+        ],
+        [
+          "Reference",
+          19
+        ],
+        [
+          "Text",
+          2
+        ],
+        [
+          "ListGroup",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 12,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          124
+        ],
+        [
+          "Line",
+          60
+        ],
+        [
+          "Text",
+          7
+        ],
+        [
+          "Reference",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 13,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          428
+        ],
+        [
+          "TableCell",
+          204
+        ],
+        [
+          "Line",
+          99
+        ],
+        [
+          "Text",
+          13
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "SectionHeader",
+          1
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "Table",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "TableGroup",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 14,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          291
+        ],
+        [
+          "Line",
+          131
+        ],
+        [
+          "TableCell",
+          30
+        ],
+        [
+          "Text",
+          7
+        ],
+        [
+          "SectionHeader",
+          2
+        ],
+        [
+          "Caption",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Figure",
+          1
+        ],
+        [
+          "Table",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "FigureGroup",
+          1
+        ],
+        [
+          "TableGroup",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 15,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          331
+        ],
+        [
+          "Line",
+          97
+        ],
+        [
+          "Text",
+          9
+        ],
+        [
+          "SectionHeader",
+          8
+        ],
+        [
+          "Equation",
+          3
+        ],
+        [
+          "ListItem",
+          3
+        ],
+        [
+          "Reference",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Picture",
+          1
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "PictureGroup",
+          1
+        ],
+        [
+          "ListGroup",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 16,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          267
+        ],
+        [
+          "Line",
+          99
+        ],
+        [
+          "ListItem",
+          34
+        ],
+        [
+          "ListGroup",
+          4
+        ],
+        [
+          "SectionHeader",
+          3
+        ],
+        [
+          "Text",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 17,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          211
+        ],
+        [
+          "Line",
+          83
+        ],
+        [
+          "ListItem",
+          22
+        ],
+        [
+          "Text",
+          8
+        ],
+        [
+          "SectionHeader",
+          6
+        ],
+        [
+          "ListGroup",
+          4
+        ],
+        [
+          "Picture",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 18,
+      "text_extraction_method": "surya",
+      "block_counts": [
+        [
+          "TableCell",
+          117
+        ],
+        [
+          "Line",
+          107
+        ],
+        [
+          "Span",
+          42
+        ],
+        [
+          "Text",
+          25
+        ],
+        [
+          "SectionHeader",
+          3
+        ],
+        [
+          "Reference",
+          3
+        ],
+        [
+          "Caption",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Figure",
+          1
+        ],
+        [
+          "Table",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "FigureGroup",
+          1
+        ],
+        [
+          "TableGroup",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 19,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          272
+        ],
+        [
+          "Line",
+          109
+        ],
+        [
+          "Text",
+          12
+        ],
+        [
+          "SectionHeader",
+          4
+        ],
+        [
+          "Figure",
+          2
+        ],
+        [
+          "Caption",
+          2
+        ],
+        [
+          "FigureGroup",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 20,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          448
+        ],
+        [
+          "Line",
+          80
+        ],
+        [
+          "Text",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 21,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          475
+        ],
+        [
+          "Line",
+          79
+        ],
+        [
+          "Picture",
+          7
+        ],
+        [
+          "Caption",
+          7
+        ],
+        [
+          "PictureGroup",
+          7
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Text",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 22,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          432
+        ],
+        [
+          "Line",
+          79
+        ],
+        [
+          "Picture",
+          8
+        ],
+        [
+          "Caption",
+          7
+        ],
+        [
+          "PictureGroup",
+          7
+        ],
+        [
+          "Text",
+          3
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    },
+    {
+      "page_id": 23,
+      "text_extraction_method": "pdftext",
+      "block_counts": [
+        [
+          "Span",
+          343
+        ],
+        [
+          "Line",
+          83
+        ],
+        [
+          "Text",
+          5
+        ],
+        [
+          "SectionHeader",
+          2
+        ],
+        [
+          "PageHeader",
+          1
+        ],
+        [
+          "Figure",
+          1
+        ],
+        [
+          "Caption",
+          1
+        ],
+        [
+          "PageFooter",
+          1
+        ],
+        [
+          "FigureGroup",
+          1
+        ],
+        [
+          "Reference",
+          1
+        ]
+      ],
+      "block_metadata": {
+        "llm_request_count": 0,
+        "llm_error_count": 0,
+        "llm_tokens_used": 0,
+        "previous_text": "",
+        "previous_type": "",
+        "previous_order": 0
+      }
+    }
+  ],
+  "debug_data_path": "debug_data/80c20b7b-ead6-454a-849e-56702a6c828f"
+}

icml26/80c20b7b-ead6-454a-849e-56702a6c828f/model_text_v3.txt

@@ -0,0 +1,209 @@
+[p. 1 | section: Abstract | type: Text]
+The scarcity of high-quality data remains a primary bottleneck in adapting multimodal generative models for medical image editing. Existing medical image editing datasets often suffer from limited diversity, neglect of medical image understanding and inability to balance quality with scalability. To address these gaps, we propose MieDB-100k, a large-scale, high-quality and diverse dataset for text-guided medical image editing. It categorizes editing tasks into perspectives of Perception, Modification and Transformation, considering both understanding and generation abilities. We construct MieDB-100k via a data curation pipeline leveraging both modality-specific expert models and rule-based data synthetic methods, followed by rigorous manual inspection to ensure clinical fidelity. Extensive experiments demonstrate that model trained with MieDB-100k consistently outperform both open-source and proprietary models while exhibiting strong generalization ability. We anticipate that this dataset will serve as a cornerstone for future advancements in specialized medical image editing. Dataset and code are publicly available at ub.com/Raiiyf/MieDB-100k
+
+[p. 1 | section: 1. Introduction | type: Text]
+Multimodal generative models (Wu et al., 2025a ;b; Liu et al., 2025c) have developed rapidly in recent years. In natural image domains, generative models are not only gradually unifying text-guided generation and editing tasks, but also progressively expanding their capabilities to encompass image modification and image understanding (Deng et al., 2025; Tong et al., 2025) . However, in medical image domains, their performance remains conspicuously limited, especially in the area of unified editing tasks (Liu et al.,
+
+[p. 1 | section: 1. Introduction | type: Text]
+2025b; Yang et al., 2025) . We attribute this performance degradation primarily to a fundamental scarcity of specialized medical image-editing data.
+
+[p. 1 | section: 1. Introduction | type: Text]
+While a few contemporary studies have proposed benchmarks or datasets for medical image editing, they remain insufficient in three key aspects: (1) limited diversity in medical image modalities. Unlike general computer vision, clinical imaging encompasses diverse modalities with distinct physical and structural foundations. However, existing research and datasets are restricted to a narrow range of imaging modalities (Chen & Feng, 2025; Liu et al., 2025b) , typically the widely available modalities such as Chest Xrays and CTs, which cannot adequately train or evaluate a model's ability across diverse clinical settings.
+
+[p. 1 | section: 1. Introduction | type: ListGroup]
+(2) Neglect of medical image understanding. Almost all medical image editing works only focus on conceptual modification and stylistic transformation tasks, but ignore visual perception tasks ( e.g. organ/lesion detection), which has been considered to be beneficial to the generation of image editing models (Huang et al., 2025; Deng et al., 2025) . Additionally, this clinical grounding ensures interpretability and corrects 'right-for-the-wrong-reason' edits, which is vital for safety-critical medical applications. Finally, neglecting understanding tasks also hinders the development of unified medical models that bridge understanding and generation. (3) Failure to ensure both data quality and scalability. Collection of medical image editing data is hindered by the difficulty of generating ground-truth counterfactuals. Some existing studies (Chen & Feng, 2025; Yang et al., 2025) distills general-purpose generative models for quick data scaling. However, these models are not tailored for medical use and hence produce results that lack clinical reliability and explainability. Conversely, previous work (Liu et al., 2025b) relies on extensive human involvement to manually collect real medical image pairs, which is notoriously difficult to scale up. Moreover, real-world longitudinal data often exhibits spatial misalignment and background inconsistency, as obtaining perfectly calibrated scan pairs is rare in practical medical settings.
+
+[p. 1 | section: 1. Introduction | type: Text]
+In this paper, we address the aforementioned limitations in previous research by introducing MieDB-100k, a largescale, high-quality and diverse dataset for text-guided medical image editing. MieDB-100k includes 112, 228 editing
+
+[p. 2 | section: 1. Introduction | type: FigureGroup]
+Figure 1. MieDB-100k overview. It categorizes medical image editing tasks into three perspectives, covering diverse medical modalities.
+
+[p. 2 | section: 1. Introduction | type: Text]
+data, covering 69 distinct editing targets and 10 diverse medical image modalities. We categorize editing tasks into three types: Perception , Modification and Transformation , which consider both model's intrinsic understanding and generation abilities on medical images. To enhance the data fidelity while preserving the scalability, we propose a data curation pipeline leveraging both modality-specific expert models and rule-based data synthetic methods. Additionally, for some complex tasks such as lesion modification, we introduce individuals with medical knowledge to perform manual quality checks on the data to ensure data quality. Finally, we introduced task-specific evaluation metrics to facilitate a comprehensive assessment of the editing models' performance.
+
+[p. 2 | section: 1. Introduction | type: Text]
+We evaluate existing open-source and closed-source multimodal generative models on MieDB-100k and argue that most of them cannot perform well in medical image editing. To further validate the reliability and utility of MieDB-100k, we finetune the OmniGen2 baseline on our dataset. Experimental results demonstrate that MieDB-100k facilitates a substantial performance leap in medical image editing tasks, surpassing or matching SOTA models including Nano Banana Pro. It also exhibits strong generalization ability driven by the synergy of understanding and generation tasks. We anticipate that this dataset will serve as a cornerstone for future advancements in specialized medical image editing.
+
+[p. 2 | section: 1. Introduction | type: Text]
+Our contributions can be summarized as follows:
+
+[p. 2 | section: 1. Introduction | type: ListGroup]
+(1) We propose a credible and scalable data curation pipeline to construct MieDB-100k , a large-scale, high-quality and highly diverse dataset for medical image editing with 69 targets and 10 medical image modalities. (2) We first unify the medical image understanding and generation into the paradigm of edit, and find that joint
+
+[p. 2 | section: 1. Introduction | type: Text]
+training yields performance gains for specific tasks.
+
+[p. 2 | section: 1. Introduction | type: Text]
+(3) We evaluate popular open-source and closed-source multimodal generative models on MieDB-100k , and observe that training with our data can significantly strengthens the model's capacity for medical image editing.
+
+[p. 2 | section: 2.1. Data Research for Medical Image Editing | type: Text]
+As an emerging area, multimodal medical generative modeling is currently supported by relatively few publicly available datasets for training and benchmarking (Tab. 1). In these works, the primary challenge lies in the construction of high-quality image-edit pairs. MedEBench (Liu et al., 2025b), an early benchmarking effort, curated pairs by manually collecting related images from medical documents. While this ensures clinical validity, the approach lacks scalability. Furthermore, the resulting image pairs often exhibit background inconsistencies, as achieving strict spatial calibration in real-world clinical settings is virtually impossible. Conversely, Med-banana-50K (Chen & Feng, 2025) proposed a fully autonomous pipeline where data construction and quality control were managed by Gemini. However, applying general-purpose models to specialized medical scenarios may introduce factual errors or inconsistent edits, raising concerns about data fidelity. Finally, MedGEN-Bench (Yang et al., 2025) introduced image-edit pairs using a mix of rule-based and model-based methods; however, the lack of specific architectural details hinders a thorough evaluation of their data quality. Moreover, existing benchmarks only focus on content generation evaluation, overlooking the critical aspect of medical image understanding.
+
+[p. 3 | section: 2.1. Data Research for Medical Image Editing | type: TableGroup]
+Table 1. Comparison of contemporary medical image editing benchmarks and datasets. In 'Perspective' column, P stands for Perception, M stands for Modification, and T stands for Transformation. Benchmark Size Modalities Targets Perspectives Source Human Inspection MedE-Bench (Liu et al., 2025b) \sim 1 k 4 13 M Real ✓ Med-banana-50K (Chen & Feng, 2025) \sim 50k 3 23 M Synthetic X MedGEN-Bench (Yang et al., 2025) \sim 6k 6 16 M, T Real & Synthetic ✓ MieDB-100k (Ours) \sim 100k 10 69 P, M, T Real & Synthetic ✓
+
+[p. 3 | section: 2.1. Data Research for Medical Image Editing | type: FigureGroup]
+Figure 2. Modality distribution (a) and prompt word cloud (b).
+
+[p. 3 | section: 2.2. Multimodal Generative Model | type: Text]
+129130
+
+[p. 3 | section: 2.2. Multimodal Generative Model | type: Text]
+131132
+
+[p. 3 | section: 2.2. Multimodal Generative Model | type: Text]
+Multimodal generative models (Liu et al., 2025c; Brooks et al., 2023) accept both images and natural language instructions as input, performing edits by translating semantic commands into precise visual manipulations. Recent studies (Wu et al., 2025a) often leverage vision-language model encoder and large-scale vision-language pretraining to align the semantic instruction with image modification. For instance, OmniGen2 (Wu et al., 2025b) utilizes Qwen2.5-VL (Bai et al., 2025b) to extract latent representations for semantic alignment, supported by a large-scale, multi-task training strategy. Furthermore, many recent studies (Deng et al., 2025) integrate image understanding and editing within a unified architecture. Exploiting these synergies is essential for creating robust models that are capable of performing both multimodal understanding and visual generation. On the commercial front, SOTA proprietary models like Gemini-3-Pro-Image (Nano Banana Pro) (DeepMind, 2025a) exhibit sophisticated image manipulation abilities, further realizing the real-world potential of multi-modal generative models. Despite these advancements, current models still struggle with the complexities of medical imaging(Liu et al., 2025b; Yang et al., 2025), highlighting the urgent need for comprehensive datasets to accelerate their adaptation to clinical domains.
+
+[p. 3 | section: 3. MieDB-100k | type: Text]
+This section introduces MieDB-100k, a high-quality, rigorous, and highly diverse dataset for medical image editing with more than 69 associated medical targets. It contains
+
+[p. 3 | section: 3. MieDB-100k | type: Text]
+112, 228 image-editing triplets. Figure 2(a) summarizes the distribution of samples across 10 imaging modalities.
+
+[p. 3 | section: 3.1. Data Definition | type: Text]
+Each entry in MieDB-100k is a triplet (I, P, O), where I is the input medical image, P is the textual prompt that describes edit operation, and O is the target image.
+
+[p. 3 | section: 3.2. Three Perspectives of MieDB-100k | type: Text]
+MieDB-100k is constructed under a novel categorization of three perspectives, considering both understanding and generation capabilities: (1) Perception tasks, which focus on model's intrinsic medical knowledge via pixel-wise identification of prompted clinical targets in the input image; (2) Modification tasks, which require the model to locate and alter specific medical features; and (3) Transformation tasks, involving medical image restoration, enhancement, and other low-level transformation. To ensure the rigor of the data triplets while maintaining scalability, we designed and implemented a specialized data construction pipeline for MieDB-100k (Fig 3), and we list all source datasets used for construction in App. A.
+
+[p. 3 | section: 3.2.1. Perception | type: Text]
+Perception tasks focus on medical image understanding, and we we formulate it as an editing task by instructing model to generate masks over regions of interest (ROIs), such as specific organs or lesions, through textual prompts. Notably, to align with image editing paradigm, the model is prompted to overlay the localization mask directly onto the source image rather than generating a standalone binary mask. This task serves two primary functions: First, since the mask-painting task only requires minimal pixel manipulation (typically modifying a single channel within a specific region), it serves as a direct assessment of the medical knowledge embedded in the generative model, isolating its perceptual accuracy from complex synthesis capabilities. Second, it introduces a promising application for multimodal generative models in the medical domain: assisted interpretation in multimodal manner. By allowing users to highlight specific targets in medical image through natural language prompts, this approach can assist
+
+[p. 4 | section: 3.2.1. Perception | type: Text]
+patients in understanding their diagnostic images, aid medical students in their education, and reduce screening time for senior clinicians.
+
+[p. 4 | section: 3.2.1. Perception | type: Text]
+The rule-based construction process for the Perception task's data triplets is illustrated in Fig. 3. Specifically, for a segmentation dataset, the original image serves as the input I. The output image O is synthesized by overlaying the ground-truth segmentation label, which is rendered in a randomly selected color (red, green, or blue), onto the input image with an alpha-blending transparency of 0.6. The ROIs of perception can be classified into three types: anatomical structure (organ, organism and so on), lesion area and holistic segmentation (segment all visible and clinically significant structures). We specifies the perception target and visualizing color scheme in the textual prompt P. Since this part of the data is constructed following a definite rule, it can be readily scaled up to a diverse set of medical knowledge assessments and to the associated training dataset by leveraging the extensive body of existing medical segmentation research. Finally, to ensure a high-quality final benchmark, we manually filtered the initial data pool to remove trivial, redundant, or incorrectly labeled samples.
+
+[p. 4 | section: 3.2.2. MODIFICATION | type: Text]
+167168
+
+[p. 4 | section: 3.2.2. MODIFICATION | type: Text]
+171172
+
+[p. 4 | section: 3.2.2. MODIFICATION | type: Text]
+177178179
+
+[p. 4 | section: 3.2.2. MODIFICATION | type: Text]
+The perspective of Modification is specifically designed for semantically modifying medical contents, so as to address the diverse requirements of editing beyond just locate them. However, constructing modification data triplets is
+
+[p. 4 | section: 3.2.2. MODIFICATION | type: Text]
+challenging because counterfactual image pairs cannot be captured simultaneously in the real world. While one could theoretically leverage general-purpose generative models (e.g., Nano Banana Pro or Qwen-Image-Edit) to produce these edits, such models are not specialized for the medical domain, and therefore are prone to severe hallucinations, which is unacceptable in a healthcare context. To construct rigorous edit triplets and preserve scalability, we propose a four-stage process (Fig. 3) designed to bridge the gap between task complexity and model competence so as to fully utilize these automatic tools.
+
+[p. 4 | section: 3.2.2. MODIFICATION | type: Text]
+Stage I: We develop a suite of modality-specific expert models for healthy tissue inpainting, built upon the FLUX.1-Fill-dev model. This strategy is based on the observation that generating healthy anatomical structures is more stable and predictable than generating lesions, as the former exhibits more tractable patterns and textures. For each modality, we curate a training dataset consisting exclusively of non-pathological samples from existing medical image repositories. Through parameter-efficient finetuning, these models learn to inpaint masked areas with high clinical accuracy. We further apply background restoration and edge blending to correct any unintended modifications made by the FLUX model outside the mask, ensuring the edited region blends seamlessly into the original image.
+
+[p. 4 | section: 3.2.2. MODIFICATION | type: Text]
+Stage II: We leverage these expert models to modify lesion-bearing images (L) into their counterfactual 'healthy' results (H). Specifically, we fill the lesion area
+
+[p. 5 | section: 3.2.2. MODIFICATION | type: Text]
+in L using white pixels based on its ground-truth segmentation label. This masked image and its corresponding binary mask are then processed by the modality-matched expert model to synthesize H, where healthy tissue replaces lesion. Compared to distilling general-purpose generative models, our modality-specific approach not only restricts the high-variance generative process to a localized region to guarantee background consistency during the edit, but also ensures that tasks remain within the model's learned distribution, thereby significantly reducing hallucinations. Furthermore, unlike manual data collection from the internet, our approach provides superior scalability and efficiency.
+
+[p. 5 | section: 3.2.2. MODIFICATION | type: Text]
+Stage III: We implement a rejection sampling mechanism for the generated 'healthy' images (H) to further enhance the data quality within the Modification tasks. For modalities that resemble natural images (e.g., endoscopy and dermoscopy), we prompt the Qwen3-VL-32B-Instruct model (Bai et al., 2025a) to filter out H that still contain lesions, exhibit artifacts, or are of low quality. For other modalities, we train separate nnUNet models (Isensee et al., 2021) for lesion segmentation and discard H where lesions remain detectable.
+
+[p. 5 | section: 3.2.2. MODIFICATION | type: Text]
+Stage IV: Triplet combination. Using these high quality 'lesion-healthy' counterfactual pairs, we generate diverse Modification task data by swapping L and H from niche of input and output and varying the textual prompts P.
+
+[p. 5 | section: 3.2.3. TRANSFORMATION | type: Text]
+Transformation tasks include a wide array of low-level medical image processing operations. Unlike the localized edits found in Perception and Modification categories, tasks in this category typically require a holistic transformation of the entire input image.
+
+[p. 5 | section: 3.2.3. TRANSFORMATION | type: Text]
+The rule-based construction pipeline of Transformation tasks is shown in Fig. 3. From public repositories, we compile medical image pairs (I and O) representing 17 distinct transformation targets under four typical low-level vision categories. We then design specialized textual prompts P for each task to unify diverse medical image processing functions into a consistent image editing framework.
+
+[p. 5 | section: 3.2.4. POST PROCESSING | type: Text]
+Prompt rephrasing. To enhance linguistic diversity, we utilize the Qwen-Max model to rephrase the prompts P for each data triplet. We also illustrate the linguistic diversity of our prompts via a word cloud in Figure 2( b).
+
+[p. 5 | section: 3.2.4. POST PROCESSING | type: Text]
+Benchmark curation. The training and test split of source dataset are strictly followed during the construction of MieDB-100k to preclude any data leakage. Furthermore, we recruit three people with clinical background to manually evaluate and curate 3, 485 of the most representative
+
+[p. 5 | section: 3.2.4. POST PROCESSING | type: Text]
+samples characterized by high clinical fidelity from raw data test split to serve as the benchmark of MieDB-100k, and we keep their original image size to minimize information loss.
+
+[p. 5 | section: 3.2.4. POST PROCESSING | type: Text]
+Train split construction. For train split, we establish three resolution bins (128, 256, and 512) and resize images to their nearest corresponding value. To check the fidelity of training split, we randomly select 6, 000 triplets for clinician evaluation, and over 95% are viewed as high quality.
+
+[p. 5 | section: 3.3. MieDB-100k Evaluation | type: Text]
+We evaluate MieDB-100k through two distinct approaches: (1) verifiable metrics for the Perception and Transformation tasks, amenable to reward design in prevailing reinforcement learning algorithms (Shao et al., 2024; Liu et al., 2025a) ; and (2) more subjective evaluations for the Modification tasks, reflecting their greater complexity.
+
+[p. 5 | section: 3.3.1. VERIFIABLE EVALUATION | type: Text]
+Localization Accuracy Metric. We use the DICE Score for evaluating the spatial overlapping performance in Perception tasks. Notably, reconstructing a binary mask from the colored regions of an edited image is mathematically feasible when the background image and overlay color are known, and we detail this process in App. D.1. This procedure is applied to both model's output O M and the ground truth images O to derive the mask of model's perceptual region and the ground truth region for DICE calculation.
+
+[p. 5 | section: 3.3.1. VERIFIABLE EVALUATION | type: Text]
+To differentiate between models that accurately identify specific medical targets and those that merely generate coarsegrained masks, we further propose Perception Accuracy. Under this metric, a result is considered successful only if the DICE score exceeds a threshold of τ = 0.8. This metric allows us to analyze whether a model possesses the specialized medical knowledge required for image understanding.
+
+[p. 5 | section: 3.3.1. VERIFIABLE EVALUATION | type: Text]
+Image Similarity Metrics. We utilize PSNR and SSIM (Wang et al., 2004) to evaluate the similarity between the ground-truth and edited images at both the pixel and structural levels. For evaluations within the Perception perspective, we mask out the pixels corresponding to the groundtruth segmentation in both images. This allows us to specifically assess the model's ability to preserve the background while performing the requested edit.
+
+[p. 5 | section: 3.3.2. EVALUATION FOR MODIFICATION TASKS | type: Text]
+Vision-Language Model Rubric Scoring. Automating reliable assessments in the Modification tasks is inherently challenging, as edits are defined semantically and cannot be evaluated via deterministic rules. Existing benchmarks often leverage Vision-Language Models (VLMs) for this purpose, and we standardize the process and mitigate potential critic hallucinations by implementing a rubric-based
+
+[p. 6 | section: 3.3.2. EVALUATION FOR MODIFICATION TASKS | type: TableGroup]
+Table 2. Overall result on MieDB-100k benchmark. P-ACC means Perception Accuracy; B-PSNR and B-SSIM mean only calculate PSNR and SSIM on background pixels respectively; Rubric-S stands for the Rubric Score from VLM and Pref-Rank stands for human preference ranking. Best values are marked in red while second bests are in blue. Perception Modification Trasnformation Model Name Size DICE P-ACC B-PSNR B-SSIM Rubric-S Pref-Rank PSNR SSIM Open-Source SDXL-turbo (Sauer et al., 2024) 3.5B 0.002 0.000 16.6 0.467 8.4 7.7 15.9 0.397 Bagel (Deng et al., 2025) 7B 0.263 0.069 13.9 0.620 34.4 6.2 12.7 0.442 OmniGen2 (Wu et al., 2025b) 7B 0.248 0.065 11.9 0.541 29.1 7.1 8.3 0.280 Step1X-Edit (Liu et al., 2025c) 21B 0.332 0.126 15.5 0.727 35.6 4.5 16.6 0.539 Qwen-Image-Edit (Wu et al., 2025a) 27B 0.387 0.153 15.4 0.722 32.2 5.5 18.9 0.606 FLUX.1-Kontext-dev (Labs et al., 2025) 12B 0.341 0.126 15.4 0.701 37.8 6.2 17.9 0.543 OmniGen2-MIE (Ours) 7B 0.831 0.737 28.1 0.917 65.9 1.4 22.6 0.685 Closed-Source GPT-Image-1 (OpenAI, 2025) 0.467 0.221 16.3 0.510 42.8 4.8 14.4 0.451 Nano Banana Pro (DeepMind, 2025a) 0.426 0.202 12.8 0.413 63.4 2.0 20.0 0.610 Imagen4 (DeepMind, 2025b) 0.142 0.000 8.9 0.210 19.7 7.4 7.9 0.174
+
+[p. 6 | section: 3.3.2. EVALUATION FOR MODIFICATION TASKS | type: Text]
+scoring system. Specifically, we provide the VLM with the input image I, edit instruction P, reference output O, and the model's generated result OM. Guided by the rubric, the VLM then performs a holistic evaluation of OM.
+
+[p. 6 | section: 3.3.2. EVALUATION FOR MODIFICATION TASKS | type: Text]
+We design a comprehensive scoring rubric (App. D.2) that assesses both the fulfillment of the editing intent and the model's ability to preserve the background. We utilize GPT-5.2 as an automated evaluator for this process, and map the final score to [0, 100].
+
+[p. 6 | section: 3.3.2. EVALUATION FOR MODIFICATION TASKS | type: Text]
+Human Preference Ranking. For each test case, we present the original triplet (I, P, O) and the outputs of all tested models simultaneously to evaluators, who are then asked to rank the various model-generated results according to their preference. By forcing this comparative ordering of all models, we are able to move beyond absolute quality scores and capture the relative strengths and weaknesses of current generative frameworks in a clinical setting. Specifically, we recruit 3 evaluators with clinical backgrounds to assess and rank the images edited by the benchmarked models, and compute the average ranking.
+
+[p. 6 | section: 4.1. Baselines | type: Text]
+We evaluate nine models on MieDB-100k, comprising six open-source models: Qwen-Image-Edit-2511 (Wu et al., 2025a) , Bagel (Deng et al., 2025) , OmniGen2 (Wu et al., 2025b) , Step1X-Edit-v1p2 (Liu et al., 2025c) , FLUX.1- Kontext-dev (Labs et al., 2025) and SDXL-turbo (Sauer et al., 2024) , plus three closed-source models: Nano Banana Pro (DeepMind, 2025a) , GPT-Image-1 (OpenAI, 2025) , and Imagen4 (DeepMind, 2025b) . We implement open-source models following their official inference settings.
+
+[p. 6 | section: 4.1. Baselines | type: Text]
+To validate the effectiveness of MieDB-100k, we finetune
+
+[p. 6 | section: 4.1. Baselines | type: Text]
+the OmniGen2 baseline on the training split and subject it to the same evaluation protocol as the other models. Specifically, we train the Diffusion Transformer (DiT) component for 20,000 iterations, employing a global batch size of 64 and a learning rate of 1e-4.
+
+[p. 6 | section: 4.2. Quantitative Results | type: Text]
+We report the benchmarking results of MieDB-100k in Tab. 2. First, the extremely low perception accuracy indicates that all tested models except ours fail to accurately comprehend and localize the specified anatomical targets under our evaluation protocol. Consequently, in Modification tasks, most of them are unable to generate clinically meaningful edits. Although a few models, such as Nano Banana Pro, achieve competitive results, we are indeed observing the 'right-for-the-wrong-reason' phenomenon, a risk that must be strictly avoided in clinical settings. Since the poor performance in Perception tasks expose their intrinsic lack of necessary medical knowledge, their edits cannot be justified. Notably in Transformation tasks, Nano Banana Pro also presents competitive results in certain cases. This may be attributed to the similarity between tasks like denoising or artifact removal and general-purpose low-level vision tasks, for which the model already possesses some capability (Zuo et al., 2025) . Alternatively, it is possible that similar medical image processing tasks were included in its training set. Regardless, its absolute performance remains insufficient for practical clinical deployment. In summary, the benchmark result demonstrates that current multimodal generative model cannot meet the requirement of medical imaging editing.
+
+[p. 6 | section: 4.2. Quantitative Results | type: Text]
+Conversely, after training on MieDB-100k, a standard baseline model can achieve superior medical editing capabilities. As shown in Tab. 2, the OmniGen2-MIE model delivers the best performance across all three edit-
+
+[p. 7 | section: 4.2. Quantitative Results | type: FigureGroup]
+[Edit Instruction] Process the T1 MRI to mitigate motion effects without changing the appearance of brain tissue, tumors, or other clinical features.
+
+[p. 7 | section: 4.2. Quantitative Results | type: Caption]
+Figure 4. Qualitative editing result comparison.
+
+[p. 7 | section: 4.2. Quantitative Results | type: Text]
+ing perspectives. The most significant improvements are observed in the Perception perspective, which demonstrate that MieDB-100k can effectively inject essential medical knowledge, thereby enhancing the interpretability of downstream editing tasks. Furthermore, in the Modification and Transformation tasks, where general-purpose editing abilities transfer more readily, our enhanced model still yields superior editing results compared to Nano Banana Pro, the SOTA multi-modal generative models. These findings highlight the pivotal role of our dataset in domain adaptation and
+
+[p. 7 | section: 4.2. Quantitative Results | type: Text]
+establish a foundation for the development of understandinggeneration unified medical models.
+
+[p. 7 | section: 4.3. Qualitative Results | type: Text]
+Fig. 4 presents qualitative editing results for several baseline models across the diverse modalities and tasks in MieDB-100k. These results demonstrate that the finetuned model exhibits an enhanced capability in both understanding and generation, allowing it to navigate the inherent complexities
+
+[p. 8 | section: 4.3. Qualitative Results | type: TableGroup]
+Table 3. Ablation study result on MieDB-100k. P stands for Perception, M stands for Modification, and T stands for Transformation. Best values are marked in red, second bests are in blue. Perce ption Modification Trasnformation Training Data DICE ACC RubricScore PSNR SSIM Baseline (No train) 0.248 0.065 29.1 8.3 0.280 P-only 0.833 0.740 37.8 19.7 0.631 M-only 0.001 0.000 57.5 19.8 0.631 T-only 0.034 0.000 15.0 23.7 0.702 MieDB-100k 0.831 0.737 65.9 22.6 0.685
+
+[p. 8 | section: 4.3. Qualitative Results | type: Text]
+of medical image editing. Moreover, despite being explicitly prompted, even sophisticated closed-source models such as Nano Banana Pro fail to maintain background consistency in certain tasks. While their instruction-following proficiency stems from large-scale pre-training on natural image pairs, these capabilities tend to degrade when the distribution of medical modalities deviates significantly from the natural images seen during pre-training. To further study the impact of modality deviation, we conduct a modality-wise analysis in App. E.1, and the results prove our judgment. This observation underscores the necessity of a highly diverse dataset like MieDB-100k to equip models with the capacity to handle a vast range of medical imaging modalities.
+
+[p. 8 | section: 4.4. Ablation Study | type: Text]
+To investigate the contribution of each task category, we conduct an ablation study by training models on individual perspective of MieDB-100k. We again utilize OmniGen2 as baseline model, following the training recipe described above while varying only the training data. As shown in Tab 3, each specialized model significantly outperforms the original baseline in its respective domain, validating the high information density and clinical relevance of our data. For the model trained on the full dataset, it achieves comparable or even better performance on all three perspectives, showing the effectiveness of the joint training. More importantly, we observe significant performance improvement in the Modification perspective, demonstrating visual understanding ability has the potential to enhance visual generation ability. In summary, the ablation study shows that MieDB-100k can provide a synergistic training signal, enabling the development of a versatile model capable of handling diverse medical editing tasks simultaneously.
+
+[p. 8 | section: 4.5. Generalization Test | type: Text]
+To further investigate the cross-task synergy and the resulting generalization capabilities, we conduct an out-ofdistribution (OOD) editing experiment. Specifically, we target 'bone metastasis', a medical target included in Perception tasks but strictly excluded from the Modification training data. We then prompt the OmniGen2-MIE model
+
+[p. 8 | section: 4.5. Generalization Test | type: FigureGroup]
+Figure 5. Generalization test assessment. (a) and (b): Edit samples output by different models on bone metastasis addition (a) and removal (b) tasks. Red bounding boxes are added post-hoc to highlight the edited regions for visualization; (c): Quantitative assessments following the recipe of Modification task evaluation.
+
+[p. 8 | section: 4.5. Generalization Test | type: Text]
+to perform metastasis addition and removal in CT scans.
+
+[p. 8 | section: 4.5. Generalization Test | type: Text]
+As shown in Fig. 5, OmniGen2-MIE significantly outperforms OmniGen2 on this unseen task, demonstrating that our unified training on MieDB-100k can enhance the model's generalization capabilities across editing tasks. We also observe that Nano Banana Pro achieves the best OOD editing performance, marginally surpassing OmniGen2-MIE. We attribute this performance to the utilization of massive-scale general and medical editing data, which further underscores the necessity of scaling up medical editing data.
+
+[p. 8 | section: 5. Conclusion | type: Text]
+In this paper, we introduce MieDB-100k, a large-scale and diverse dataset for text-guided medical image editing. By unifying Perception, Modification, and Transformation tasks into the paradigm of editing, our dataset bridges the gap between medical image understanding and generation. We develop a robust curation pipeline, integrating modalityspecific expert models with rule-based synthesis, and enforce rigorous manual quality control to ensure clinical fidelity across all data. Extensive benchmarking demonstrates that model trained on MieDB-100k consistently outperform both SOTA open-source and proprietary multimodal models while exhibiting exceptional generalization to unseen clinical tasks. Our work thus provides the data foundation to support the development and evaluation of multimodal generative models for clinical applications.
+
+[p. 9 | section: Impact Statement | type: Text]
+This paper presents work whose goal is to advance the field of Machine Learning. There are many potential societal consequences of our work, none which we feel must be specifically highlighted here.

icml26/80c20b7b-ead6-454a-849e-56702a6c828f/paper.md

@@ -0,0 +1,1136 @@
+
+
+{0}------------------------------------------------
+
+# MieDB-100k: A Comprehensive Dataset for Medical Image Editing
+
+# Anonymous Authors<sup>1</sup>
+
+# Abstract
+
+The scarcity of high-quality data remains a primary bottleneck in adapting multimodal generative models for medical image editing. Existing medical image editing datasets often suffer from limited diversity, neglect of medical image understanding and inability to balance quality with scalability. To address these gaps, we propose MieDB-100k, a large-scale, high-quality and diverse dataset for text-guided medical image editing. It categorizes editing tasks into perspectives of Perception, Modification and Transformation, considering both understanding and generation abilities. We construct MieDB-100k via a data curation pipeline leveraging both modality-specific expert models and rule-based data synthetic methods, followed by rigorous manual inspection to ensure clinical fidelity. Extensive experiments demonstrate that model trained with MieDB-100k consistently outperform both open-source and proprietary models while exhibiting strong generalization ability. We anticipate that this dataset will serve as a cornerstone for future advancements in specialized medical image editing. Dataset and code are publicly available at [https://gith](https://github.com/Raiiyf/MieDB-100k) [ub.com/Raiiyf/MieDB-100k](https://github.com/Raiiyf/MieDB-100k)
+
+# 1. Introduction
+
+Multimodal generative models [\(Wu et al.,](#page-11-0) [2025a](#page-11-0)[;b;](#page-11-1) [Liu](#page-10-0) [et al.,](#page-10-0) [2025c\)](#page-10-0) have developed rapidly in recent years. In natural image domains, generative models are not only gradually unifying text-guided generation and editing tasks, but also progressively expanding their capabilities to encompass image modification and image understanding [\(Deng](#page-9-0) [et al.,](#page-9-0) [2025;](#page-9-0) [Tong et al.,](#page-11-2) [2025\)](#page-11-2). However, in medical image domains, their performance remains conspicuously limited, especially in the area of unified editing tasks [\(Liu et al.,](#page-10-1)
+
+Preliminary work. Under review by the International Conference on Machine Learning (ICML). Do not distribute.
+
+[2025b;](#page-10-1) [Yang et al.,](#page-11-3) [2025\)](#page-11-3). We attribute this performance degradation primarily to a fundamental scarcity of specialized medical image-editing data.
+
+While a few contemporary studies have proposed benchmarks or datasets for medical image editing, they remain insufficient in three key aspects: (1) limited diversity in medical image modalities. Unlike general computer vision, clinical imaging encompasses diverse modalities with distinct physical and structural foundations. However, existing research and datasets are restricted to a narrow range of imaging modalities [\(Chen & Feng,](#page-8-0) [2025;](#page-8-0) [Liu et al.,](#page-10-1) [2025b\)](#page-10-1), typically the widely available modalities such as Chest Xrays and CTs, which cannot adequately train or evaluate a model's ability across diverse clinical settings.
+
+- (2) Neglect of medical image understanding. Almost all medical image editing works only focus on conceptual modification and stylistic transformation tasks, but ignore visual perception tasks (*e.g.* organ/lesion detection), which has been considered to be beneficial to the generation of image editing models [\(Huang et al.,](#page-9-1) [2025;](#page-9-1) [Deng et al.,](#page-9-0) [2025\)](#page-9-0). Additionally, this clinical grounding ensures interpretability and corrects 'right-for-the-wrong-reason' edits, which is vital for safety-critical medical applications. Finally, neglecting understanding tasks also hinders the development of unified medical models that bridge understanding and generation.
+- (3) Failure to ensure both data quality and scalability. Collection of medical image editing data is hindered by the difficulty of generating ground-truth counterfactuals. Some existing studies [\(Chen & Feng,](#page-8-0) [2025;](#page-8-0) [Yang et al.,](#page-11-3) [2025\)](#page-11-3) distills general-purpose generative models for quick data scaling. However, these models are not tailored for medical use and hence produce results that lack clinical reliability and explainability. Conversely, previous work [\(Liu et al.,](#page-10-1) [2025b\)](#page-10-1) relies on extensive human involvement to manually collect real medical image pairs, which is notoriously difficult to scale up. Moreover, real-world longitudinal data often exhibits spatial misalignment and background inconsistency, as obtaining perfectly calibrated scan pairs is rare in practical medical settings.
+
+In this paper, we address the aforementioned limitations in previous research by introducing MieDB-100k, a largescale, high-quality and diverse dataset for text-guided medical image editing. MieDB-100k includes 112, 228 editing
+
+<sup>1</sup>Anonymous Institution, Anonymous City, Anonymous Region, Anonymous Country. Correspondence to: Anonymous Author <anon.email@domain.com>.
+
+{1}------------------------------------------------
+
+![](_page_1_Figure_1.jpeg)
+
+Figure 1. MieDB-100k overview. It categorizes medical image editing tasks into three perspectives, covering diverse medical modalities.
+
+data, covering **69** distinct editing targets and **10** diverse medical image modalities. We categorize editing tasks into three types: **Perception**, **Modification** and **Transformation**, which consider both model's intrinsic understanding and generation abilities on medical images. To enhance the data fidelity while preserving the scalability, we propose a data curation pipeline leveraging both modality-specific expert models and rule-based data synthetic methods. Additionally, for some complex tasks such as lesion modification, we introduce individuals with medical knowledge to perform manual quality checks on the data to ensure data quality. Finally, we introduced task-specific evaluation metrics to facilitate a comprehensive assessment of the editing models' performance.
+
+061
+
+062
+
+063 064 065
+
+066
+
+067
+
+068
+
+071 072
+
+074 075
+
+076
+
+077
+
+078
+
+081
+
+082
+
+083
+
+084
+
+085
+
+087
+
+088
+
+089
+
+090
+
+091
+
+092
+
+093
+
+094
+
+095
+
+096
+
+097
+
+098
+
+099
+
+100
+
+104
+
+105
+
+106
+
+108 109 We evaluate existing open-source and closed-source multimodal generative models on MieDB-100k and argue that most of them cannot perform well in medical image editing. To further validate the reliability and utility of MieDB-100k, we finetune the OmniGen2 baseline on our dataset. Experimental results demonstrate that MieDB-100k facilitates a substantial performance leap in medical image editing tasks, surpassing or matching SOTA models including Nano Banana Pro. It also exhibits strong generalization ability driven by the synergy of understanding and generation tasks. We anticipate that this dataset will serve as a cornerstone for future advancements in specialized medical image editing.
+
+Our contributions can be summarized as follows:
+
+- (1) We propose a credible and scalable data curation pipeline to construct **MieDB-100k**, a large-scale, high-quality and highly diverse dataset for medical image editing with 69 targets and 10 medical image modalities.
+- (2) We first unify the medical image understanding and generation into the paradigm of edit, and find that joint
+
+training yields performance gains for specific tasks.
+
+(3) We evaluate popular open-source and closed-source multimodal generative models on **MieDB-100k**, and observe that training with our data can significantly strengthens the model's capacity for medical image editing.
+
+#### 2. Related Work
+
+### 2.1. Data Research for Medical Image Editing
+
+As an emerging area, multimodal medical generative modeling is currently supported by relatively few publicly available datasets for training and benchmarking (Tab. 1). In these works, the primary challenge lies in the construction of high-quality image-edit pairs. MedEBench (Liu et al., 2025b), an early benchmarking effort, curated pairs by manually collecting related images from medical documents. While this ensures clinical validity, the approach lacks scalability. Furthermore, the resulting image pairs often exhibit background inconsistencies, as achieving strict spatial calibration in real-world clinical settings is virtually impossible. Conversely, Med-banana-50K (Chen & Feng, 2025) proposed a fully autonomous pipeline where data construction and quality control were managed by Gemini. However, applying general-purpose models to specialized medical scenarios may introduce factual errors or inconsistent edits, raising concerns about data fidelity. Finally, MedGEN-Bench (Yang et al., 2025) introduced image-edit pairs using a mix of rule-based and model-based methods; however, the lack of specific architectural details hinders a thorough evaluation of their data quality. Moreover, existing benchmarks only focus on content generation evaluation, overlooking the critical aspect of medical image understanding.
+
+{2}------------------------------------------------
+
+<span id="page-2-0"></span>*Table 1.* Comparison of contemporary medical image editing benchmarks and datasets. In 'Perspective' column, P stands for Perception, M stands for Modification, and T stands for Transformation.
+
+| Benchmark                          | Size        | Modalities | Targets | Perspectives | Source           | <b>Human Inspection</b> |
+|------------------------------------|-------------|------------|---------|--------------|------------------|-------------------------|
+| MedE-Bench (Liu et al., 2025b)     | $\sim 1 k$  | 4          | 13      | M            | Real             | ✓                       |
+| Med-banana-50K (Chen & Feng, 2025) | $\sim$ 50k  | 3          | 23      | M            | Synthetic        | X                       |
+| MedGEN-Bench (Yang et al., 2025)   | $\sim$ 6k   | 6          | 16      | M, T         | Real & Synthetic | ✓                       |
+| MieDB-100k (Ours)                  | $\sim$ 100k | 10         | 69      | P, M, T      | Real & Synthetic | <b>✓</b>                |
+
+<span id="page-2-1"></span>![](_page_2_Figure_3.jpeg)
+
+Figure 2. Modality distribution (a) and prompt word cloud (b).
+
+#### 2.2. Multimodal Generative Model
+
+110
+
+120
+
+121
+
+122
+
+123
+
+124
+
+125
+
+126
+
+127
+
+128
+
+129130
+
+131132
+
+133
+
+134
+
+135
+
+136
+
+137
+
+138
+
+139
+
+140
+
+141
+
+142
+
+143
+
+144
+
+145
+
+146
+
+147
+
+148
+
+149
+
+150
+
+151
+
+152
+
+153
+
+154
+
+155
+
+156
+
+157
+
+158
+
+159
+
+160
+
+161
+
+162
+
+163
+
+164
+
+Multimodal generative models (Liu et al., 2025c; Brooks et al., 2023) accept both images and natural language instructions as input, performing edits by translating semantic commands into precise visual manipulations. Recent studies (Wu et al., 2025a) often leverage vision-language model encoder and large-scale vision-language pretraining to align the semantic instruction with image modification. For instance, OmniGen2 (Wu et al., 2025b) utilizes Qwen2.5-VL (Bai et al., 2025b) to extract latent representations for semantic alignment, supported by a large-scale, multi-task training strategy. Furthermore, many recent studies (Deng et al., 2025) integrate image understanding and editing within a unified architecture. Exploiting these synergies is essential for creating robust models that are capable of performing both multimodal understanding and visual generation. On the commercial front, SOTA proprietary models like Gemini-3-Pro-Image (Nano Banana Pro) (DeepMind, 2025a) exhibit sophisticated image manipulation abilities, further realizing the real-world potential of multi-modal generative models. Despite these advancements, current models still struggle with the complexities of medical imaging(Liu et al., 2025b; Yang et al., 2025), highlighting the urgent need for comprehensive datasets to accelerate their adaptation to clinical domains.
+
+#### 3. MieDB-100k
+
+This section introduces MieDB-100k, a high-quality, rigorous, and highly diverse dataset for medical image editing with more than 69 associated medical targets. It contains
+
+112, 228 image-editing triplets. Figure 2(a) summarizes the distribution of samples across 10 imaging modalities.
+
+#### 3.1. Data Definition
+
+Each entry in MieDB-100k is a triplet (I, P, O), where I is the input medical image, P is the textual prompt that describes edit operation, and O is the target image.
+
+#### 3.2. Three Perspectives of MieDB-100k
+
+MieDB-100k is constructed under a novel categorization of three perspectives, considering both understanding and generation capabilities: (1) **Perception** tasks, which focus on model's intrinsic medical knowledge via pixel-wise identification of prompted clinical targets in the input image; (2) **Modification** tasks, which require the model to locate and alter specific medical features; and (3) **Transformation** tasks, involving medical image restoration, enhancement, and other low-level transformation. To ensure the rigor of the data triplets while maintaining scalability, we designed and implemented a specialized data construction pipeline for MieDB-100k (Fig 3), and we list all source datasets used for construction in App. A.
+
+# 3.2.1. Perception
+
+Perception tasks focus on medical image understanding, and we we formulate it as an editing task by instructing model to generate masks over regions of interest (ROIs), such as specific organs or lesions, through textual prompts. Notably, to align with image editing paradigm, the model is prompted to overlay the localization mask directly onto the source image rather than generating a standalone binary mask. This task serves two primary functions: First, since the mask-painting task only requires minimal pixel manipulation (typically modifying a single channel within a specific region), it serves as a direct assessment of the medical knowledge embedded in the generative model, isolating its perceptual accuracy from complex synthesis capabilities. Second, it introduces a promising application for multimodal generative models in the medical domain: assisted interpretation in multimodal manner. By allowing users to highlight specific targets in medical image through natural language prompts, this approach can assist
+
+{3}------------------------------------------------
+
+<span id="page-3-0"></span>![](_page_3_Figure_1.jpeg)
+
+patients in understanding their diagnostic images, aid medical students in their education, and reduce screening time for senior clinicians.
+
+The rule-based construction process for the Perception task's data triplets is illustrated in Fig. 3. Specifically, for a segmentation dataset, the original image serves as the input I. The output image O is synthesized by overlaying the ground-truth segmentation label, which is rendered in a randomly selected color (red, green, or blue), onto the input image with an alpha-blending transparency of 0.6. The ROIs of perception can be classified into three types: anatomical structure (organ, organism and so on), lesion area and holistic segmentation (segment all visible and clinically significant structures). We specifies the perception target and visualizing color scheme in the textual prompt P. Since this part of the data is constructed following a definite rule, it can be readily scaled up to a diverse set of medical knowledge assessments and to the associated training dataset by leveraging the extensive body of existing medical segmentation research. Finally, to ensure a high-quality final benchmark, we manually filtered the initial data pool to remove trivial, redundant, or incorrectly labeled samples.
+
+#### 3.2.2. MODIFICATION
+
+166
+
+167168
+
+169
+
+170
+
+171172
+
+174
+
+175
+
+176
+
+177178179
+
+180
+
+181 182
+
+183
+
+184
+
+185
+
+186 187
+
+188 189 190
+
+191
+
+192
+
+194
+
+195
+
+196
+
+197
+
+198
+
+199
+
+200
+
+204
+
+205
+
+206
+
+208
+
+209
+
+210
+
+211
+
+212
+
+213
+
+214
+
+215
+
+216
+
+217
+
+218
+
+219
+
+The perspective of Modification is specifically designed for semantically modifying medical contents, so as to address the diverse requirements of editing beyond just locate them. However, constructing modification data triplets is challenging because counterfactual image pairs cannot be captured simultaneously in the real world. While one could theoretically leverage general-purpose generative models (e.g., Nano Banana Pro or Qwen-Image-Edit) to produce these edits, such models are not specialized for the medical domain, and therefore are prone to severe hallucinations, which is unacceptable in a healthcare context. To construct rigorous edit triplets and preserve scalability, we propose a four-stage process (Fig. 3) designed to bridge the gap between task complexity and model competence so as to fully utilize these automatic tools.
+
+Stage I: We develop a suite of modality-specific expert models for healthy tissue inpainting, built upon the FLUX.1-Fill-dev model. This strategy is based on the observation that generating healthy anatomical structures is more stable and predictable than generating lesions, as the former exhibits more tractable patterns and textures. For each modality, we curate a training dataset consisting exclusively of non-pathological samples from existing medical image repositories. Through parameter-efficient finetuning, these models learn to inpaint masked areas with high clinical accuracy. We further apply background restoration and edge blending to correct any unintended modifications made by the FLUX model outside the mask, ensuring the edited region blends seamlessly into the original image.
+
+Stage II: We leverage these expert models to modify lesion-bearing images (L) into their counterfactual 'healthy' results (H). Specifically, we fill the lesion area
+
+{4}------------------------------------------------
+
+in L using white pixels based on its ground-truth segmentation label. This masked image and its corresponding binary mask are then processed by the modality-matched expert model to synthesize H, where healthy tissue replaces lesion. Compared to distilling general-purpose generative models, our modality-specific approach not only restricts the high-variance generative process to a localized region to guarantee background consistency during the edit, but also ensures that tasks remain within the model's learned distribution, thereby significantly reducing hallucinations. Furthermore, unlike manual data collection from the internet, our approach provides superior scalability and efficiency.
+
+Stage III: We implement a rejection sampling mechanism for the generated 'healthy' images (H) to further enhance the data quality within the Modification tasks. For modalities that resemble natural images (e.g., endoscopy and dermoscopy), we prompt the Qwen3-VL-32B-Instruct model [\(Bai et al.,](#page-8-3) [2025a\)](#page-8-3) to filter out H that still contain lesions, exhibit artifacts, or are of low quality. For other modalities, we train separate nnUNet models [\(Isensee et al.,](#page-9-3) [2021\)](#page-9-3) for lesion segmentation and discard H where lesions remain detectable.
+
+Stage IV: Triplet combination. Using these high quality 'lesion-healthy' counterfactual pairs, we generate diverse Modification task data by swapping L and H from niche of input and output and varying the textual prompts P.
+
+#### 3.2.3. TRANSFORMATION
+
+220
+
+224
+
+226
+
+228 229 230
+
+233 234
+
+236
+
+254
+
+256
+
+264
+
+274
+
+Transformation tasks include a wide array of low-level medical image processing operations. Unlike the localized edits found in Perception and Modification categories, tasks in this category typically require a holistic transformation of the entire input image.
+
+The rule-based construction pipeline of Transformation tasks is shown in Fig. [3.](#page-3-0) From public repositories, we compile medical image pairs (I and O) representing 17 distinct transformation targets under four typical low-level vision categories. We then design specialized textual prompts P for each task to unify diverse medical image processing functions into a consistent image editing framework.
+
+#### 3.2.4. POST PROCESSING
+
+Prompt rephrasing. To enhance linguistic diversity, we utilize the Qwen-Max model to rephrase the prompts P for each data triplet. We also illustrate the linguistic diversity of our prompts via a word cloud in Figure [2\(](#page-2-1)b).
+
+Benchmark curation. The training and test split of source dataset are strictly followed during the construction of MieDB-100k to preclude any data leakage. Furthermore, we recruit three people with clinical background to manually evaluate and curate 3, 485 of the most representative
+
+samples characterized by high clinical fidelity from raw data test split to serve as the benchmark of MieDB-100k, and we keep their original image size to minimize information loss.
+
+Train split construction. For train split, we establish three resolution bins (128, 256, and 512) and resize images to their nearest corresponding value. To check the fidelity of training split, we randomly select 6, 000 triplets for clinician evaluation, and over 95% are viewed as high quality.
+
+### 3.3. MieDB-100k Evaluation
+
+We evaluate MieDB-100k through two distinct approaches: (1) verifiable metrics for the Perception and Transformation tasks, amenable to reward design in prevailing reinforcement learning algorithms [\(Shao et al.,](#page-11-4) [2024;](#page-11-4) [Liu et al.,](#page-10-2) [2025a\)](#page-10-2); and (2) more subjective evaluations for the Modification tasks, reflecting their greater complexity.
+
+#### 3.3.1. VERIFIABLE EVALUATION
+
+Localization Accuracy Metric. We use the DICE Score for evaluating the spatial overlapping performance in Perception tasks. Notably, reconstructing a binary mask from the colored regions of an edited image is mathematically feasible when the background image and overlay color are known, and we detail this process in App. [D.1.](#page-15-0) This procedure is applied to both model's output O<sup>M</sup> and the ground truth images O to derive the mask of model's perceptual region and the ground truth region for DICE calculation.
+
+To differentiate between models that accurately identify specific medical targets and those that merely generate coarsegrained masks, we further propose Perception Accuracy. Under this metric, a result is considered successful only if the DICE score exceeds a threshold of τ = 0.8. This metric allows us to analyze whether a model possesses the specialized medical knowledge required for image understanding.
+
+Image Similarity Metrics. We utilize PSNR and SSIM [\(Wang et al.,](#page-11-5) [2004\)](#page-11-5) to evaluate the similarity between the ground-truth and edited images at both the pixel and structural levels. For evaluations within the Perception perspective, we mask out the pixels corresponding to the groundtruth segmentation in both images. This allows us to specifically assess the model's ability to preserve the background while performing the requested edit.
+
+### 3.3.2. EVALUATION FOR MODIFICATION TASKS
+
+Vision-Language Model Rubric Scoring. Automating reliable assessments in the Modification tasks is inherently challenging, as edits are defined semantically and cannot be evaluated via deterministic rules. Existing benchmarks often leverage Vision-Language Models (VLMs) for this purpose, and we standardize the process and mitigate potential critic hallucinations by implementing a rubric-based
+
+{5}------------------------------------------------
+
+316
+
+328 329
+
+<span id="page-5-0"></span>*Table 2.* Overall result on MieDB-100k benchmark. P-ACC means Perception Accuracy; B-PSNR and B-SSIM mean only calculate PSNR and SSIM on background pixels respectively; Rubric-S stands for the Rubric Score from VLM and Pref-Rank stands for human preference ranking. Best values are marked in red while second bests are in blue.
+
+|                                        |      |       |       | Perception |        | Modification |           | Trasnformation |       |  |
+|----------------------------------------|------|-------|-------|------------|--------|--------------|-----------|----------------|-------|--|
+| Model Name                             | Size | DICE  | P-ACC | B-PSNR     | B-SSIM | Rubric-S     | Pref-Rank | PSNR           | SSIM  |  |
+| Open-Source                            |      |       |       |            |        |              |           |                |       |  |
+| SDXL-turbo (Sauer et al., 2024)        | 3.5B | 0.002 | 0.000 | 16.6       | 0.467  | 8.4          | 7.7       | 15.9           | 0.397 |  |
+| Bagel (Deng et al., 2025)              | 7B   | 0.263 | 0.069 | 13.9       | 0.620  | 34.4         | 6.2       | 12.7           | 0.442 |  |
+| OmniGen2 (Wu et al., 2025b)            | 7B   | 0.248 | 0.065 | 11.9       | 0.541  | 29.1         | 7.1       | 8.3            | 0.280 |  |
+| Step1X-Edit (Liu et al., 2025c)        | 21B  | 0.332 | 0.126 | 15.5       | 0.727  | 35.6         | 4.5       | 16.6           | 0.539 |  |
+| Qwen-Image-Edit (Wu et al., 2025a)     | 27B  | 0.387 | 0.153 | 15.4       | 0.722  | 32.2         | 5.5       | 18.9           | 0.606 |  |
+| FLUX.1-Kontext-dev (Labs et al., 2025) | 12B  | 0.341 | 0.126 | 15.4       | 0.701  | 37.8         | 6.2       | 17.9           | 0.543 |  |
+| OmniGen2-MIE (Ours)                    | 7B   | 0.831 | 0.737 | 28.1       | 0.917  | 65.9         | 1.4       | 22.6           | 0.685 |  |
+| Closed-Source                          |      |       |       |            |        |              |           |                |       |  |
+| GPT-Image-1 (OpenAI, 2025)             |      | 0.467 | 0.221 | 16.3       | 0.510  | 42.8         | 4.8       | 14.4           | 0.451 |  |
+| Nano Banana Pro (DeepMind, 2025a)      |      | 0.426 | 0.202 | 12.8       | 0.413  | 63.4         | 2.0       | 20.0           | 0.610 |  |
+| Imagen4 (DeepMind, 2025b)              |      | 0.142 | 0.000 | 8.9        | 0.210  | 19.7         | 7.4       | 7.9            | 0.174 |  |
+
+scoring system. Specifically, we provide the VLM with the input image I, edit instruction P, reference output O, and the model's generated result OM. Guided by the rubric, the VLM then performs a holistic evaluation of OM.
+
+We design a comprehensive scoring rubric (App. [D.2\)](#page-15-1) that assesses both the fulfillment of the editing intent and the model's ability to preserve the background. We utilize GPT-5.2 as an automated evaluator for this process, and map the final score to [0, 100].
+
+Human Preference Ranking. For each test case, we present the original triplet (I, P, O) and the outputs of all tested models simultaneously to evaluators, who are then asked to rank the various model-generated results according to their preference. By forcing this comparative ordering of all models, we are able to move beyond absolute quality scores and capture the relative strengths and weaknesses of current generative frameworks in a clinical setting. Specifically, we recruit 3 evaluators with clinical backgrounds to assess and rank the images edited by the benchmarked models, and compute the average ranking.
+
+# 4. Experiments
+
+# 4.1. Baselines
+
+We evaluate nine models on MieDB-100k, comprising six open-source models: Qwen-Image-Edit-2511 [\(Wu et al.,](#page-11-0) [2025a\)](#page-11-0), Bagel [\(Deng et al.,](#page-9-0) [2025\)](#page-9-0), OmniGen2 [\(Wu et al.,](#page-11-1) [2025b\)](#page-11-1), Step1X-Edit-v1p2 [\(Liu et al.,](#page-10-0) [2025c\)](#page-10-0), FLUX.1- Kontext-dev [\(Labs et al.,](#page-10-4) [2025\)](#page-10-4) and SDXL-turbo [\(Sauer](#page-10-3) [et al.,](#page-10-3) [2024\)](#page-10-3), plus three closed-source models: Nano Banana Pro [\(DeepMind,](#page-9-2) [2025a\)](#page-9-2), GPT-Image-1 [\(OpenAI,](#page-10-5) [2025\)](#page-10-5), and Imagen4 [\(DeepMind,](#page-9-4) [2025b\)](#page-9-4). We implement open-source models following their official inference settings.
+
+To validate the effectiveness of MieDB-100k, we finetune
+
+the OmniGen2 baseline on the training split and subject it to the same evaluation protocol as the other models. Specifically, we train the Diffusion Transformer (DiT) component for 20,000 iterations, employing a global batch size of 64 and a learning rate of 1e-4.
+
+## 4.2. Quantitative Results
+
+We report the benchmarking results of MieDB-100k in Tab. [2.](#page-5-0) First, the extremely low perception accuracy indicates that all tested models except ours fail to accurately comprehend and localize the specified anatomical targets under our evaluation protocol. Consequently, in Modification tasks, most of them are unable to generate clinically meaningful edits. Although a few models, such as Nano Banana Pro, achieve competitive results, we are indeed observing the 'right-for-the-wrong-reason' phenomenon, a risk that must be strictly avoided in clinical settings. Since the poor performance in Perception tasks expose their intrinsic lack of necessary medical knowledge, their edits cannot be justified. Notably in Transformation tasks, Nano Banana Pro also presents competitive results in certain cases. This may be attributed to the similarity between tasks like denoising or artifact removal and general-purpose low-level vision tasks, for which the model already possesses some capability [\(Zuo et al.,](#page-12-0) [2025\)](#page-12-0). Alternatively, it is possible that similar medical image processing tasks were included in its training set. Regardless, its absolute performance remains insufficient for practical clinical deployment. In summary, the benchmark result demonstrates that current multimodal generative model cannot meet the requirement of medical imaging editing.
+
+Conversely, after training on MieDB-100k, a standard baseline model can achieve superior medical editing capabilities. As shown in Tab. [2,](#page-5-0) the OmniGen2-MIE model delivers the best performance across all three edit-
+
+{6}------------------------------------------------
+
+<span id="page-6-0"></span>![](_page_6_Figure_1.jpeg)
+
+**[Edit Instruction]** Process the T1 MRI to mitigate motion effects without changing the appearance of brain tissue, tumors, or other clinical features.
+
+*Figure 4.* Qualitative editing result comparison.
+
+ing perspectives. The most significant improvements are observed in the Perception perspective, which demonstrate that MieDB-100k can effectively inject essential medical knowledge, thereby enhancing the interpretability of downstream editing tasks. Furthermore, in the Modification and Transformation tasks, where general-purpose editing abilities transfer more readily, our enhanced model still yields superior editing results compared to Nano Banana Pro, the SOTA multi-modal generative models. These findings highlight the pivotal role of our dataset in domain adaptation and
+
+330
+
+334
+
+336
+
+354
+
+356
+
+374
+
+376
+
+establish a foundation for the development of understandinggeneration unified medical models.
+
+#### 4.3. Qualitative Results
+
+Fig. [4](#page-6-0) presents qualitative editing results for several baseline models across the diverse modalities and tasks in MieDB-100k. These results demonstrate that the finetuned model exhibits an enhanced capability in both understanding and generation, allowing it to navigate the inherent complexities
+
+{7}------------------------------------------------
+
+385 387 388 389 390
+
+395
+
+396
+
+398
+
+399
+
+400
+
+401
+
+402
+
+403
+
+404
+
+405
+
+406
+
+407
+
+408
+
+409
+
+410 411
+
+412
+
+413
+
+414
+
+415
+
+416
+
+417
+
+418
+
+419
+
+420
+
+421
+
+422
+
+423
+
+424
+
+425
+
+426
+
+427
+
+428
+
+429
+
+430
+
+431
+
+432
+
+433
+
+434
+
+435
+
+436
+
+437
+
+438
+
+439
+
+<span id="page-7-0"></span>Table 3. Ablation study result on MieDB-100k. P stands for Perception, M stands for Modification, and T stands for Transformation. Best values are marked in red, second bests are in blue.
+
+|                     | Perce    | ption | Modification | Trasnformation |       |  |  |
+|---------------------|----------|-------|--------------|----------------|-------|--|--|
+| Training Data       | DICE ACC |       | RubricScore  | PSNR           | SSIM  |  |  |
+| Baseline (No train) | 0.248    | 0.065 | 29.1         | 8.3            | 0.280 |  |  |
+| P-only              | 0.833    | 0.740 | 37.8         | 19.7           | 0.631 |  |  |
+| M-only              | 0.001    | 0.000 | 57.5         | 19.8           | 0.631 |  |  |
+| T-only              | 0.034    | 0.000 | 15.0         | 23.7           | 0.702 |  |  |
+| MieDB-100k          | 0.831    | 0.737 | 65.9         | 22.6           | 0.685 |  |  |
+
+of medical image editing. Moreover, despite being explicitly prompted, even sophisticated closed-source models such as Nano Banana Pro fail to maintain background consistency in certain tasks. While their instruction-following proficiency stems from large-scale pre-training on natural image pairs, these capabilities tend to degrade when the distribution of medical modalities deviates significantly from the natural images seen during pre-training. To further study the impact of modality deviation, we conduct a modality-wise analysis in App. E.1, and the results prove our judgment. This observation underscores the necessity of a highly diverse dataset like MieDB-100k to equip models with the capacity to handle a vast range of medical imaging modalities.
+
+#### 4.4. Ablation Study
+
+To investigate the contribution of each task category, we conduct an ablation study by training models on individual perspective of MieDB-100k. We again utilize OmniGen2 as baseline model, following the training recipe described above while varying only the training data. As shown in Tab 3, each specialized model significantly outperforms the original baseline in its respective domain, validating the high information density and clinical relevance of our data. For the model trained on the full dataset, it achieves comparable or even better performance on all three perspectives, showing the effectiveness of the joint training. More importantly, we observe significant performance improvement in the Modification perspective, demonstrating visual understanding ability has the potential to enhance visual generation ability. In summary, the ablation study shows that MieDB-100k can provide a synergistic training signal, enabling the development of a versatile model capable of handling diverse medical editing tasks simultaneously.
+
+### <span id="page-7-2"></span>4.5. Generalization Test
+
+To further investigate the cross-task synergy and the resulting generalization capabilities, we conduct an out-ofdistribution (OOD) editing experiment. Specifically, we target 'bone metastasis', a medical target included in Perception tasks but strictly excluded from the Modification training data. We then prompt the OmniGen2-MIE model
+
+<span id="page-7-1"></span>![](_page_7_Figure_10.jpeg)
+
+Figure 5. Generalization test assessment. (a) and (b): Edit samples output by different models on bone metastasis addition (a) and removal (b) tasks. Red bounding boxes are added post-hoc to highlight the edited regions for visualization; (c): Quantitative assessments following the recipe of Modification task evaluation.
+
+to perform metastasis addition and removal in CT scans.
+
+As shown in Fig. 5, OmniGen2-MIE significantly outperforms OmniGen2 on this unseen task, demonstrating that our unified training on MieDB-100k can enhance the model's generalization capabilities across editing tasks. We also observe that Nano Banana Pro achieves the best OOD editing performance, marginally surpassing OmniGen2-MIE. We attribute this performance to the utilization of massive-scale general and medical editing data, which further underscores the necessity of scaling up medical editing data.
+
+#### 5. Conclusion
+
+In this paper, we introduce MieDB-100k, a large-scale and diverse dataset for text-guided medical image editing. By unifying Perception, Modification, and Transformation tasks into the paradigm of editing, our dataset bridges the gap between medical image understanding and generation. We develop a robust curation pipeline, integrating modalityspecific expert models with rule-based synthesis, and enforce rigorous manual quality control to ensure clinical fidelity across all data. Extensive benchmarking demonstrates that model trained on MieDB-100k consistently outperform both SOTA open-source and proprietary multimodal models while exhibiting exceptional generalization to unseen clinical tasks. Our work thus provides the data foundation to support the development and evaluation of multimodal generative models for clinical applications.
+
+{8}------------------------------------------------
+
+# Impact Statement
+
+This paper presents work whose goal is to advance the field of Machine Learning. There are many potential societal consequences of our work, none which we feel must be specifically highlighted here.
+
+# References
+
+- <span id="page-8-9"></span>AAPM. Low dose ct grand challenge, 2016. URL [https:](https://www.aapm.org/grandchallenge/lowdosect/#trainingData) [//www.aapm.org/grandchallenge/lowdos](https://www.aapm.org/grandchallenge/lowdosect/#trainingData) [ect/#trainingData](https://www.aapm.org/grandchallenge/lowdosect/#trainingData). Accessed: 2026-01-14.
+- <span id="page-8-14"></span>Abdi, A. H., Kasaei, S., and Mehdizadeh, M. Automatic segmentation of mandible in panoramic x-ray. *Journal of Medical Imaging*, 2(4):044003–044003, 2015.
+- <span id="page-8-10"></span>Ahmed, Z., Panhwar, S. Q., Baqai, A., Umrani, F. A., Ahmed, M., and Khan, A. Deep learning based automated detection of intraretinal cystoid fluid. *International Journal of Imaging Systems and Technology*, 32(3):902–917, 2022.
+- <span id="page-8-4"></span>Al-Dhabyani, W., Gomaa, M., Khaled, H., and Fahmy, A. Dataset of breast ultrasound images. *Data in brief*, 28: 104863, 2020.
+- <span id="page-8-15"></span>Ali, S., Ghatwary, N., Jha, D., Isik-Polat, E., Polat, G., Yang, C., Li, W., Galdran, A., Ballester, M.-A. G., Thambawita, ´ V., et al. Assessing generalisability of deep learningbased polyp detection and segmentation methods through a computer vision challenge. *Scientific Reports*, 14(1): 2032, 2024.
+- <span id="page-8-7"></span>Angiographics. Chuac dataset. URL [https://figsha](https://figshare.com/s/4d24cf3d14bc901a94bf) [re.com/s/4d24cf3d14bc901a94bf](https://figshare.com/s/4d24cf3d14bc901a94bf). Accessed: 2026-01-14.
+- <span id="page-8-13"></span>Antonelli, M., Reinke, A., Bakas, S., Farahani, K., Kopp-Schneider, A., Landman, B. A., Litjens, G., Menze, B., Ronneberger, O., Summers, R. M., et al. The medical segmentation decathlon. *Nature communications*, 13(1): 4128, 2022.
+- <span id="page-8-16"></span>Aysen, D., Muhammad, C., Tawsifur, R., Amith, K., Yazan, Q., Mete, A., Anas M., T., and serkan, k. Qata-cov19 dataset, 2024. URL [https://www.kaggle.com/d](https://www.kaggle.com/datasets/aysendegerli/qatacov19-dataset) [atasets/aysendegerli/qatacov19-datas](https://www.kaggle.com/datasets/aysendegerli/qatacov19-dataset) [et](https://www.kaggle.com/datasets/aysendegerli/qatacov19-dataset). Accessed: 2026-01-14.
+- <span id="page-8-3"></span>Bai, S., Cai, Y., Chen, R., Chen, K., Chen, X., Cheng, Z., Deng, L., Ding, W., Gao, C., Ge, C., Ge, W., Guo, Z., Huang, Q., Huang, J., Huang, F., Hui, B., Jiang, S., Li, Z., Li, M., Li, M., Li, K., Lin, Z., Lin, J., Liu, X., Liu, J., Liu, C., Liu, Y., Liu, D., Liu, S., Lu, D., Luo, R., Lv, C., Men, R., Meng, L., Ren, X., Ren, X., Song, S., Sun, Y., Tang, J., Tu, J., Wan, J., Wang, P., Wang, P.,
+
+- Wang, Q., Wang, Y., Xie, T., Xu, Y., Xu, H., Xu, J., Yang, Z., Yang, M., Yang, J., Yang, A., Yu, B., Zhang, F., Zhang, H., Zhang, X., Zheng, B., Zhong, H., Zhou, J., Zhou, F., Zhou, J., Zhu, Y., and Zhu, K. Qwen3-vl technical report. *Preprint at arXiv*, 2025a. URL [https:](https://arxiv.org/abs/2511.21631) [//arxiv.org/abs/2511.21631](https://arxiv.org/abs/2511.21631).
+- <span id="page-8-2"></span>Bai, S., Chen, K., Liu, X., Wang, J., Ge, W., Song, S., Dang, K., Wang, P., Wang, S., Tang, J., Zhong, H., Zhu, Y., Yang, M., Li, Z., Wan, J., Wang, P., Ding, W., Fu, Z., Xu, Y., Ye, J., Zhang, X., Xie, T., Cheng, Z., Zhang, H., Yang, Z., Xu, H., and Lin, J. Qwen2.5-vl technical report. *Preprint at arXiv*, 2025b. URL [https://arxiv.or](https://arxiv.org/abs/2502.13923) [g/abs/2502.13923](https://arxiv.org/abs/2502.13923).
+- <span id="page-8-1"></span>Brooks, T., Holynski, A., and Efros, A. A. Instructpix2pix: Learning to follow image editing instructions. In *Proceedings of the IEEE/CVF conference on computer vision and pattern recognition*, pp. 18392–18402, 2023.
+- <span id="page-8-12"></span>Buda, M., Saha, A., and Mazurowski, M. A. Association of genomic subtypes of lower-grade gliomas with shape features automatically extracted by a deep learning algorithm. *Computers in biology and medicine*, 109:218–225, 2019.
+- <span id="page-8-5"></span>Caicedo, J. C., Goodman, A., Karhohs, K. W., Cimini, B. A., Ackerman, J., Haghighi, M., Heng, C., Becker, T., Doan, M., McQuin, C., et al. Nucleus segmentation across imaging experiments: the 2018 data science bowl. *Nature methods*, 16(12):1247–1253, 2019.
+- <span id="page-8-6"></span>Carballal, A., Novoa, F. J., Fernandez-Lozano, C., Garc´ıa-Guimaraes, M., Aldama-Lopez, G., Calvi ´ no-Santos, R., ˜ Vazquez-Rodriguez, J. M., and Pazos, A. Automatic multiscale vascular image segmentation algorithm for coronary angiography. *Biomedical Signal Processing and Control*, 46:1–9, 2018.
+- <span id="page-8-11"></span>Cervantes-Sanchez, F., Cruz-Aceves, I., Hernandez-Aguirre, A., Hernandez-Gonzalez, M. A., and Solorio-Meza, S. E. Automatic segmentation of coronary arteries in x-ray angiograms using multiscale analysis and artificial neural networks. *Applied Sciences*, 9(24):5507, 2019.
+- <span id="page-8-0"></span>Chen, Z. and Feng, M. Med-banana-50k: A cross-modality large-scale dataset for text-guided medical image editing. *Preprint at arXiv*, 2025. URL [https://arxiv.or](https://arxiv.org/abs/2511.00801) [g/abs/2511.00801](https://arxiv.org/abs/2511.00801).
+- <span id="page-8-8"></span>Chowdhury, M. E., Rahman, T., Khandakar, A., Mazhar, R., Kadir, M. A., Mahbub, Z. B., Islam, K. R., Khan, M. S., Iqbal, A., Al Emadi, N., et al. Can ai help in screening viral and covid-19 pneumonia? *Ieee Access*, 8: 132665–132676, 2020.
+
+{9}------------------------------------------------
+
+<span id="page-9-8"></span>495 496 497 498 499 500 Codella, N., Rotemberg, V., Tschandl, P., Celebi, M. E., Dusza, S., Gutman, D., Helba, B., Kalloo, A., Liopyris, K., Marchetti, M., et al. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic). *arXiv preprint arXiv:1902.03368*, 2019.
+
+<span id="page-9-13"></span>501
+
+504
+
+513 514
+
+516
+
+518 519 520
+
+524 525 526
+
+528 529 530
+
+534
+
+<span id="page-9-4"></span>536
+
+- Dao, B. T., Nguyen, T. V., Pham, H. H., and Nguyen, H. Q. Phase recognition in contrast-enhanced ct scans based on deep learning and random sampling. *Medical Physics*, 49 (7):4518–4528, 2022.
+- <span id="page-9-5"></span>de Verdier, M. C., Saluja, R., Gagnon, L., LaBella, D., Baid, U., Tahon, N. H., Foltyn-Dumitru, M., Zhang, J., Alafif, M., Baig, S., Chang, K., D'Anna, G., Deptula, L., Gupta, D., Haider, M. A., Hussain, A., Iv, M., Kontzialis, M., Manning, P., Moodi, F., Nunes, T., Simon, A., Sollmann, N., Vu, D., Adewole, M., Albrecht, J., Anazodo, U., Chai, R., Chung, V., Faghani, S., Farahani, K., Kazerooni, A. F., Iglesias, E., Kofler, F., Li, H., Linguraru, M. G., Menze, B., Moawad, A. W., Velichko, Y., Wiestler, B., Altes, T., Basavasagar, P., Bendszus, M., Brugnara, G., Cho, J., Dhemesh, Y., Fields, B. K. K., Garrett, F., Gass, J., Hadjiiski, L., Hattangadi-Gluth, J., Hess, C., Houk, J. L., Isufi, E., Layfield, L. J., Mastorakos, G., Mongan, J., Nedelec, P., Nguyen, U., Oliva, S., Pease, M. W., Rastogi, A., Sinclair, J., Smith, R. X., Sugrue, L. P., Thacker, J., Vidic, I., Villanueva-Meyer, J., White, N. S., Aboian, M., Conte, G. M., Dale, A., Sabuncu, M. R., Seibert, T. M., Weinberg, B., Abayazeed, A., Huang, R., Turk, S., Rauschecker, A. M., Farid, N., Vollmuth, P., Nada, A., Bakas, S., Calabrese, E., and Rudie, J. D. The 2024 brain tumor segmentation (brats) challenge: Glioma segmentation on post-treatment mri. *Preprint at arXiv*, 2024. URL [https://arxiv.org/abs/2405.1](https://arxiv.org/abs/2405.18368) [8368](https://arxiv.org/abs/2405.18368).
+- <span id="page-9-2"></span>DeepMind, G. Gemini 3 pro: High-precision multimodal reasoning, 2025a. URL [https://deepmind.goo](https://deepmind.google/models/gemini-image/pro/) [gle/models/gemini-image/pro/](https://deepmind.google/models/gemini-image/pro/). Accessed: 2026-01-14.
+- DeepMind, G. Imagen 4.0: Model documentation and generation guide, 2025b. URL [https://docs.c](https://docs.cloud.google.com/vertex-ai/generative-ai/docs/models/imagen/4-0-generate) [loud.google.com/vertex-ai/generative](https://docs.cloud.google.com/vertex-ai/generative-ai/docs/models/imagen/4-0-generate) [-ai/docs/models/imagen/4-0-generate](https://docs.cloud.google.com/vertex-ai/generative-ai/docs/models/imagen/4-0-generate). Accessed: 2026-01-14.
+- <span id="page-9-0"></span>Deng, C., Zhu, D., Li, K., Gou, C., Li, F., Wang, Z., Zhong, S., Yu, W., Nie, X., Song, Z., Shi, G., and Fan, H. Emerging properties in unified multimodal pretraining. *Preprint at arXiv*, 2025. URL [https:](https://arxiv.org/abs/2505.14683) [//arxiv.org/abs/2505.14683](https://arxiv.org/abs/2505.14683).
+- <span id="page-9-14"></span>Dietler, N., Minder, M., Gligorovski, V., Economou, A. M., Joly, D. A. H. L., Sadeghi, A., Chan, C. H. M., Kozinski, ´
+
+- M., Weigert, M., Bitbol, A.-F., et al. A convolutional neural network segments yeast microscopy images with high accuracy. *Nature communications*, 11(1):5723, 2020.
+- <span id="page-9-12"></span>Duong, V. H., Vu, H., Phan, H. D., Nguyen, D. Q., Pham, D. H., Le, Q. T., Nguyen, B. S., Do, T. D., Dinh, V. S., Nguyen, T. C., et al. Thyroidxl: Advancing thyroid nodule diagnosis with an expert-labeled, pathology-validated dataset. In *International Conference on Medical Image Computing and Computer-Assisted Intervention*, pp. 616– 626. Springer, 2025.
+- <span id="page-9-10"></span>Fang, H., Li, F., Wu, J., Fu, H., Sun, X., Son, J., Yu, S., Zhang, M., Yuan, C., Bian, C., Lei, B., Zhao, B., Xu, X., Li, S., Fumero, F., Sigut, J., Almubarak, H., Bazi, Y., Guo, Y., Zhou, Y., Baid, U., Innani, S., Guo, T., Yang, J., Orlando, J. I., Bogunovic, H., Zhang, X., and Xu, Y. ´ Refuge2 challenge: A treasure trove for multi-dimension analysis and evaluation in glaucoma screening. *Preprint at arXiv*, 2022. URL [https://arxiv.org/abs/](https://arxiv.org/abs/2202.08994) [2202.08994](https://arxiv.org/abs/2202.08994).
+- <span id="page-9-11"></span>Garcia-Peraza-Herrera, L. C., Fidon, L., D'Ettorre, C., Stoyanov, D., Vercauteren, T., and Ourselin, S. Image compositing for segmentation of surgical tools without manual annotations. *IEEE transactions on medical imaging*, 40(5):1450–1460, 2021.
+- <span id="page-9-7"></span>Gutman, D., Codella, N. C., Celebi, E., Helba, B., Marchetti, M., Mishra, N., and Halpern, A. Skin lesion analysis toward melanoma detection: A challenge at the international symposium on biomedical imaging (isbi) 2016, hosted by the international skin imaging collaboration (isic). *arXiv preprint arXiv:1605.01397*, 2016.
+- <span id="page-9-6"></span>Haneda, E., Peters, N., Zhang, J., Karageorgos, G., Xia, W., Paganetti, H., Wang, G., Guo, Y., Ma, J., Park, H. S., et al. Aapm ct metal artifact reduction grand challenge. *Medical physics*, 52(10):e70050, 2025.
+- <span id="page-9-1"></span>Huang, Z., Zheng, D., Zou, C., Liu, R., Wang, X., Ji, K., Chai, W., Sun, J., Wang, L., Lv, Y., Huang, T., Liu, J., Guo, Q., Yang, M., Chen, J., and Zhou, J. Ming-univision: Joint image understanding and generation with a unified continuous tokenizer. *Preprint at arXiv*, 2025. URL <https://arxiv.org/abs/2510.06590>.
+- <span id="page-9-3"></span>Isensee, F., Jaeger, P. F., Kohl, S. A., Petersen, J., and Maier-Hein, K. H. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. *Nature methods*, 18(2):203–211, 2021.
+- <span id="page-9-9"></span>Jha, D., Smedsrud, P. H., Riegler, M. A., Halvorsen, P., De Lange, T., Johansen, D., and Johansen, H. D. Kvasirseg: A segmented polyp dataset. In *International conference on multimedia modeling*, pp. 451–462. Springer, 2019.
+
+{10}------------------------------------------------
+
+<span id="page-10-20"></span>550 Kus¸, Z. and Aydin, M. Medsegbench: A comprehensive benchmark for medical image segmentation in diverse data modalities. *Scientific Data*, 11(1):1283, 2024.
+
+<span id="page-10-4"></span>554
+
+556
+
+<span id="page-10-2"></span>574
+
+576
+
+<span id="page-10-6"></span>594
+
+596
+
+- Labs, B. F., Batifol, S., Blattmann, A., Boesel, F., Consul, S., Diagne, C., Dockhorn, T., English, J., English, Z., Esser, P., Kulal, S., Lacey, K., Levi, Y., Li, C., Lorenz, D., Muller, J., Podell, D., Rombach, R., Saini, H., Sauer, ¨ A., and Smith, L. Flux.1 kontext: Flow matching for in-context image generation and editing in latent space. *Preprint at arXiv*, 2025. URL [https://arxiv.or](https://arxiv.org/abs/2506.15742) [g/abs/2506.15742](https://arxiv.org/abs/2506.15742).
+- <span id="page-10-15"></span>Lei, W., Chen, H., Zhang, Z., Luo, L., Xiao, Q., Gu, Y., Gao, P., Jiang, Y., Wang, C., Wu, G., Xu, T., Zhang, Y., Rajpurkar, P., Zhang, X., Zhang, S., and Wang, Z. A synthetic data-driven radiology foundation model for pantumor clinical diagnosis. *Preprint at arXiv*, 2025. URL <https://arxiv.org/abs/2502.06171>.
+- <span id="page-10-16"></span>Litjens, G., Toth, R., Van De Ven, W., Hoeks, C., Kerkstra, S., Van Ginneken, B., Vincent, G., Guillard, G., Birbeck, N., Zhang, J., et al. Evaluation of prostate segmentation algorithms for mri: the promise12 challenge. *Medical image analysis*, 18(2):359–373, 2014.
+- Liu, J., Liu, G., Liang, J., Li, Y., Liu, J., Wang, X., Wan, P., Zhang, D., and Ouyang, W. Flow-grpo: Training flow matching models via online rl. *arXiv preprint at arXiv*, 2025a. URL [https://arxiv.org/abs/2505.0](https://arxiv.org/abs/2505.05470) [5470](https://arxiv.org/abs/2505.05470).
+- <span id="page-10-1"></span>Liu, M., He, Z., Fan, Z., Wang, Q., and Fung, Y. R. Medebench: Diagnosing reliability in text-guided medical image editing. In *Findings of the Association for Computational Linguistics: EMNLP 2025*, pp. 767–791, 2025b.
+- <span id="page-10-0"></span>Liu, S., Han, Y., Xing, P., Yin, F., Wang, R., Cheng, W., Liao, J., Wang, Y., Fu, H., Han, C., Li, G., Peng, Y., Sun, Q., Wu, J., Cai, Y., Ge, Z., Ming, R., Xia, L., Zeng, X., Zhu, Y., Jiao, B., Zhang, X., Yu, G., and Jiang, D. Step1xedit: A practical framework for general image editing. *Preprint at arXiv*, 2025c. URL [https://arxiv.or](https://arxiv.org/abs/2504.17761) [g/abs/2504.17761](https://arxiv.org/abs/2504.17761).
+- Ljosa, V., Sokolnicki, K. L., and Carpenter, A. E. Annotated high-throughput microscopy image sets for validation. *Nature methods*, 9(7):637, 2012.
+- <span id="page-10-12"></span>Lu, Y., Zhou, M., Zhi, D., Zhou, M., Jiang, X., Qiu, R., Ou, Z., Wang, H., Qiu, D., Zhong, M., et al. The jnu-ifm dataset for segmenting pubic symphysis-fetal head. *Data in brief*, 41:107904, 2022.
+- <span id="page-10-8"></span>Mathieu, G., Bachir, E. D., et al. Brifiseg: a deep learningbased method for semantic and instance segmentation
+
+- of nuclei in brightfield images. *Preprint at arXiv*, 2022. URL <https://arxiv.org/abs/2211.03072>.
+- <span id="page-10-18"></span>Montoya, A. Ultrasound nerve segmentation, 2026. URL [https://www.kaggle.com/competitions/](https://www.kaggle.com/competitions/ultrasound-nerve-segmentation) [ultrasound-nerve-segmentation](https://www.kaggle.com/competitions/ultrasound-nerve-segmentation). Accessed: 2026-01-14.
+- <span id="page-10-11"></span>Morozov, S. P., Andreychenko, A. E., Pavlov, N. A., Vladzymyrskyy, A., Ledikhova, N. V., Gombolevskiy, V. A., Blokhin, I. A., Gelezhe, P. B., Gonchar, A., and Chernina, V. Y. Mosmeddata: Chest ct scans with covid-19 related findings dataset. *arXiv preprint arXiv:2005.06465*, 2020.
+- <span id="page-10-14"></span>Narai, ´ A., Hermann, P., Auer, T., Kemenczky, P., Szalma, ´ J., Homolya, I., Somogyi, E., Vakli, P., Weiss, B., and Vidnyanszky, Z. Movement-related artefacts (mr-art) ´ dataset of matched motion-corrupted and clean structural mri brain scans. *Scientific data*, 9(1):630, 2022.
+- <span id="page-10-17"></span>Naylor, P., Lae, M., Reyal, F., and Walter, T. Segmentation ´ of nuclei in histopathology images by deep regression of the distance map. *IEEE transactions on medical imaging*, 38(2):448–459, 2018.
+- <span id="page-10-7"></span>Ngoc Lan, P., An, N. S., Hang, D. V., Long, D. V., Trung, T. Q., Thuy, N. T., and Sang, D. V. Neounet: Towards accurate colon polyp segmentation and neoplasm detection. In *International Symposium on Visual Computing*, pp. 15–28. Springer, 2021.
+- <span id="page-10-5"></span>OpenAI. Gpt image 1 model documentation, 2025. URL [https://platform.openai.com/docs/mod](https://platform.openai.com/docs/models/gpt-image-1) [els/gpt-image-1](https://platform.openai.com/docs/models/gpt-image-1). Accessed: 2026-01-14.
+- <span id="page-10-10"></span>Pandey, N. Chest xray masks and labels, 2025. URL [ht](https://www.kaggle.com/datasets/nikhilpandey360/chest-xray-masks-and-labels) [tps://www.kaggle.com/datasets/nikhil](https://www.kaggle.com/datasets/nikhilpandey360/chest-xray-masks-and-labels) [pandey360/chest-xray-masks-and-labels](https://www.kaggle.com/datasets/nikhilpandey360/chest-xray-masks-and-labels). Accessed: 2026-01-14.
+- <span id="page-10-19"></span>pazhoulab. Low-dose ct reconstruction contest, 2024. URL [https://iacc.pazhoulab-huangpu.com/c](https://iacc.pazhoulab-huangpu.com/contestdetail?id=667e0c687ff47da8cc827679&award=1,000,000) [ontestdetail?id=667e0c687ff47da8cc82](https://iacc.pazhoulab-huangpu.com/contestdetail?id=667e0c687ff47da8cc827679&award=1,000,000) [7679&award=1,000,000](https://iacc.pazhoulab-huangpu.com/contestdetail?id=667e0c687ff47da8cc827679&award=1,000,000). Accessed: 2026-01-14.
+- <span id="page-10-9"></span>Polo, M. Chest ct segmentation, 2025. URL [https:](https://www.kaggle.com/datasets/polomarco/chest-ct-segmentation) [//www.kaggle.com/datasets/polomarco/](https://www.kaggle.com/datasets/polomarco/chest-ct-segmentation) [chest-ct-segmentation](https://www.kaggle.com/datasets/polomarco/chest-ct-segmentation). Accessed: 2026-01-14.
+- <span id="page-10-13"></span>Porwal, P., Pachade, S., Kamble, R., Kokare, M., Deshmukh, G., Sahasrabuddhe, V., and Meriaudeau, F. Indian diabetic retinopathy image dataset (idrid): a database for diabetic retinopathy screening research. *Data*, 3(3):25, 2018.
+- <span id="page-10-3"></span>Sauer, A., Lorenz, D., Blattmann, A., and Rombach, R. Adversarial diffusion distillation. In *European Conference on Computer Vision*, pp. 87–103. Springer, 2024.
+
+{11}------------------------------------------------
+
+<span id="page-11-4"></span>Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., Bi, X., Zhang, H., Zhang, M., Li, Y. K., Wu, Y., and Guo, D. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. *Preprint at arXiv*, 2024. URL <https://arxiv.org/abs/2402.03300>.
+
+- <span id="page-11-16"></span>Song, Y., Zheng, J., Lei, L., Ni, Z., Zhao, B., and Hu, Y. Ct2us: Cross-modal transfer learning for kidney segmentation in ultrasound images with synthesized data. *Ultrasonics*, 122:106706, 2022.
+- <span id="page-11-8"></span>Spahn, C., Gomez-de Mariscal, E., Laine, R. F., Pereira, ´ P. M., von Chamier, L., Conduit, M., Pinho, M. G., Jacquemet, G., Holden, S., Heilemann, M., et al. Deepbacs for multi-task bacterial image analysis using opensource deep learning approaches. *Communications Biology*, 5(1):688, 2022.
+- <span id="page-11-9"></span>Staal, J., Abramoff, M. D., Niemeijer, M., Viergever, M. A., ` and Van Ginneken, B. Ridge-based vessel segmentation in color images of the retina. *IEEE transactions on medical imaging*, 23(4):501–509, 2004.
+- <span id="page-11-7"></span>Tahir, A. M., Chowdhury, M. E., Khandakar, A., Rahman, T., Qiblawey, Y., Khurshid, U., Kiranyaz, S., Ibtehaz, N., Rahman, M. S., Al-Maadeed, S., et al. Covid-19 infection localization and severity grading from chest x-ray images. *Computers in biology and medicine*, 139:105002, 2021.
+- <span id="page-11-2"></span>Tong, S., Fan, D., Li, J., Xiong, Y., Chen, X., Sinha, K., Rabbat, M., LeCun, Y., Xie, S., and Liu, Z. Metamorph: Multimodal understanding and generation via instruction tuning. In *Proceedings of the IEEE/CVF International Conference on Computer Vision*, pp. 17001–17012, 2025.
+- <span id="page-11-10"></span>Van Valen, D. A., Kudo, T., Lane, K. M., Macklin, D. N., Quach, N. T., DeFelice, M. M., Maayan, I., Tanouchi, Y., Ashley, E. A., and Covert, M. W. Deep learning automates the quantitative analysis of individual cells in live-cell imaging experiments. *PLoS computational biology*, 12(11):e1005177, 2016.
+- <span id="page-11-12"></span>Verma, R., Kumar, N., Patil, A., Kurian, N. C., Rane, S., Graham, S., Vu, Q. D., Zwager, M., Raza, S. E. A., Rajpoot, N., et al. Monusac2020: A multi-organ nuclei segmentation and classification challenge. *IEEE Transactions on Medical Imaging*, 40(12):3413–3423, 2021.
+- <span id="page-11-17"></span>Vision and Lab, I. P. Skin cancer detection, 2024. URL [https://vip.uwaterloo.ca/skin-cance](https://vip.uwaterloo.ca/skin-cancer-detection/) [r-detection/](https://vip.uwaterloo.ca/skin-cancer-detection/). Accessed: 2026-01-14.
+- <span id="page-11-6"></span>Vitale, S., Orlando, J. I., Iarussi, E., and Larrabide, I. Improving realism in patient-specific abdominal ultrasound simulation using cyclegans. *International journal of computer assisted radiology and surgery*, 15(2):183–192, 2020.
+
+- <span id="page-11-11"></span>Wang, Y. and Shi, F. KMAR-50K. Mendeley Data, V6, 2025. URL [https://doi.org/10.17632/xw7](https://doi.org/10.17632/xw7mrg7ntg.6) [mrg7ntg.6](https://doi.org/10.17632/xw7mrg7ntg.6). Accessed: 2026-01-14.
+- <span id="page-11-5"></span>Wang, Z., Bovik, A. C., Sheikh, H. R., and Simoncelli, E. P. Image quality assessment: from error visibility to structural similarity. *IEEE transactions on image processing*, 13(4):600–612, 2004.
+- <span id="page-11-15"></span>Wasserthal, J., Breit, H.-C., Meyer, M. T., Pradella, M., Hinck, D., Sauter, A. W., Heye, T., Boll, D. T., Cyriac, J., Yang, S., et al. Totalsegmentator: robust segmentation of 104 anatomic structures in ct images. *Radiology: Artificial Intelligence*, 5(5):e230024, 2023.
+- <span id="page-11-0"></span>Wu, C., Li, J., Zhou, J., Lin, J., Gao, K., Yan, K., ming Yin, S., Bai, S., Xu, X., Chen, Y., Chen, Y., Tang, Z., Zhang, Z., Wang, Z., Yang, A., Yu, B., Cheng, C., Liu, D., Li, D., Zhang, H., Meng, H., Wei, H., Ni, J., Chen, K., Cao, K., Peng, L., Qu, L., Wu, M., Wang, P., Yu, S., Wen, T., Feng, W., Xu, X., Wang, Y., Zhang, Y., Zhu, Y., Wu, Y., Cai, Y., and Liu, Z. Qwen-image technical report. *Preprint at arXiv*, 2025a. URL [https://arxiv.org/abs/25](https://arxiv.org/abs/2508.02324) [08.02324](https://arxiv.org/abs/2508.02324).
+- <span id="page-11-1"></span>Wu, C., Zheng, P., Yan, R., Xiao, S., Luo, X., Wang, Y., Li, W., Jiang, X., Liu, Y., Zhou, J., et al. Omnigen2: Exploration to advanced multimodal generation. *Preprint at arXiv*, 2025b. URL [https://arxiv.org/abs/](https://arxiv.org/abs/2506.18871) [2506.18871](https://arxiv.org/abs/2506.18871).
+- <span id="page-11-3"></span>Yang, J., Yan, Y., Wu, G., Wang, Y., Liang, R., Jiang, X., Wan, X., Fan, F., Zhang, Y., Qin, F., and Wang, C. Medgen-bench: Contextually entangled benchmark for open-ended multimodal medical generation. *Preprint at arXiv*, 2025. URL [https://arxiv.org/abs/25](https://arxiv.org/abs/2511.13135) [11.13135](https://arxiv.org/abs/2511.13135).
+- <span id="page-11-13"></span>Yang, L., Ghosh, R. P., Franklin, J. M., Chen, S., You, C., Narayan, R. R., Melcher, M. L., and Liphardt, J. T. Nuset: A deep learning tool for reliably separating and analyzing crowded cells. *PLoS computational biology*, 16(9):e1008193, 2020.
+- <span id="page-11-14"></span>Younus Akon, Z. S. S. Paired ct and mri dataset for medical applications, 2025. URL [https://www.kaggle.c](https://www.kaggle.com/datasets/29c3607295965ebb030f2d158fec487412d84c82528dd44f8ef956aef35541aa) [om/datasets/29c3607295965ebb030f2d15](https://www.kaggle.com/datasets/29c3607295965ebb030f2d158fec487412d84c82528dd44f8ef956aef35541aa) [8fec487412d84c82528dd44f8ef956aef355](https://www.kaggle.com/datasets/29c3607295965ebb030f2d158fec487412d84c82528dd44f8ef956aef35541aa) [41aa](https://www.kaggle.com/datasets/29c3607295965ebb030f2d158fec487412d84c82528dd44f8ef956aef35541aa). Accessed: 2026-01-14.
+- <span id="page-11-18"></span>Zhang, S., Zhang, Q., Zhang, S., Liu, X., Yue, J., Lu, M., Xu, H., Yao, J., Wei, X., Cao, J., et al. A generalist foundation model and database for open-world medical image segmentation. *Nature Biomedical Engineering*, pp. 1–16, 2025.
+
+{12}------------------------------------------------
+
+<span id="page-12-1"></span>Zheng, X., Wang, Y., Wang, G., and Liu, J. Fast and robust segmentation of white blood cell images by selfsupervised learning. *Micron*, 107:55–71, 2018.
+
+ <span id="page-12-0"></span>Zuo, J., Deng, H., Zhou, H., Zhu, J., Zhang, Y., Zhang, Y., Yan, Y., Huang, K., Chen, W., Deng, Y., Jin, R., Sang, N., and Gao, C. Is nano banana pro a low-level vision all-rounder? a comprehensive evaluation on 14 tasks and 40 datasets. *Preprint at arXiv*, 2025. URL <https://arxiv.org/abs/2512.15110>.
+
+{13}------------------------------------------------
+
+# <span id="page-13-0"></span>A. Data Sources
+
+716
+
+718 719 720
+
+724
+
+726
+
+734
+
+736
+
+754
+
+756
+
+Our work is compiled based on following public medical image repositories:
+
+*Table 4.* Summary of public medical datasets utilized in the construction of MieDB-100k. The columns #Train and #Benchmark denote the number of samples allocated to our training and benchmark splits respectively from each source dataset.
+
+| DatasetName                                           | #Train | #Benchmark | Modality       |
+|-------------------------------------------------------|--------|------------|----------------|
+| AbdomenUS (Vitale et al., 2020)                       | 569    | 62         | Ultrasound     |
+| Bbbc010 (Ljosa et al., 2012)                          | 70     | 20         | Microscopy     |
+| Bkai-Igh (Ngoc Lan et al., 2021)                      | 700    | 81         | Endoscopy      |
+| Brats-gli (de Verdier et al., 2024)                   | 1529   | 80         | MRI            |
+| BriFiSeg (Mathieu et al., 2022)                       | 1005   | 40         | Microscopy     |
+| BUSI (Al-Dhabyani et al., 2020)                       | 452    | 80         | Ultrasound     |
+| CellNuclei (Caicedo et al., 2019)                     | 469    | 51         | Microscopy     |
+| ChaseDB1 (Carballal et al., 2018)                     | 19     | 7          | Fundus         |
+| Chest-ct-segmentation (Polo, 2025)                    | 278    | 19         | CT             |
+| Chest-xray-masks-and-labels (Pandey, 2025)            | 666    | 32         | Xray           |
+| CHUAC (Angiographics)                                 | 17     | 5          | Fundus         |
+| COVID-19 Radiography Dataset (Chowdhury et al., 2020) | 2010   | 95         | Xray           |
+| COVID-19-CT-SCAN-Lesion (Morozov et al., 2020)        | 255    | 15         | CT             |
+| CovidQU (Tahir et al., 2021)                          | 5684   | 122        | Xray           |
+| CT MAR (Haneda et al., 2025)                          | 1595   | 82         | CT             |
+| CT-Low-Dose-Reconstruction (AAPM, 2016)               | 867    | 51         | CT             |
+| CystoidFluid (Ahmed et al., 2022)                     | 703    | 59         | OCT            |
+| Dca1 (Cervantes-Sanchez et al., 2019)                 | 93     | 28         | Fundus         |
+| Deepbacs (Spahn et al., 2022)                         | 17     | 10         | Microscopy     |
+| Drive (Staal et al., 2004)                            | 18     | 20         | Fundus         |
+| DynamicNuclear (Van Valen et al., 2016)               | 50     | 17         | Microscopy     |
+| FHPsAOP (Lu et al., 2022)                             | 2800   | 80         | Ultrasound     |
+| IDRiD (Porwal et al., 2018)                           | 47     | 27         | Fundus         |
+| ISIC2016 (Gutman et al., 2016)                        | 810    | 80         | Dermoscopy     |
+| ISIC2018 (Codella et al., 2019)                       | 9973   | 115        | Dermoscopy     |
+| KMAR-50K (Wang & Shi, 2025)                           | 651    | 47         | MRI            |
+| Kvasir (Jha et al., 2019)                             | 4429   | 139        | Endoscopy      |
+| Lgg-mri-segmentation (Buda et al., 2019)              | 1669   | 55         | MRI            |
+| MoNuSAC (Verma et al., 2021)                          | 0      | 21         | Microscopy     |
+| MR-ART (Narai et al. ´ , 2022)                        | 820    | 18         | MRI            |
+| MSD (Antonelli et al., 2022)                          | 797    | 3912       | MRI            |
+| NuSeT (Yang et al., 2020)                             | 2383   | 40         | Microscopy     |
+| Paried MRI CT (Younus Akon, 2025)                     | 1974   | 72         | CT, MRI        |
+| Pandental (Abdi et al., 2015)                         | 81     | 24         | Xray           |
+| Pasta-GEN (Lei et al., 2025)                          | 32299  | 731        | CT             |
+| PolypGen (Ali et al., 2024)                           | 984    | 75         | Endoscopy      |
+| PROMISE12 (Litjens et al., 2014)                      | 1031   | 80         | MRI            |
+| QaTa-COV19 (Aysen et al., 2024)                       | 3573   | 85         | Xray           |
+| Refuge (Fang et al., 2022)                            | 80     | 80         | Fundus         |
+| RoboTool (Garcia-Peraza-Herrera et al., 2021)         | 350    | 76         | Surgical Photo |
+| ThyroidXL (Duong et al., 2025)                        | 7029   | 138        | Ultrasound     |
+| Tnbcnuclei (Naylor et al., 2018)                      | 35     | 10         | Microscopy     |
+| TotalSegmentator (Wasserthal et al., 2023)            | 5206   | 154        | CT, MRI        |
+| UltrasoundNerve (Montoya, 2026)                       | 1651   | 50         | Ultrasound     |
+| USforKidney (Song et al., 2022)                       | 4351   | 50         | Ultrasound     |
+| UWSkinCancer (Vision & Lab, 2024)                     | 143    | 44         | Dermoscopy     |
+| VinDr-Multiphase (Dao et al., 2022)                   | 3486   | 44         | CT             |
+| WBC (Zheng et al., 2018)                              | 280    | 40         | Microscopy     |
+| YeaZ (Dietler et al., 2020)                           | 358    | 51         | Microscopy     |
+| YGA low dose ct (pazhoulab, 2024)                     | 4387   | 44         | CT             |
+
+We also appreciate MedSegBench(Kus¸ [& Aydin,](#page-10-20) [2024\)](#page-10-20) and MedSegDB[\(Zhang et al.,](#page-11-18) [2025\)](#page-11-18) for collecting and pre-processing some of these datasets.
+
+{14}------------------------------------------------
+
+# B. Construction Details
+
+770 771
+
+774
+
+776
+
+794
+
+796
+
+![](_page_14_Figure_2.jpeg)
+
+*Figure 6.* Construction details of three perspective. We manually curate the benchmark split to uphold high clinical standards. The remaining training data is validated through sampling-based quality checks, establishing a high-quality data proportion exceeding 95%.
+
+# C. Implementation Details of OmniGen2-MIE
+
+| Hyper-Parameter               | Value                     |
+|-------------------------------|---------------------------|
+| Finetuning method             | Full-Parameter Finetuning |
+| snr type                      | lognorm                   |
+| do shift                      | True                      |
+| dynamic time shift            | True                      |
+| Steps                         | 20, 000                   |
+| #GPUs                         | 8                         |
+| Per-device batch size         | 8                         |
+| Gradient accumulation         | 1                         |
+| Global batch size (effective) | 64                        |
+| Learning rate                 | 1 × 10−4                  |
+| LR scheduler                  | timm constant with warmup |
+| Warm-up t                     | 500                       |
+| Precision                     | BF16                      |
+| Random seed                   | 2233                      |
+
+*Table 5.* Training hyper-parameters used for finetuning OmniGen2-MIE on our dataset.
+
+{15}------------------------------------------------
+
+# <span id="page-15-0"></span>D. Evaluation Details
+
+#### D.1. Mask Reconstruction via Alpha De-blending
+
+### D.1.1. MATHEMATICS
+
+To recover the segmentation mask from the visualized output, we model the edited image O as a linear interpolation between the original background image B (a.k.a. the input image I) and a known overlay color C (red, green or blue). This relationship is governed by the per-pixel alpha channel α ∈ [0, 1], according to the standard alpha blending equation:
+
+$$\mathbf{O} = (1 - \alpha)\mathbf{B} + \alpha\mathbf{C} \tag{1}$$
+
+By rearranging the terms as O − B = α(C − B), the scalar value α can be interpreted as the projection of the observed color shift onto the vector representing the maximum possible color change. To account for potential noise in the RGB space, we solve for α at each pixel using the least-squares solution:
+
+$$\alpha = \frac{(\mathbf{O} - \mathbf{B}) \cdot (\mathbf{C} - \mathbf{B})}{|\mathbf{C} - \mathbf{B}|^2}$$
+ (2)
+
+The continuous alpha map is subsequently binarized to produce the final segmentation mask M. This is achieved by applying a global threshold τ , such that:
+
+$$M_{i,j} = \begin{cases} 1 & \text{if } \alpha_{i,j} > \tau \\ 0 & \text{otherwise} \end{cases}$$
+ (3)
+
+In our implementation, a threshold of τ = 0.5 is utilized to effectively separate the predicted regions from the background.
+
+#### D.1.2. CASE OF MASK RECONSTRUCTION
+
+![](_page_15_Picture_12.jpeg)
+
+*Figure 7.* Case of perception mask reconstruction.
+
+#### <span id="page-15-1"></span>D.2. VLM Automatic Scoring
+
+#### D.2.1. VLM SCORING RUBRIC
+
+# Scoring Rubric for Modification Tasks
+
+You are a helpful assistant in evaluating medical image editing result.
+
+You will be provided with an edit instruction and a collage image where the leftmost is origin image, center is edited image and rightmost is the reference ground truth image.
+
+You should score how well an edited image matches the intended edit while preserving clinical realism and image integrity based on following scoring rubrics:
+
+#### 1) Edit Goal Fulfillment (Edit Correctness): Assesses whether the intended lesion change is achieved.
+
+Scoring reference:
+
+- *5: Lesion added/removed exactly as intended; no residuals or unintended remnants.*
+- *4: Mostly correct; slight residual signal after removal or slight under/over-addition.*
+- *3: Partial success; lesion still partially present (removal) or incomplete/incorrect lesion (addition).*
+
+{16}------------------------------------------------
+
+> 932 933 934
+
+- *2: Wrong area or wrong type of change; target lesion largely unchanged.*
+- *1: No effective edit or opposite edit performed.*
+- 2) Edit Area Morphology (Shape, Margins, Internal Structure): Evaluates whether edit area matches expected morphology and/or reference.
+
+### Scoring reference:
+
+- *5: Shape, border characteristics, and internal texture are highly consistent.*
+- *4: Minor border/shape irregularities; still plausible.*
+- *3: Morphology is generic/unconvincing; borders/texture inconsistent.*
+- *2: Clearly artificial morphology (blocky, repeated patterns, unnatural contours).*
+- *1: Morphology nonsensical or misleading (e.g., appears like different pathology).*
+- 3) Intensity / Signal / Attenuation Consistency: Checks whether edited region match modality-specific intensities.
+
+#### Scoring reference:
+
+- *5: Intensities match local tissue statistics; no intensity discontinuities.*
+- *4: Slight intensity mismatch detectable with careful viewing.*
+- *3: Obvious mismatch (too bright/dark), inconsistent with modality or anatomy.*
+- *2: Strong intensity discontinuity; clearly edited.*
+- *1: Severe intensity errors that invalidate the image (e.g., saturation/clipping, inverted contrast).*
+- 4) Boundary Blending & Transition Naturalness: Rates edge blending and transitions between edited and unedited regions.
+
+#### Scoring reference:
+
+- *5: Seamless blending; no halos, ringing, cut-paste edges.*
+- *4: Minor halo/edge artifacts only on close inspection.*
+- *3: Visible seams; boundary looks edited.*
+- *2: Strong cutout appearance or blur patches.*
+- *1: Boundary artifacts dominate the image.*
+- 5) Background / Non-target Preservation: Measures unintended changes outside the lesion edit region. Scoring reference:
+  - *5: Non-target anatomy and background unchanged (within expected noise).*
+  - *4: Small unintended changes but not clinically meaningful.*
+  - *3: Noticeable unintended alterations in nearby structures.*
+  - *2: Large unintended modifications to anatomy or overall image.*
+  - *1: Global corruption or major anatomical distortions.*
+- 6) Anatomical Plausibility & Clinical Coherence: Assesses whether result respects anatomy and pathology logic (e.g., lesion doesn't cross impossible boundaries). Scoring reference:
+  - *5: Fully plausible; consistent with organ boundaries and expected presentation.*
+  - *4: Mostly plausible; minor oddity but acceptable.*
+  - *3: Questionable plausibility (e.g., lesion overlaps structures unnaturally).*
+  - *2: Clearly implausible anatomy/pathology relationship.*
+  - *1: Clinically nonsensical or misleading.*
+- 7) Artifact Introduction (Noise, Texture, Aliasing, Compression, Repetition): Evaluates new artifacts introduced by editing. Scoring reference:
+
+{17}------------------------------------------------
+
+- *5: No new artifacts; noise texture consistent with original.*
+- *4: Minor artifacts (subtle smoothing/grain mismatch).*
+- *3: Artifacts visible and distracting.*
+- *2: Strong artifacts (banding, checkerboard, repeated texture).*
+- *1: Severe artifacts preventing clinical use.*
+- 8) Image Quality & Acquisition Consistency: Checks consistency with scanner characteristics (resolution, blur, point spread, slice thickness cues, motion). Scoring reference:
+  - *5: Matches acquisition characteristics; sharpness/noise consistent.*
+  - *4: Slight mismatch in sharpness or noise level.*
+  - *3: Clear mismatch (over-smoothed or over-sharpened region).*
+  - *2: Strong mismatch; edited region appears from different source.*
+  - *1: Completely inconsistent with acquisition; unusable.*
+
+#### Return Format: Return a JSON dictionary with two fields:
+
+- "conclusion": A brief conclusion to the edited image.
+- "score list": The scores of the eight aspects in a JSON list.
+
+#### D.2.2. CASE
+
+![](_page_17_Picture_26.jpeg)
+
+# **Input collage VLM Output**
+
+**conclusion**: The edited image successfully removes the renal cyst and replaces it with plausible parenchyma closely matching the reference, with only mild residual smoothing/texture mismatch in the filled region and minimal impact on surrounding anatomy.
+
+**score\_list**: [5, 4, 4, 4, 5, 5, 4, 4]
+
+![](_page_17_Picture_31.jpeg)
+
+**conclusion**: The edited image does not introduce a clinically plausible gallstone. Instead, it contains multiple obvious synthetic circular overlays and line artifacts across the abdomen, with major non-target corruption and mismatch to CT texture; it does not match the reference which shows no such artifacts.
+
+**score\_list**: [1, 1, 1, 1, 1, 1, 1, 1]
+
+*Figure 8.* Cases of VLM rubric scoring.
+
+{18}------------------------------------------------
+
+# **E. Supplementary Experiments**
+
+#### <span id="page-18-1"></span><span id="page-18-0"></span>E.1. Modality-Wise Performance Analysis
+
+![](_page_18_Figure_3.jpeg)
+
+Figure 9. Modality-wise performance analysis result within perception perspective. Left: DICE score; right: PSNR score.
+
+To investigate the impact of modality deviation, we conduct a modality-wise analysis of the benchmarking results within the Perception perspective. Specifically, we report the DICE and PSNR scores of six representative models across all medical imaging modalities included in MieDB-100k. As illustrated in Fig. 9, the experimental results are consistent with our hypotheses. For the baseline models, performance is unevenly distributed across the various modalities: They achieve relatively strong results on modalities that resemble natural images, such as Endoscopy, Dermoscopy, and Surgical Photo. However, on non-optical modalities (e.g., CT, MRI, Ultrasound), their performance degrades drastically. In contrast, the model trained on our dataset exhibits balanced and superior performance across all imaging types. Collectively, these results demonstrate that a diverse dataset like MieDB-100k is essential for successfully adapting multi-modal generative models to the medical domain.
+
+#### <span id="page-18-2"></span>E.2. Multi-Round Generation
+
+Table 6. Multi-round generation result. Best values are marked in Bold
+
+|                     | Perception |        |        |        |        |        |        |        | Trasnformation |        |        |        |  |
+|---------------------|------------|--------|--------|--------|--------|--------|--------|--------|----------------|--------|--------|--------|--|
+|                     | DICE       |        | P-A    | P-ACC  |        | B-PSNR |        | B-SSIM |                | PSNR   |        | SSIM   |  |
+|                     | Pass@1     | Pass@3 | Pass@1 | Pass@3 | Pass@1 | Pass@3 | Pass@1 | Pass@3 | Pass@1         | Pass@3 | Pass@1 | Pass@3 |  |
+| Open-Source         |            |        |        |        |        |        |        |        |                |        |        |        |  |
+| SDXL-turbo          | 0.002      | 0.003  | 0.000  | 0.000  | 16.6   | 17.0   | 0.467  | 0.484  | 15.9           | 16.1   | 0.397  | 0.427  |  |
+| Bagel               | 0.263      | 0.383  | 0.069  | 0.137  | 13.9   | 16.1   | 0.620  | 0.703  | 12.7           | 15.2   | 0.442  | 0.548  |  |
+| OmniGen2            | 0.248      | 0.357  | 0.065  | 0.125  | 11.9   | 14.4   | 0.541  | 0.628  | 8.3            | 16.0   | 0.280  | 0.551  |  |
+| Step1X-Edit         | 0.332      | 0.369  | 0.126  | 0.143  | 15.5   | 16.4   | 0.727  | 0.748  | 16.6           | 17.1   | 0.539  | 0.558  |  |
+| FLUX.1-Kontext-dev  | 0.347      | 0.41   | 0.126  | 0.174  | 15.4   | 16.5   | 0.701  | 0.761  | 17.9           | 19.5   | 0.543  | 0.602  |  |
+| Qwen-Image-Edit     | 0.387      | 0.493  | 0.153  | 0.249  | 15.4   | 17.4   | 0.722  | 0.795  | 18.9           | 20.3   | 0.606  | 0.652  |  |
+| OmniGen2-MIE (Ours) | 0.831      | 0.856  | 0.737  | 0.789  | 28.1   | 28.8   | 0.917  | 0.921  | 22.6           | 23.3   | 0.685  | 0.711  |  |
+
+To mitigate the inherent variance of the generative process, we report Pass@3 scores for the open-source models on Perception tasks. Specifically, we generate three independent outputs for each editing task and select the highest-performing sample to represent the task's score. These results are then averaged across all tasks to provide a robust assessment of overall performance.
+
+The results of the multi-round generation tests are summarized in Table 6. While multi-round generation improves the absolute scores for baseline models, it does not alter the underlying fact that these models lack essential medical knowledge. Furthermore, the significant fluctuations across rounds expose the high-variance nature of these baselines, undermining their reliability under clinical applications. In contrast, our model exhibits remarkable stability across all three trials. This
+
+{19}------------------------------------------------
+
+consistency suggests that model trained on MieDB-100k has developed a deterministic understanding of medical concepts rather than relying on fortuitous generation.
+
+#### E.3. Out-Of-Distribution Image Edit
+
+1054
+
+1056
+
+1058
+
+1071
+
+1074 1075 1076
+
+1078
+
+While Section [4.5](#page-7-2) demonstrates that the model trained on MieDB-100k generalizes effectively to OOD editing targets, we further evaluate its robustness by performing edits on 'in-the-wild' medical images sourced from the internet (Fig. [10\)](#page-19-0).
+
+<span id="page-19-0"></span>![](_page_19_Figure_4.jpeg)
+
+*Figure 10.* Examples of Out-Of-Distribution Editing.
+
+The results indicate that our model is capable of readily adapting to medical images outside of datasets. This suggests that the diversity of MieDB-100k has successfully decoupled the model from specific data distribution, allowing it to internalize generalizable edit operations that are applicable to real-world clinical scenarios.
+
+# F. Data Gallery and More Qualitative Result
+
+#### F.1. Examples of Healthy Tissue Inpainting
+
+![](_page_19_Figure_9.jpeg)
+
+*Figure 11.* Examples of Inpainting. We train different inpainting models on each medical modalities. H: the Healthy image; L: the Lesion-bearing image.
+
+#### F.2. Extended Examples of Qualitative result
+
+{20}------------------------------------------------
+
+ Input Image Bagel FLUX.1 KontextOmniGen2OmniGen2-MIE Reference Nano Banana ProGPT Qwen Image Edit -Image **[Edit Instruction]** Illustrate all nuclei instances by painting GREEN masks directly onto the microscopy image, with zero modification to the background. Edit **[Edit Instruction]** Use a RED mask to highlight the polyp(s) by drawing it directly onto the given image, preserving the background exactly as is. Edit **[Edit Instruction]** Overlay unique RED masks on each organ in the ultrasound image by painting them directly, keeping the original background intact. Edit **[Edit Instruction]** Visually segment Caenorhabditis elegans by applying BLUE masks directly onto the microscope image without affecting the background. **[Edit Instruction]** Mark the region corresponding to BLADDER TUMOR in the Enhanced CT image using a solid RED mask applied directly, while ensuring the background stays unmodified. **[Edit Instruction]** Encode the lesion location with a GREEN mask drawn directly on the image, maintaining original background appearance. **[Edit Instruction]** Cover the infected zones with a RED mask by painting it directly onto the chest X-ray, ensuring background fidelity. Edit Edit Edit Edit
+
+{21}------------------------------------------------
+
+![](_page_21_Picture_2.jpeg)
+
+**[Edit Instruction]** Delete all visible manifestations of COVID pneumonia from the X-ray image and replace them with normal lung appearance, maintaining vascular and bronchial structures.
+
+![](_page_21_Picture_4.jpeg)
+
+**[Edit Instruction]** Eliminate the Tumor in the Stomach while ensuring the reconstructed zone blends naturally with adjacent healthy tissue in terms of Hounsfield units and spatial patterns.
+
+![](_page_21_Picture_6.jpeg)
+
+**[Edit Instruction]** Simulate a focal mucosal lesion in the white masked region that could represent a common colonic or gastric polyp, based on realistic endoscopic criteria.
+
+![](_page_21_Picture_8.jpeg)
+
+**[Edit Instruction]** Reconstruct the affected skin area by removing the lesion and generating realistic, symmetryconsistent healthy tissue based on surrounding context.
+
+![](_page_21_Picture_10.jpeg)
+
+**[Edit Instruction]** Convert the white masked area into a realistic representation of healthy skin, avoiding artificial smoothness or color mismatches.
+
+![](_page_21_Picture_12.jpeg)
+
+**[Edit Instruction]** Synthetically insert a thyroid nodule into the given ultrasound scan, ensuring it exhibits clinically plausible features such as shape, margin, echogenicity, and vascularity.
+
+![](_page_21_Picture_14.jpeg)
+
+**[Edit Instruction]** Add a convincing brain tumor to the FLAIR MRI while maintaining correct contrast dynamics—such as suppressed CSF and bright pathological signal.
+
+{22}------------------------------------------------
+
+![](_page_22_Picture_2.jpeg)
+
+![](_page_22_Picture_3.jpeg)
+
+**[Edit Instruction]** Improve diagnostic clarity of the CT by removing metal degradation while ensuring zero change to medically relevant content outside the artifact zone.
+
+![](_page_22_Picture_5.jpeg)
+
+**[Edit Instruction]** Filter out motion artifacts from the PD MRI scan without smoothing, warping, or otherwise altering true tissue signals.
+
+![](_page_22_Picture_7.jpeg)
+
+**[Edit Instruction]** Render the CT with Abdominal window parameters to mimic how it would appear on a PACS viewer configured for that tissue type.
+
+![](_page_22_Picture_9.jpeg)
+
+**[Edit Instruction]** Synthesize a Arterial phase-equivalent CT image from the provided Non-Contrast phase input, ensuring realistic vascular and tissue enhancement patterns.
+
+![](_page_22_Picture_11.jpeg)
+
+**[Edit Instruction]** Apply post-processing to remove metal artifacts from the CT image without smoothing, interpolating, or altering real anatomical details unnecessarily.
+
+![](_page_22_Picture_13.jpeg)
+
+**[Edit Instruction]** Convert the given T2-MRI image to resemble a CT image, maintaining all underlying medical information intact.
+
+![](_page_22_Picture_15.jpeg)
+
+**[Edit Instruction]** Process the low-dose CT using a SOFT kernel to suppress graininess while preserving low-contrast diagnostic details.
+
+{23}------------------------------------------------
+
+# <span id="page-23-0"></span>G. Failure Cases
+
+![](_page_23_Figure_2.jpeg)
+
+*Figure 12.* Failure Cases.
+
+Fig [12](#page-23-0) illustrates several representative failure cases of OmniGen2-MIE. The most frequent failure modes include: (1) semantic confusion between targeted anatomical features and morphologically similar background tissues; (2) intensity inconsistency, where the brightness of the edited region deviates from the surrounding context in a physically implausible manner; and (3) background inconsistency, especially after holistic transformations. These limitations underscore the need for more sophisticated multimodal architectures capable of preserving fine-grained details, as well as even more comprehensive training datasets to satisfy the requirements of rigorous clinical applications.
+
+# H. Limitations
+
+Despite MieDB-100k provides a large-scale and diverse dataset for medical image editing, the primary limitation lies in the inherent difficulty of capturing ALL possible medical imaging modalities, and the relative scarcity of data for rare clinical cases. Continuous efforts to enrich these underrepresented categories will be vital for enhancing the dataset's diversity and effectivity. Furthermore, while our work establishes a foundation for unified understanding and generation, it focuses exclusively on editing tasks. Integrating medical VQA and text-to-image datasets represents a natural progression of this research direction, resulting in a more comprehensive resource for the development of holistic medical models.

icml26/80c20b7b-ead6-454a-849e-56702a6c828f/parse_report.json

@@ -0,0 +1,78 @@
+{
+  "paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f",
+  "pipeline": "marker_non_llm_v3",
+  "parser": "marker_single",
+  "formats": [
+    "markdown",
+    "chunks"
+  ],
+  "llm_enabled": false,
+  "pdf_path": "/network/scratch/j/jianan.zhao/ReviewAgent/data/processed_papers/icml26_20260429_1952_duequeue/raw/80c20b7b-ead6-454a-849e-56702a6c828f.pdf",
+  "pdf_sha256": "c7b830098f394c3dbbcfa2dfbcb365651fbc6f09845bddf972672e735ecbdd51",
+  "bytes": 11333727,
+  "source": "https://koala.science/storage/pdfs/80c20b7b-ead6-454a-849e-56702a6c828f.pdf",
+  "page_count": 24,
+  "ok": true,
+  "elapsed_seconds": 180.07999999999998,
+  "paper2markdown_v3": {
+    "paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f",
+    "pipeline": "Paper2Markdown-V3",
+    "ok": true,
+    "page_count": 24,
+    "chunk_count": 140,
+    "main_body_chunk_count": 70,
+    "appendix_chunk_count": 57,
+    "reference_chunk_count": 13,
+    "model_text_chars": 34560,
+    "raw_markdown_chars": 87319,
+    "sanitized_chars": 77147,
+    "page_provenance": {
+      "min_page": 1,
+      "max_page": 24,
+      "invalid_count": 0
+    },
+    "marker_block_type_counts": {
+      "Caption": 3,
+      "Equation": 3,
+      "Figure": 1,
+      "FigureGroup": 9,
+      "Footnote": 1,
+      "ListGroup": 19,
+      "PageFooter": 24,
+      "PageHeader": 31,
+      "Picture": 3,
+      "PictureGroup": 15,
+      "SectionHeader": 55,
+      "TableGroup": 6,
+      "Text": 367
+    },
+    "asset_count_raw": 28,
+    "asset_count_model_kept": 13,
+    "asset_count_rejected": 15,
+    "asset_reject_reasons": {
+      "extreme_aspect_ratio": 15,
+      "kept": 13
+    },
+    "artifact_leak_audit": {
+      "ok": true,
+      "hits": {
+        "Anonymous Authors": [],
+        "ACKNOWLEDGMENT": [],
+        "OpenReview": [],
+        "\"accept_label\"": [],
+        "\"decision\"": [],
+        "\"decision_tier\"": [],
+        "\"source_status\"": [],
+        "Meta-review": [],
+        "Official Review": [],
+        "official_reviews": [],
+        "meta_reviews": [],
+        "suggested_verdict_score": []
+      },
+      "artifact_count": 2
+    },
+    "default_model_input": "model_text_v3.txt",
+    "appendix_input": "appendix_text_v3.txt",
+    "reference_input": "reference_text_v3.txt"
+  }
+}

icml26/80c20b7b-ead6-454a-849e-56702a6c828f/reference_chunks.jsonl

@@ -0,0 +1,13 @@
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0070", "section": "References", "page_start": 9, "page_end": 9, "type": "ListGroup", "text": "AAPM. Low dose ct grand challenge, 2016. URL https: // ect/#trainingData . Accessed: 2026-01-14. Abdi, A. H., Kasaei, S., and Mehdizadeh, M. Automatic segmentation of mandible in panoramic x-ray. Journal of Medical Imaging , 2(4):044003–044003, 2015. Ahmed, Z., Panhwar, S. Q., Baqai, A., Umrani, F. A., Ahmed, M., and Khan, A. Deep learning based automated detection of intraretinal cystoid fluid. International Jour nal of Imaging Systems and Technology , 32(3):902–917, 2022. Al-Dhabyani, W., Gomaa, M., Khaled, H., and Fahmy, A. Dataset of breast ultrasound images. Data in brief , 28: 104863, 2020. Ali, S., Ghatwary, N., Jha, D., Isik-Polat, E., Polat, G., Yang, C., Li, W., Galdran, A., Ballester, M.-A. G., Thambawita, ´ V., et al. Assessing generalisability of deep learningbased polyp detection and segmentation methods through a computer vision challenge. Scientific Reports , 14(1): 2032, 2024. Angiographics. Chuac dataset. URL re.com/s/4d24cf3d14bc901a94bf . Accessed: 2026-01-14. Antonelli, M., Reinke, A., Bakas, S., Farahani, K., Kopp-Schneider, A., Landman, B. A., Litjens, G., Menze, B., Ronneberger, O., Summers, R. M., et al. The medical segmentation decathlon. Nature communications , 13(1): 4128, 2022. Aysen, D., Muhammad, C., Tawsifur, R., Amith, K., Yazan, Q., Mete, A., Anas M., T., and serkan, k. Qata-cov19 dataset, 2024. URL atasets/aysendegerli/qatacov19-datas et . Accessed: 2026-01-14. Bai, S., Cai, Y., Chen, R., Chen, K., Chen, X., Cheng, Z., Deng, L., Ding, W., Gao, C., Ge, C., Ge, W., Guo, Z., Huang, Q., Huang, J., Huang, F., Hui, B., Jiang, S., Li, Z., Li, M., Li, M., Li, K., Lin, Z., Lin, J., Liu, X., Liu, J., Liu, C., Liu, Y., Liu, D., Liu, S., Lu, D., Luo, R., Lv, C., Men, R., Meng, L., Ren, X., Ren, X., Song, S., Sun, Y., Tang, J., Tu, J., Wan, J., Wang, P., Wang, P.,", "source": "marker_v2", "marker_block_id": "/page/8/ListGroup/530"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0071", "section": "References", "page_start": 9, "page_end": 9, "type": "ListGroup", "text": "Wang, Q., Wang, Y., Xie, T., Xu, Y., Xu, H., Xu, J., Yang, Z., Yang, M., Yang, J., Yang, A., Yu, B., Zhang, F., Zhang, H., Zhang, X., Zheng, B., Zhong, H., Zhou, J., Zhou, F., Zhou, J., Zhu, Y., and Zhu, K. Qwen3-vl technical report. Preprint at arXiv , 2025a. URL https: //arxiv.org/abs/2511.21631 . Bai, S., Chen, K., Liu, X., Wang, J., Ge, W., Song, S., Dang, K., Wang, P., Wang, S., Tang, J., Zhong, H., Zhu, Y., Yang, M., Li, Z., Wan, J., Wang, P., Ding, W., Fu, Z., Xu, Y., Ye, J., Zhang, X., Xie, T., Cheng, Z., Zhang, H., Yang, Z., Xu, H., and Lin, J. Qwen2.5-vl technical report. Preprint at arXiv , 2025b. URL g/abs/2502.13923 . Brooks, T., Holynski, A., and Efros, A. A. Instructpix2pix: Learning to follow image editing instructions. In Pro ceedings of the IEEE/CVF conference on computer vision and pattern recognition , pp. 18392–18402, 2023. Buda, M., Saha, A., and Mazurowski, M. A. Association of genomic subtypes of lower-grade gliomas with shape features automatically extracted by a deep learning algorithm. Computers in biology and medicine , 109:218–225, 2019. Caicedo, J. C., Goodman, A., Karhohs, K. W., Cimini, B. A., Ackerman, J., Haghighi, M., Heng, C., Becker, T., Doan, M., McQuin, C., et al. Nucleus segmentation across imaging experiments: the 2018 data science bowl. Nature methods , 16(12):1247–1253, 2019. Carballal, A., Novoa, F. J., Fernandez-Lozano, C., Garc´ıa-Guimaraes, M., Aldama-Lopez, G., Calvi ´ no-Santos, R., ˜ Vazquez-Rodriguez, J. M., and Pazos, A. Automatic multiscale vascular image segmentation algorithm for coronary angiography. Biomedical Signal Processing and Control , 46:1–9, 2018. Cervantes-Sanchez, F., Cruz-Aceves, I., Hernandez-Aguirre, A., Hernandez-Gonzalez, M. A., and Solorio-Meza, S. E. Automatic segmentation of coronary arteries in x-ray angiograms using multiscale analysis and artificial neural networks. Applied Sciences , 9(24):5507, 2019. Chen, Z. and Feng, M. Med-banana-50k: A cross-modality large-scale dataset for text-guided medical image editing. Preprint at arXiv , 2025. URL g/abs/2511.00801 . Chowdhury, M. E., Rahman, T., Khandakar, A., Mazhar, R., Kadir, M. A., Mahbub, Z. B., Islam, K. R., Khan, M. S., Iqbal, A., Al Emadi, N., et al. Can ai help in screening viral and covid-19 pneumonia? Ieee Access , 8: 132665–132676, 2020.", "source": "marker_v2", "marker_block_id": "/page/8/ListGroup/531"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0072", "section": "References", "page_start": 10, "page_end": 10, "type": "Text", "text": "495 496 497 498 499 500 Codella, N., Rotemberg, V., Tschandl, P., Celebi, M. E., Dusza, S., Gutman, D., Helba, B., Kalloo, A., Liopyris, K., Marchetti, M., et al. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic). arXiv preprint arXiv:1902.03368 , 2019.", "source": "marker_v2", "marker_block_id": "/page/9/Text/501"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0073", "section": "References", "page_start": 10, "page_end": 10, "type": "ListGroup", "text": "Dao, B. T., Nguyen, T. V., Pham, H. H., and Nguyen, H. Q. Phase recognition in contrast-enhanced ct scans based on deep learning and random sampling. Medical Physics , 49 (7):4518–4528, 2022. de Verdier, M. C., Saluja, R., Gagnon, L., LaBella, D., Baid, U., Tahon, N. H., Foltyn-Dumitru, M., Zhang, J., Alafif, M., Baig, S., Chang, K., D'Anna, G., Deptula, L., Gupta, D., Haider, M. A., Hussain, A., Iv, M., Kontzialis, M., Manning, P., Moodi, F., Nunes, T., Simon, A., Sollmann, N., Vu, D., Adewole, M., Albrecht, J., Anazodo, U., Chai, R., Chung, V., Faghani, S., Farahani, K., Kazerooni, A. F., Iglesias, E., Kofler, F., Li, H., Linguraru, M. G., Menze, B., Moawad, A. W., Velichko, Y., Wiestler, B., Altes, T., Basavasagar, P., Bendszus, M., Brugnara, G., Cho, J., Dhemesh, Y., Fields, B. K. K., Garrett, F., Gass, J., Hadjiiski, L., Hattangadi-Gluth, J., Hess, C., Houk, J. L., Isufi, E., Layfield, L. J., Mastorakos, G., Mongan, J., Nedelec, P., Nguyen, U., Oliva, S., Pease, M. W., Rastogi, A., Sinclair, J., Smith, R. X., Sugrue, L. P., Thacker, J., Vidic, I., Villanueva-Meyer, J., White, N. S., Aboian, M., Conte, G. M., Dale, A., Sabuncu, M. R., Seibert, T. M., Weinberg, B., Abayazeed, A., Huang, R., Turk, S., Rauschecker, A. M., Farid, N., Vollmuth, P., Nada, A., Bakas, S., Calabrese, E., and Rudie, J. D. The 2024 brain tumor segmentation (brats) challenge: Glioma segmentation on post-treatment mri. Preprint at arXiv , 2024. URL 8368 . DeepMind, G. Gemini 3 pro: High-precision multimodal reasoning, 2025a. URL gle/models/gemini-image/pro/ . Accessed: 2026-01-14. DeepMind, G. Imagen 4.0: Model documentation and generation guide, 2025b. URL loud.google.com/vertex-ai/generative -ai/docs/models/imagen/4-0-generate . Accessed: 2026-01-14. Deng, C., Zhu, D., Li, K., Gou, C., Li, F., Wang, Z., Zhong, S., Yu, W., Nie, X., Song, Z., Shi, G., and Fan, H. Emerging properties in unified multimodal pretraining. Preprint at arXiv , 2025. URL https: //arxiv.org/abs/2505.14683 . Dietler, N., Minder, M., Gligorovski, V., Economou, A. M., Joly, D. A. H. L., Sadeghi, A., Chan, C. H. M., Kozinski, ´", "source": "marker_v2", "marker_block_id": "/page/9/ListGroup/499"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0074", "section": "References", "page_start": 10, "page_end": 10, "type": "ListGroup", "text": "M., Weigert, M., Bitbol, A.-F., et al. A convolutional neural network segments yeast microscopy images with high accuracy. Nature communications , 11(1):5723, 2020. Duong, V. H., Vu, H., Phan, H. D., Nguyen, D. Q., Pham, D. H., Le, Q. T., Nguyen, B. S., Do, T. D., Dinh, V. S., Nguyen, T. C., et al. Thyroidxl: Advancing thyroid nodule diagnosis with an expert-labeled, pathology-validated dataset. In International Conference on Medical Image Computing and Computer-Assisted Intervention , pp. 616– 626. Springer, 2025. Fang, H., Li, F., Wu, J., Fu, H., Sun, X., Son, J., Yu, S., Zhang, M., Yuan, C., Bian, C., Lei, B., Zhao, B., Xu, X., Li, S., Fumero, F., Sigut, J., Almubarak, H., Bazi, Y., Guo, Y., Zhou, Y., Baid, U., Innani, S., Guo, T., Yang, J., Orlando, J. I., Bogunovic, H., Zhang, X., and Xu, Y. ´ Refuge2 challenge: A treasure trove for multi-dimension analysis and evaluation in glaucoma screening. Preprint at arXiv , 2022. URL 2202.08994 . Garcia-Peraza-Herrera, L. C., Fidon, L., D'Ettorre, C., Stoyanov, D., Vercauteren, T., and Ourselin, S. Image compositing for segmentation of surgical tools without manual annotations. IEEE transactions on medical imaging , 40(5):1450–1460, 2021. Gutman, D., Codella, N. C., Celebi, E., Helba, B., Marchetti, M., Mishra, N., and Halpern, A. Skin lesion analysis toward melanoma detection: A challenge at the international symposium on biomedical imaging (isbi) 2016, hosted by the international skin imaging collaboration (isic). arXiv preprint arXiv:1605.01397 , 2016. Haneda, E., Peters, N., Zhang, J., Karageorgos, G., Xia, W., Paganetti, H., Wang, G., Guo, Y., Ma, J., Park, H. S., et al. Aapm ct metal artifact reduction grand challenge. Medical physics , 52(10):e70050, 2025. Huang, Z., Zheng, D., Zou, C., Liu, R., Wang, X., Ji, K., Chai, W., Sun, J., Wang, L., Lv, Y., Huang, T., Liu, J., Guo, Q., Yang, M., Chen, J., and Zhou, J. Ming-univision: Joint image understanding and generation with a unified continuous tokenizer. Preprint at arXiv , 2025. URL . Isensee, F., Jaeger, P. F., Kohl, S. A., Petersen, J., and Maier-Hein, K. H. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods , 18(2):203–211, 2021. Jha, D., Smedsrud, P. H., Riegler, M. A., Halvorsen, P., De Lange, T., Johansen, D., and Johansen, H. D. Kvasirseg: A segmented polyp dataset. In International con ference on multimedia modeling , pp. 451–462. Springer, 2019.", "source": "marker_v2", "marker_block_id": "/page/9/ListGroup/500"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0075", "section": "References", "page_start": 11, "page_end": 11, "type": "Text", "text": "550 Kus¸, Z. and Aydin, M. Medsegbench: A comprehensive benchmark for medical image segmentation in diverse data modalities. Scientific Data , 11(1):1283, 2024.", "source": "marker_v2", "marker_block_id": "/page/10/Text/520"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0076", "section": "References", "page_start": 11, "page_end": 11, "type": "ListGroup", "text": "Labs, B. F., Batifol, S., Blattmann, A., Boesel, F., Consul, S., Diagne, C., Dockhorn, T., English, J., English, Z., Esser, P., Kulal, S., Lacey, K., Levi, Y., Li, C., Lorenz, D., Muller, J., Podell, D., Rombach, R., Saini, H., Sauer, ¨ A., and Smith, L. Flux.1 kontext: Flow matching for in-context image generation and editing in latent space. Preprint at arXiv , 2025. URL g/abs/2506.15742 . Lei, W., Chen, H., Zhang, Z., Luo, L., Xiao, Q., Gu, Y., Gao, P., Jiang, Y., Wang, C., Wu, G., Xu, T., Zhang, Y., Rajpurkar, P., Zhang, X., Zhang, S., and Wang, Z. A synthetic data-driven radiology foundation model for pantumor clinical diagnosis. Preprint at arXiv , 2025. URL . Litjens, G., Toth, R., Van De Ven, W., Hoeks, C., Kerkstra, S., Van Ginneken, B., Vincent, G., Guillard, G., Birbeck, N., Zhang, J., et al. Evaluation of prostate segmentation algorithms for mri: the promise12 challenge. Medical image analysis , 18(2):359–373, 2014. Liu, J., Liu, G., Liang, J., Li, Y., Liu, J., Wang, X., Wan, P., Zhang, D., and Ouyang, W. Flow-grpo: Training flow matching models via online rl. arXiv preprint at arXiv , 2025a. URL 5470 . Liu, M., He, Z., Fan, Z., Wang, Q., and Fung, Y. R. Medebench: Diagnosing reliability in text-guided medical image editing. In Findings of the Association for Computational Linguistics: EMNLP 2025 , pp. 767–791, 2025b. Liu, S., Han, Y., Xing, P., Yin, F., Wang, R., Cheng, W., Liao, J., Wang, Y., Fu, H., Han, C., Li, G., Peng, Y., Sun, Q., Wu, J., Cai, Y., Ge, Z., Ming, R., Xia, L., Zeng, X., Zhu, Y., Jiao, B., Zhang, X., Yu, G., and Jiang, D. Step1xedit: A practical framework for general image editing. Preprint at arXiv , 2025c. URL g/abs/2504.17761 . Ljosa, V., Sokolnicki, K. L., and Carpenter, A. E. Annotated high-throughput microscopy image sets for validation. Nature methods , 9(7):637, 2012. Lu, Y., Zhou, M., Zhi, D., Zhou, M., Jiang, X., Qiu, R., Ou, Z., Wang, H., Qiu, D., Zhong, M., et al. The jnu-ifm dataset for segmenting pubic symphysis-fetal head. Data in brief , 41:107904, 2022. Mathieu, G., Bachir, E. D., et al. Brifiseg: a deep learningbased method for semantic and instance segmentation", "source": "marker_v2", "marker_block_id": "/page/10/ListGroup/518"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0077", "section": "References", "page_start": 11, "page_end": 11, "type": "ListGroup", "text": "of nuclei in brightfield images. Preprint at arXiv , 2022. URL . Montoya, A. Ultrasound nerve segmentation, 2026. URL ultrasound-nerve-segmentation . Accessed: 2026-01-14. Morozov, S. P., Andreychenko, A. E., Pavlov, N. A., Vladzymyrskyy, A., Ledikhova, N. V., Gombolevskiy, V. A., Blokhin, I. A., Gelezhe, P. B., Gonchar, A., and Chernina, V. Y. Mosmeddata: Chest ct scans with covid-19 related findings dataset. arXiv preprint arXiv:2005.06465 , 2020. Narai, ´ A., Hermann, P., Auer, T., Kemenczky, P., Szalma, ´ J., Homolya, I., Somogyi, E., Vakli, P., Weiss, B., and Vidnyanszky, Z. Movement-related artefacts (mr-art) ´ dataset of matched motion-corrupted and clean structural mri brain scans. Scientific data , 9(1):630, 2022. Naylor, P., Lae, M., Reyal, F., and Walter, T. Segmentation ´ of nuclei in histopathology images by deep regression of the distance map. IEEE transactions on medical imaging , 38(2):448–459, 2018. Ngoc Lan, P., An, N. S., Hang, D. V., Long, D. V., Trung, T. Q., Thuy, N. T., and Sang, D. V. Neounet: Towards accurate colon polyp segmentation and neoplasm detection. In International Symposium on Visual Computing , pp. 15–28. Springer, 2021. OpenAI. Gpt image 1 model documentation, 2025. URL els/gpt-image-1 . Accessed: 2026-01-14. Pandey, N. Chest xray masks and labels, 2025. URL ht tps:// pandey360/chest-xray-masks-and-labels . Accessed: 2026-01-14. pazhoulab. Low-dose ct reconstruction contest, 2024. URL ontestdetail?id=667e0c687ff47da8cc82 7679&award=1,000,000 . Accessed: 2026-01-14. Polo, M. Chest ct segmentation, 2025. URL https: // chest-ct-segmentation . Accessed: 2026-01-14. Porwal, P., Pachade, S., Kamble, R., Kokare, M., Deshmukh, G., Sahasrabuddhe, V., and Meriaudeau, F. Indian diabetic retinopathy image dataset (idrid): a database for diabetic retinopathy screening research. Data , 3(3):25, 2018. Sauer, A., Lorenz, D., Blattmann, A., and Rombach, R. Adversarial diffusion distillation. In European Conference on Computer Vision , pp. 87–103. Springer, 2024.", "source": "marker_v2", "marker_block_id": "/page/10/ListGroup/519"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0078", "section": "References", "page_start": 12, "page_end": 12, "type": "Text", "text": "Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., Bi, X., Zhang, H., Zhang, M., Li, Y. K., Wu, Y., and Guo, D. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. Preprint at arXiv , 2024. URL .", "source": "marker_v2", "marker_block_id": "/page/11/Text/533"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0079", "section": "References", "page_start": 12, "page_end": 12, "type": "ListGroup", "text": "Song, Y., Zheng, J., Lei, L., Ni, Z., Zhao, B., and Hu, Y. Ct2us: Cross-modal transfer learning for kidney segmentation in ultrasound images with synthesized data. Ultrasonics , 122:106706, 2022. Spahn, C., Gomez-de Mariscal, E., Laine, R. F., Pereira, ´ P. M., von Chamier, L., Conduit, M., Pinho, M. G., Jacquemet, G., Holden, S., Heilemann, M., et al. Deepbacs for multi-task bacterial image analysis using opensource deep learning approaches. Communications Biol ogy , 5(1):688, 2022. Staal, J., Abramoff, M. D., Niemeijer, M., Viergever, M. A., ` and Van Ginneken, B. Ridge-based vessel segmentation in color images of the retina. IEEE transactions on medi cal imaging , 23(4):501–509, 2004. Tahir, A. M., Chowdhury, M. E., Khandakar, A., Rahman, T., Qiblawey, Y., Khurshid, U., Kiranyaz, S., Ibtehaz, N., Rahman, M. S., Al-Maadeed, S., et al. Covid-19 infection localization and severity grading from chest x-ray images. Computers in biology and medicine , 139:105002, 2021. Tong, S., Fan, D., Li, J., Xiong, Y., Chen, X., Sinha, K., Rabbat, M., LeCun, Y., Xie, S., and Liu, Z. Metamorph: Multimodal understanding and generation via instruction tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pp. 17001–17012, 2025. Van Valen, D. A., Kudo, T., Lane, K. M., Macklin, D. N., Quach, N. T., DeFelice, M. M., Maayan, I., Tanouchi, Y., Ashley, E. A., and Covert, M. W. Deep learning automates the quantitative analysis of individual cells in live-cell imaging experiments. PLoS computational biology , 12(11):e1005177, 2016. Verma, R., Kumar, N., Patil, A., Kurian, N. C., Rane, S., Graham, S., Vu, Q. D., Zwager, M., Raza, S. E. A., Rajpoot, N., et al. Monusac2020: A multi-organ nuclei segmentation and classification challenge. IEEE Transac tions on Medical Imaging , 40(12):3413–3423, 2021. Vision and Lab, I. P. Skin cancer detection, 2024. URL r-detection/ . Accessed: 2026-01-14. Vitale, S., Orlando, J. I., Iarussi, E., and Larrabide, I. Improving realism in patient-specific abdominal ultrasound simulation using cyclegans. International journal of computer assisted radiology and surgery , 15(2):183–192, 2020.", "source": "marker_v2", "marker_block_id": "/page/11/ListGroup/531"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0080", "section": "References", "page_start": 12, "page_end": 12, "type": "ListGroup", "text": "Wang, Y. and Shi, F. KMAR-50K. Mendeley Data, V6, 2025. URL mrg7ntg.6 . Accessed: 2026-01-14. Wang, Z., Bovik, A. C., Sheikh, H. R., and Simoncelli, E. P. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing , 13(4):600–612, 2004. Wasserthal, J., Breit, H.-C., Meyer, M. T., Pradella, M., Hinck, D., Sauter, A. W., Heye, T., Boll, D. T., Cyriac, J., Yang, S., et al. Totalsegmentator: robust segmentation of 104 anatomic structures in ct images. Radiology: Artificial Intelligence , 5(5):e230024, 2023. Wu, C., Li, J., Zhou, J., Lin, J., Gao, K., Yan, K., ming Yin, S., Bai, S., Xu, X., Chen, Y., Chen, Y., Tang, Z., Zhang, Z., Wang, Z., Yang, A., Yu, B., Cheng, C., Liu, D., Li, D., Zhang, H., Meng, H., Wei, H., Ni, J., Chen, K., Cao, K., Peng, L., Qu, L., Wu, M., Wang, P., Yu, S., Wen, T., Feng, W., Xu, X., Wang, Y., Zhang, Y., Zhu, Y., Wu, Y., Cai, Y., and Liu, Z. Qwen-image technical report. Preprint at arXiv , 2025a. URL 08.02324 . Wu, C., Zheng, P., Yan, R., Xiao, S., Luo, X., Wang, Y., Li, W., Jiang, X., Liu, Y., Zhou, J., et al. Omnigen2: Exploration to advanced multimodal generation. Preprint at arXiv , 2025b. URL 2506.18871 . Yang, J., Yan, Y., Wu, G., Wang, Y., Liang, R., Jiang, X., Wan, X., Fan, F., Zhang, Y., Qin, F., and Wang, C. Medgen-bench: Contextually entangled benchmark for open-ended multimodal medical generation. Preprint at arXiv , 2025. URL 11.13135 . Yang, L., Ghosh, R. P., Franklin, J. M., Chen, S., You, C., Narayan, R. R., Melcher, M. L., and Liphardt, J. T. Nuset: A deep learning tool for reliably separating and analyzing crowded cells. PLoS computational biology , 16(9):e1008193, 2020. Younus Akon, Z. S. S. Paired ct and mri dataset for medical applications, 2025. URL om/datasets/29c3607295965ebb030f2d15 8fec487412d84c82528dd44f8ef956aef355 41aa . Accessed: 2026-01-14. Zhang, S., Zhang, Q., Zhang, S., Liu, X., Yue, J., Lu, M., Xu, H., Yao, J., Wei, X., Cao, J., et al. A generalist foundation model and database for open-world medical image segmentation. Nature Biomedical Engineering , pp. 1–16, 2025.", "source": "marker_v2", "marker_block_id": "/page/11/ListGroup/532"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0081", "section": "References", "page_start": 13, "page_end": 13, "type": "Text", "text": "Zheng, X., Wang, Y., Wang, G., and Liu, J. Fast and robust segmentation of white blood cell images by selfsupervised learning. Micron , 107:55–71, 2018.", "source": "marker_v2", "marker_block_id": "/page/12/Text/1"}
+{"paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f", "chunk_id": "80c20b7b-ead6-454a-849e-56702a6c828f:0082", "section": "References", "page_start": 13, "page_end": 13, "type": "Text", "text": "Zuo, J., Deng, H., Zhou, H., Zhu, J., Zhang, Y., Zhang, Y., Yan, Y., Huang, K., Chen, W., Deng, Y., Jin, R., Sang, N., and Gao, C. Is nano banana pro a low-level vision all-rounder? a comprehensive evaluation on 14 tasks and 40 datasets. Preprint at arXiv , 2025. URL .", "source": "marker_v2", "marker_block_id": "/page/12/Text/2"}

icml26/80c20b7b-ead6-454a-849e-56702a6c828f/reference_text_v3.txt

@@ -0,0 +1,38 @@
+[p. 9 | section: References | type: ListGroup]
+AAPM. Low dose ct grand challenge, 2016. URL https: // ect/#trainingData . Accessed: 2026-01-14. Abdi, A. H., Kasaei, S., and Mehdizadeh, M. Automatic segmentation of mandible in panoramic x-ray. Journal of Medical Imaging , 2(4):044003–044003, 2015. Ahmed, Z., Panhwar, S. Q., Baqai, A., Umrani, F. A., Ahmed, M., and Khan, A. Deep learning based automated detection of intraretinal cystoid fluid. International Jour nal of Imaging Systems and Technology , 32(3):902–917, 2022. Al-Dhabyani, W., Gomaa, M., Khaled, H., and Fahmy, A. Dataset of breast ultrasound images. Data in brief , 28: 104863, 2020. Ali, S., Ghatwary, N., Jha, D., Isik-Polat, E., Polat, G., Yang, C., Li, W., Galdran, A., Ballester, M.-A. G., Thambawita, ´ V., et al. Assessing generalisability of deep learningbased polyp detection and segmentation methods through a computer vision challenge. Scientific Reports , 14(1): 2032, 2024. Angiographics. Chuac dataset. URL re.com/s/4d24cf3d14bc901a94bf . Accessed: 2026-01-14. Antonelli, M., Reinke, A., Bakas, S., Farahani, K., Kopp-Schneider, A., Landman, B. A., Litjens, G., Menze, B., Ronneberger, O., Summers, R. M., et al. The medical segmentation decathlon. Nature communications , 13(1): 4128, 2022. Aysen, D., Muhammad, C., Tawsifur, R., Amith, K., Yazan, Q., Mete, A., Anas M., T., and serkan, k. Qata-cov19 dataset, 2024. URL atasets/aysendegerli/qatacov19-datas et . Accessed: 2026-01-14. Bai, S., Cai, Y., Chen, R., Chen, K., Chen, X., Cheng, Z., Deng, L., Ding, W., Gao, C., Ge, C., Ge, W., Guo, Z., Huang, Q., Huang, J., Huang, F., Hui, B., Jiang, S., Li, Z., Li, M., Li, M., Li, K., Lin, Z., Lin, J., Liu, X., Liu, J., Liu, C., Liu, Y., Liu, D., Liu, S., Lu, D., Luo, R., Lv, C., Men, R., Meng, L., Ren, X., Ren, X., Song, S., Sun, Y., Tang, J., Tu, J., Wan, J., Wang, P., Wang, P.,
+
+[p. 9 | section: References | type: ListGroup]
+Wang, Q., Wang, Y., Xie, T., Xu, Y., Xu, H., Xu, J., Yang, Z., Yang, M., Yang, J., Yang, A., Yu, B., Zhang, F., Zhang, H., Zhang, X., Zheng, B., Zhong, H., Zhou, J., Zhou, F., Zhou, J., Zhu, Y., and Zhu, K. Qwen3-vl technical report. Preprint at arXiv , 2025a. URL https: //arxiv.org/abs/2511.21631 . Bai, S., Chen, K., Liu, X., Wang, J., Ge, W., Song, S., Dang, K., Wang, P., Wang, S., Tang, J., Zhong, H., Zhu, Y., Yang, M., Li, Z., Wan, J., Wang, P., Ding, W., Fu, Z., Xu, Y., Ye, J., Zhang, X., Xie, T., Cheng, Z., Zhang, H., Yang, Z., Xu, H., and Lin, J. Qwen2.5-vl technical report. Preprint at arXiv , 2025b. URL g/abs/2502.13923 . Brooks, T., Holynski, A., and Efros, A. A. Instructpix2pix: Learning to follow image editing instructions. In Pro ceedings of the IEEE/CVF conference on computer vision and pattern recognition , pp. 18392–18402, 2023. Buda, M., Saha, A., and Mazurowski, M. A. Association of genomic subtypes of lower-grade gliomas with shape features automatically extracted by a deep learning algorithm. Computers in biology and medicine , 109:218–225, 2019. Caicedo, J. C., Goodman, A., Karhohs, K. W., Cimini, B. A., Ackerman, J., Haghighi, M., Heng, C., Becker, T., Doan, M., McQuin, C., et al. Nucleus segmentation across imaging experiments: the 2018 data science bowl. Nature methods , 16(12):1247–1253, 2019. Carballal, A., Novoa, F. J., Fernandez-Lozano, C., Garc´ıa-Guimaraes, M., Aldama-Lopez, G., Calvi ´ no-Santos, R., ˜ Vazquez-Rodriguez, J. M., and Pazos, A. Automatic multiscale vascular image segmentation algorithm for coronary angiography. Biomedical Signal Processing and Control , 46:1–9, 2018. Cervantes-Sanchez, F., Cruz-Aceves, I., Hernandez-Aguirre, A., Hernandez-Gonzalez, M. A., and Solorio-Meza, S. E. Automatic segmentation of coronary arteries in x-ray angiograms using multiscale analysis and artificial neural networks. Applied Sciences , 9(24):5507, 2019. Chen, Z. and Feng, M. Med-banana-50k: A cross-modality large-scale dataset for text-guided medical image editing. Preprint at arXiv , 2025. URL g/abs/2511.00801 . Chowdhury, M. E., Rahman, T., Khandakar, A., Mazhar, R., Kadir, M. A., Mahbub, Z. B., Islam, K. R., Khan, M. S., Iqbal, A., Al Emadi, N., et al. Can ai help in screening viral and covid-19 pneumonia? Ieee Access , 8: 132665–132676, 2020.
+
+[p. 10 | section: References | type: Text]
+495 496 497 498 499 500 Codella, N., Rotemberg, V., Tschandl, P., Celebi, M. E., Dusza, S., Gutman, D., Helba, B., Kalloo, A., Liopyris, K., Marchetti, M., et al. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic). arXiv preprint arXiv:1902.03368 , 2019.
+
+[p. 10 | section: References | type: ListGroup]
+Dao, B. T., Nguyen, T. V., Pham, H. H., and Nguyen, H. Q. Phase recognition in contrast-enhanced ct scans based on deep learning and random sampling. Medical Physics , 49 (7):4518–4528, 2022. de Verdier, M. C., Saluja, R., Gagnon, L., LaBella, D., Baid, U., Tahon, N. H., Foltyn-Dumitru, M., Zhang, J., Alafif, M., Baig, S., Chang, K., D'Anna, G., Deptula, L., Gupta, D., Haider, M. A., Hussain, A., Iv, M., Kontzialis, M., Manning, P., Moodi, F., Nunes, T., Simon, A., Sollmann, N., Vu, D., Adewole, M., Albrecht, J., Anazodo, U., Chai, R., Chung, V., Faghani, S., Farahani, K., Kazerooni, A. F., Iglesias, E., Kofler, F., Li, H., Linguraru, M. G., Menze, B., Moawad, A. W., Velichko, Y., Wiestler, B., Altes, T., Basavasagar, P., Bendszus, M., Brugnara, G., Cho, J., Dhemesh, Y., Fields, B. K. K., Garrett, F., Gass, J., Hadjiiski, L., Hattangadi-Gluth, J., Hess, C., Houk, J. L., Isufi, E., Layfield, L. J., Mastorakos, G., Mongan, J., Nedelec, P., Nguyen, U., Oliva, S., Pease, M. W., Rastogi, A., Sinclair, J., Smith, R. X., Sugrue, L. P., Thacker, J., Vidic, I., Villanueva-Meyer, J., White, N. S., Aboian, M., Conte, G. M., Dale, A., Sabuncu, M. R., Seibert, T. M., Weinberg, B., Abayazeed, A., Huang, R., Turk, S., Rauschecker, A. M., Farid, N., Vollmuth, P., Nada, A., Bakas, S., Calabrese, E., and Rudie, J. D. The 2024 brain tumor segmentation (brats) challenge: Glioma segmentation on post-treatment mri. Preprint at arXiv , 2024. URL 8368 . DeepMind, G. Gemini 3 pro: High-precision multimodal reasoning, 2025a. URL gle/models/gemini-image/pro/ . Accessed: 2026-01-14. DeepMind, G. Imagen 4.0: Model documentation and generation guide, 2025b. URL loud.google.com/vertex-ai/generative -ai/docs/models/imagen/4-0-generate . Accessed: 2026-01-14. Deng, C., Zhu, D., Li, K., Gou, C., Li, F., Wang, Z., Zhong, S., Yu, W., Nie, X., Song, Z., Shi, G., and Fan, H. Emerging properties in unified multimodal pretraining. Preprint at arXiv , 2025. URL https: //arxiv.org/abs/2505.14683 . Dietler, N., Minder, M., Gligorovski, V., Economou, A. M., Joly, D. A. H. L., Sadeghi, A., Chan, C. H. M., Kozinski, ´
+
+[p. 10 | section: References | type: ListGroup]
+M., Weigert, M., Bitbol, A.-F., et al. A convolutional neural network segments yeast microscopy images with high accuracy. Nature communications , 11(1):5723, 2020. Duong, V. H., Vu, H., Phan, H. D., Nguyen, D. Q., Pham, D. H., Le, Q. T., Nguyen, B. S., Do, T. D., Dinh, V. S., Nguyen, T. C., et al. Thyroidxl: Advancing thyroid nodule diagnosis with an expert-labeled, pathology-validated dataset. In International Conference on Medical Image Computing and Computer-Assisted Intervention , pp. 616– 626. Springer, 2025. Fang, H., Li, F., Wu, J., Fu, H., Sun, X., Son, J., Yu, S., Zhang, M., Yuan, C., Bian, C., Lei, B., Zhao, B., Xu, X., Li, S., Fumero, F., Sigut, J., Almubarak, H., Bazi, Y., Guo, Y., Zhou, Y., Baid, U., Innani, S., Guo, T., Yang, J., Orlando, J. I., Bogunovic, H., Zhang, X., and Xu, Y. ´ Refuge2 challenge: A treasure trove for multi-dimension analysis and evaluation in glaucoma screening. Preprint at arXiv , 2022. URL 2202.08994 . Garcia-Peraza-Herrera, L. C., Fidon, L., D'Ettorre, C., Stoyanov, D., Vercauteren, T., and Ourselin, S. Image compositing for segmentation of surgical tools without manual annotations. IEEE transactions on medical imaging , 40(5):1450–1460, 2021. Gutman, D., Codella, N. C., Celebi, E., Helba, B., Marchetti, M., Mishra, N., and Halpern, A. Skin lesion analysis toward melanoma detection: A challenge at the international symposium on biomedical imaging (isbi) 2016, hosted by the international skin imaging collaboration (isic). arXiv preprint arXiv:1605.01397 , 2016. Haneda, E., Peters, N., Zhang, J., Karageorgos, G., Xia, W., Paganetti, H., Wang, G., Guo, Y., Ma, J., Park, H. S., et al. Aapm ct metal artifact reduction grand challenge. Medical physics , 52(10):e70050, 2025. Huang, Z., Zheng, D., Zou, C., Liu, R., Wang, X., Ji, K., Chai, W., Sun, J., Wang, L., Lv, Y., Huang, T., Liu, J., Guo, Q., Yang, M., Chen, J., and Zhou, J. Ming-univision: Joint image understanding and generation with a unified continuous tokenizer. Preprint at arXiv , 2025. URL . Isensee, F., Jaeger, P. F., Kohl, S. A., Petersen, J., and Maier-Hein, K. H. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods , 18(2):203–211, 2021. Jha, D., Smedsrud, P. H., Riegler, M. A., Halvorsen, P., De Lange, T., Johansen, D., and Johansen, H. D. Kvasirseg: A segmented polyp dataset. In International con ference on multimedia modeling , pp. 451–462. Springer, 2019.
+
+[p. 11 | section: References | type: Text]
+550 Kus¸, Z. and Aydin, M. Medsegbench: A comprehensive benchmark for medical image segmentation in diverse data modalities. Scientific Data , 11(1):1283, 2024.
+
+[p. 11 | section: References | type: ListGroup]
+Labs, B. F., Batifol, S., Blattmann, A., Boesel, F., Consul, S., Diagne, C., Dockhorn, T., English, J., English, Z., Esser, P., Kulal, S., Lacey, K., Levi, Y., Li, C., Lorenz, D., Muller, J., Podell, D., Rombach, R., Saini, H., Sauer, ¨ A., and Smith, L. Flux.1 kontext: Flow matching for in-context image generation and editing in latent space. Preprint at arXiv , 2025. URL g/abs/2506.15742 . Lei, W., Chen, H., Zhang, Z., Luo, L., Xiao, Q., Gu, Y., Gao, P., Jiang, Y., Wang, C., Wu, G., Xu, T., Zhang, Y., Rajpurkar, P., Zhang, X., Zhang, S., and Wang, Z. A synthetic data-driven radiology foundation model for pantumor clinical diagnosis. Preprint at arXiv , 2025. URL . Litjens, G., Toth, R., Van De Ven, W., Hoeks, C., Kerkstra, S., Van Ginneken, B., Vincent, G., Guillard, G., Birbeck, N., Zhang, J., et al. Evaluation of prostate segmentation algorithms for mri: the promise12 challenge. Medical image analysis , 18(2):359–373, 2014. Liu, J., Liu, G., Liang, J., Li, Y., Liu, J., Wang, X., Wan, P., Zhang, D., and Ouyang, W. Flow-grpo: Training flow matching models via online rl. arXiv preprint at arXiv , 2025a. URL 5470 . Liu, M., He, Z., Fan, Z., Wang, Q., and Fung, Y. R. Medebench: Diagnosing reliability in text-guided medical image editing. In Findings of the Association for Computational Linguistics: EMNLP 2025 , pp. 767–791, 2025b. Liu, S., Han, Y., Xing, P., Yin, F., Wang, R., Cheng, W., Liao, J., Wang, Y., Fu, H., Han, C., Li, G., Peng, Y., Sun, Q., Wu, J., Cai, Y., Ge, Z., Ming, R., Xia, L., Zeng, X., Zhu, Y., Jiao, B., Zhang, X., Yu, G., and Jiang, D. Step1xedit: A practical framework for general image editing. Preprint at arXiv , 2025c. URL g/abs/2504.17761 . Ljosa, V., Sokolnicki, K. L., and Carpenter, A. E. Annotated high-throughput microscopy image sets for validation. Nature methods , 9(7):637, 2012. Lu, Y., Zhou, M., Zhi, D., Zhou, M., Jiang, X., Qiu, R., Ou, Z., Wang, H., Qiu, D., Zhong, M., et al. The jnu-ifm dataset for segmenting pubic symphysis-fetal head. Data in brief , 41:107904, 2022. Mathieu, G., Bachir, E. D., et al. Brifiseg: a deep learningbased method for semantic and instance segmentation
+
+[p. 11 | section: References | type: ListGroup]
+of nuclei in brightfield images. Preprint at arXiv , 2022. URL . Montoya, A. Ultrasound nerve segmentation, 2026. URL ultrasound-nerve-segmentation . Accessed: 2026-01-14. Morozov, S. P., Andreychenko, A. E., Pavlov, N. A., Vladzymyrskyy, A., Ledikhova, N. V., Gombolevskiy, V. A., Blokhin, I. A., Gelezhe, P. B., Gonchar, A., and Chernina, V. Y. Mosmeddata: Chest ct scans with covid-19 related findings dataset. arXiv preprint arXiv:2005.06465 , 2020. Narai, ´ A., Hermann, P., Auer, T., Kemenczky, P., Szalma, ´ J., Homolya, I., Somogyi, E., Vakli, P., Weiss, B., and Vidnyanszky, Z. Movement-related artefacts (mr-art) ´ dataset of matched motion-corrupted and clean structural mri brain scans. Scientific data , 9(1):630, 2022. Naylor, P., Lae, M., Reyal, F., and Walter, T. Segmentation ´ of nuclei in histopathology images by deep regression of the distance map. IEEE transactions on medical imaging , 38(2):448–459, 2018. Ngoc Lan, P., An, N. S., Hang, D. V., Long, D. V., Trung, T. Q., Thuy, N. T., and Sang, D. V. Neounet: Towards accurate colon polyp segmentation and neoplasm detection. In International Symposium on Visual Computing , pp. 15–28. Springer, 2021. OpenAI. Gpt image 1 model documentation, 2025. URL els/gpt-image-1 . Accessed: 2026-01-14. Pandey, N. Chest xray masks and labels, 2025. URL ht tps:// pandey360/chest-xray-masks-and-labels . Accessed: 2026-01-14. pazhoulab. Low-dose ct reconstruction contest, 2024. URL ontestdetail?id=667e0c687ff47da8cc82 7679&award=1,000,000 . Accessed: 2026-01-14. Polo, M. Chest ct segmentation, 2025. URL https: // chest-ct-segmentation . Accessed: 2026-01-14. Porwal, P., Pachade, S., Kamble, R., Kokare, M., Deshmukh, G., Sahasrabuddhe, V., and Meriaudeau, F. Indian diabetic retinopathy image dataset (idrid): a database for diabetic retinopathy screening research. Data , 3(3):25, 2018. Sauer, A., Lorenz, D., Blattmann, A., and Rombach, R. Adversarial diffusion distillation. In European Conference on Computer Vision , pp. 87–103. Springer, 2024.
+
+[p. 12 | section: References | type: Text]
+Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., Bi, X., Zhang, H., Zhang, M., Li, Y. K., Wu, Y., and Guo, D. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. Preprint at arXiv , 2024. URL .
+
+[p. 12 | section: References | type: ListGroup]
+Song, Y., Zheng, J., Lei, L., Ni, Z., Zhao, B., and Hu, Y. Ct2us: Cross-modal transfer learning for kidney segmentation in ultrasound images with synthesized data. Ultrasonics , 122:106706, 2022. Spahn, C., Gomez-de Mariscal, E., Laine, R. F., Pereira, ´ P. M., von Chamier, L., Conduit, M., Pinho, M. G., Jacquemet, G., Holden, S., Heilemann, M., et al. Deepbacs for multi-task bacterial image analysis using opensource deep learning approaches. Communications Biol ogy , 5(1):688, 2022. Staal, J., Abramoff, M. D., Niemeijer, M., Viergever, M. A., ` and Van Ginneken, B. Ridge-based vessel segmentation in color images of the retina. IEEE transactions on medi cal imaging , 23(4):501–509, 2004. Tahir, A. M., Chowdhury, M. E., Khandakar, A., Rahman, T., Qiblawey, Y., Khurshid, U., Kiranyaz, S., Ibtehaz, N., Rahman, M. S., Al-Maadeed, S., et al. Covid-19 infection localization and severity grading from chest x-ray images. Computers in biology and medicine , 139:105002, 2021. Tong, S., Fan, D., Li, J., Xiong, Y., Chen, X., Sinha, K., Rabbat, M., LeCun, Y., Xie, S., and Liu, Z. Metamorph: Multimodal understanding and generation via instruction tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pp. 17001–17012, 2025. Van Valen, D. A., Kudo, T., Lane, K. M., Macklin, D. N., Quach, N. T., DeFelice, M. M., Maayan, I., Tanouchi, Y., Ashley, E. A., and Covert, M. W. Deep learning automates the quantitative analysis of individual cells in live-cell imaging experiments. PLoS computational biology , 12(11):e1005177, 2016. Verma, R., Kumar, N., Patil, A., Kurian, N. C., Rane, S., Graham, S., Vu, Q. D., Zwager, M., Raza, S. E. A., Rajpoot, N., et al. Monusac2020: A multi-organ nuclei segmentation and classification challenge. IEEE Transac tions on Medical Imaging , 40(12):3413–3423, 2021. Vision and Lab, I. P. Skin cancer detection, 2024. URL r-detection/ . Accessed: 2026-01-14. Vitale, S., Orlando, J. I., Iarussi, E., and Larrabide, I. Improving realism in patient-specific abdominal ultrasound simulation using cyclegans. International journal of computer assisted radiology and surgery , 15(2):183–192, 2020.
+
+[p. 12 | section: References | type: ListGroup]
+Wang, Y. and Shi, F. KMAR-50K. Mendeley Data, V6, 2025. URL mrg7ntg.6 . Accessed: 2026-01-14. Wang, Z., Bovik, A. C., Sheikh, H. R., and Simoncelli, E. P. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing , 13(4):600–612, 2004. Wasserthal, J., Breit, H.-C., Meyer, M. T., Pradella, M., Hinck, D., Sauter, A. W., Heye, T., Boll, D. T., Cyriac, J., Yang, S., et al. Totalsegmentator: robust segmentation of 104 anatomic structures in ct images. Radiology: Artificial Intelligence , 5(5):e230024, 2023. Wu, C., Li, J., Zhou, J., Lin, J., Gao, K., Yan, K., ming Yin, S., Bai, S., Xu, X., Chen, Y., Chen, Y., Tang, Z., Zhang, Z., Wang, Z., Yang, A., Yu, B., Cheng, C., Liu, D., Li, D., Zhang, H., Meng, H., Wei, H., Ni, J., Chen, K., Cao, K., Peng, L., Qu, L., Wu, M., Wang, P., Yu, S., Wen, T., Feng, W., Xu, X., Wang, Y., Zhang, Y., Zhu, Y., Wu, Y., Cai, Y., and Liu, Z. Qwen-image technical report. Preprint at arXiv , 2025a. URL 08.02324 . Wu, C., Zheng, P., Yan, R., Xiao, S., Luo, X., Wang, Y., Li, W., Jiang, X., Liu, Y., Zhou, J., et al. Omnigen2: Exploration to advanced multimodal generation. Preprint at arXiv , 2025b. URL 2506.18871 . Yang, J., Yan, Y., Wu, G., Wang, Y., Liang, R., Jiang, X., Wan, X., Fan, F., Zhang, Y., Qin, F., and Wang, C. Medgen-bench: Contextually entangled benchmark for open-ended multimodal medical generation. Preprint at arXiv , 2025. URL 11.13135 . Yang, L., Ghosh, R. P., Franklin, J. M., Chen, S., You, C., Narayan, R. R., Melcher, M. L., and Liphardt, J. T. Nuset: A deep learning tool for reliably separating and analyzing crowded cells. PLoS computational biology , 16(9):e1008193, 2020. Younus Akon, Z. S. S. Paired ct and mri dataset for medical applications, 2025. URL om/datasets/29c3607295965ebb030f2d15 8fec487412d84c82528dd44f8ef956aef355 41aa . Accessed: 2026-01-14. Zhang, S., Zhang, Q., Zhang, S., Liu, X., Yue, J., Lu, M., Xu, H., Yao, J., Wei, X., Cao, J., et al. A generalist foundation model and database for open-world medical image segmentation. Nature Biomedical Engineering , pp. 1–16, 2025.
+
+[p. 13 | section: References | type: Text]
+Zheng, X., Wang, Y., Wang, G., and Liu, J. Fast and robust segmentation of white blood cell images by selfsupervised learning. Micron , 107:55–71, 2018.
+
+[p. 13 | section: References | type: Text]
+Zuo, J., Deng, H., Zhou, H., Zhu, J., Zhang, Y., Zhang, Y., Yan, Y., Huang, K., Chen, W., Deng, Y., Jin, R., Sang, N., and Gao, C. Is nano banana pro a low-level vision all-rounder? a comprehensive evaluation on 14 tasks and 40 datasets. Preprint at arXiv , 2025. URL .

icml26/80c20b7b-ead6-454a-849e-56702a6c828f/sanitization_report.json

@@ -0,0 +1,61 @@
+{
+  "paper_id": "80c20b7b-ead6-454a-849e-56702a6c828f",
+  "pipeline": "Paper2Markdown-V3",
+  "ok": true,
+  "page_count": 24,
+  "chunk_count": 140,
+  "main_body_chunk_count": 70,
+  "appendix_chunk_count": 57,
+  "reference_chunk_count": 13,
+  "model_text_chars": 34560,
+  "raw_markdown_chars": 87319,
+  "sanitized_chars": 77147,
+  "page_provenance": {
+    "min_page": 1,
+    "max_page": 24,
+    "invalid_count": 0
+  },
+  "marker_block_type_counts": {
+    "Caption": 3,
+    "Equation": 3,
+    "Figure": 1,
+    "FigureGroup": 9,
+    "Footnote": 1,
+    "ListGroup": 19,
+    "PageFooter": 24,
+    "PageHeader": 31,
+    "Picture": 3,
+    "PictureGroup": 15,
+    "SectionHeader": 55,
+    "TableGroup": 6,
+    "Text": 367
+  },
+  "asset_count_raw": 28,
+  "asset_count_model_kept": 13,
+  "asset_count_rejected": 15,
+  "asset_reject_reasons": {
+    "extreme_aspect_ratio": 15,
+    "kept": 13
+  },
+  "artifact_leak_audit": {
+    "ok": true,
+    "hits": {
+      "Anonymous Authors": [],
+      "ACKNOWLEDGMENT": [],
+      "OpenReview": [],
+      "\"accept_label\"": [],
+      "\"decision\"": [],
+      "\"decision_tier\"": [],
+      "\"source_status\"": [],
+      "Meta-review": [],
+      "Official Review": [],
+      "official_reviews": [],
+      "meta_reviews": [],
+      "suggested_verdict_score": []
+    },
+    "artifact_count": 2
+  },
+  "default_model_input": "model_text_v3.txt",
+  "appendix_input": "appendix_text_v3.txt",
+  "reference_input": "reference_text_v3.txt"
+}

icml26/80c20b7b-ead6-454a-849e-56702a6c828f/sanitized_v3.txt

@@ -0,0 +1,466 @@
+{0}
+# Abstract
+The scarcity of high-quality data remains a primary bottleneck in adapting multimodal generative models for medical image editing. Existing medical image editing datasets often suffer from limited diversity, neglect of medical image understanding and inability to balance quality with scalability. To address these gaps, we propose MieDB-100k, a large-scale, high-quality and diverse dataset for text-guided medical image editing. It categorizes editing tasks into perspectives of Perception, Modification and Transformation, considering both understanding and generation abilities. We construct MieDB-100k via a data curation pipeline leveraging both modality-specific expert models and rule-based data synthetic methods, followed by rigorous manual inspection to ensure clinical fidelity. Extensive experiments demonstrate that model trained with MieDB-100k consistently outperform both open-source and proprietary models while exhibiting strong generalization ability. We anticipate that this dataset will serve as a cornerstone for future advancements in specialized medical image editing. Dataset and code are publicly available at [ [ub.com/Raiiyf/MieDB-100k](
+# 1. Introduction
+Multimodal generative models [\(Wu et al.,](#page-11-0) [2025a](#page-11-0)[;b;](#page-11-1) [Liu](#page-10-0) [et al.,](#page-10-0) [2025c\)](#page-10-0) have developed rapidly in recent years. In natural image domains, generative models are not only gradually unifying text-guided generation and editing tasks, but also progressively expanding their capabilities to encompass image modification and image understanding [\(Deng](#page-9-0) [et al.,](#page-9-0) [2025;](#page-9-0) [Tong et al.,](#page-11-2) [2025\)](#page-11-2). However, in medical image domains, their performance remains conspicuously limited, especially in the area of unified editing tasks [\(Liu et al.,](#page-10-1)
+[2025b;](#page-10-1) [Yang et al.,](#page-11-3) [2025\)](#page-11-3). We attribute this performance degradation primarily to a fundamental scarcity of specialized medical image-editing data.
+While a few contemporary studies have proposed benchmarks or datasets for medical image editing, they remain insufficient in three key aspects: (1) limited diversity in medical image modalities. Unlike general computer vision, clinical imaging encompasses diverse modalities with distinct physical and structural foundations. However, existing research and datasets are restricted to a narrow range of imaging modalities [\(Chen & Feng,](#page-8-0) [2025;](#page-8-0) [Liu et al.,](#page-10-1) [2025b\)](#page-10-1), typically the widely available modalities such as Chest Xrays and CTs, which cannot adequately train or evaluate a model's ability across diverse clinical settings.
+- (2) Neglect of medical image understanding. Almost all medical image editing works only focus on conceptual modification and stylistic transformation tasks, but ignore visual perception tasks (*e.g.* organ/lesion detection), which has been considered to be beneficial to the generation of image editing models [\(Huang et al.,](#page-9-1) [2025;](#page-9-1) [Deng et al.,](#page-9-0) [2025\)](#page-9-0). Additionally, this clinical grounding ensures interpretability and corrects 'right-for-the-wrong-reason' edits, which is vital for safety-critical medical applications. Finally, neglecting understanding tasks also hinders the development of unified medical models that bridge understanding and generation.
+- (3) Failure to ensure both data quality and scalability. Collection of medical image editing data is hindered by the difficulty of generating ground-truth counterfactuals. Some existing studies [\(Chen & Feng,](#page-8-0) [2025;](#page-8-0) [Yang et al.,](#page-11-3) [2025\)](#page-11-3) distills general-purpose generative models for quick data scaling. However, these models are not tailored for medical use and hence produce results that lack clinical reliability and explainability. Conversely, previous work [\(Liu et al.,](#page-10-1) [2025b\)](#page-10-1) relies on extensive human involvement to manually collect real medical image pairs, which is notoriously difficult to scale up. Moreover, real-world longitudinal data often exhibits spatial misalignment and background inconsistency, as obtaining perfectly calibrated scan pairs is rare in practical medical settings.
+In this paper, we address the aforementioned limitations in previous research by introducing MieDB-100k, a largescale, high-quality and diverse dataset for text-guided medical image editing. MieDB-100k includes 112, 228 editing
+{1}------------------------------------------------
+![](_page_1_Figure_1.jpeg)
+Figure 1. MieDB-100k overview. It categorizes medical image editing tasks into three perspectives, covering diverse medical modalities.
+data, covering **69** distinct editing targets and **10** diverse medical image modalities. We categorize editing tasks into three types: **Perception**, **Modification** and **Transformation**, which consider both model's intrinsic understanding and generation abilities on medical images. To enhance the data fidelity while preserving the scalability, we propose a data curation pipeline leveraging both modality-specific expert models and rule-based data synthetic methods. Additionally, for some complex tasks such as lesion modification, we introduce individuals with medical knowledge to perform manual quality checks on the data to ensure data quality. Finally, we introduced task-specific evaluation metrics to facilitate a comprehensive assessment of the editing models' performance.
+108 109 We evaluate existing open-source and closed-source multimodal generative models on MieDB-100k and argue that most of them cannot perform well in medical image editing. To further validate the reliability and utility of MieDB-100k, we finetune the OmniGen2 baseline on our dataset. Experimental results demonstrate that MieDB-100k facilitates a substantial performance leap in medical image editing tasks, surpassing or matching SOTA models including Nano Banana Pro. It also exhibits strong generalization ability driven by the synergy of understanding and generation tasks. We anticipate that this dataset will serve as a cornerstone for future advancements in specialized medical image editing.
+Our contributions can be summarized as follows:
+- (1) We propose a credible and scalable data curation pipeline to construct **MieDB-100k**, a large-scale, high-quality and highly diverse dataset for medical image editing with 69 targets and 10 medical image modalities.
+- (2) We first unify the medical image understanding and generation into the paradigm of edit, and find that joint
+training yields performance gains for specific tasks.
+(3) We evaluate popular open-source and closed-source multimodal generative models on **MieDB-100k**, and observe that training with our data can significantly strengthens the model's capacity for medical image editing.
+#### 2. Related Work
+### 2.1. Data Research for Medical Image Editing
+As an emerging area, multimodal medical generative modeling is currently supported by relatively few publicly available datasets for training and benchmarking (Tab. 1). In these works, the primary challenge lies in the construction of high-quality image-edit pairs. MedEBench (Liu et al., 2025b), an early benchmarking effort, curated pairs by manually collecting related images from medical documents. While this ensures clinical validity, the approach lacks scalability. Furthermore, the resulting image pairs often exhibit background inconsistencies, as achieving strict spatial calibration in real-world clinical settings is virtually impossible. Conversely, Med-banana-50K (Chen & Feng, 2025) proposed a fully autonomous pipeline where data construction and quality control were managed by Gemini. However, applying general-purpose models to specialized medical scenarios may introduce factual errors or inconsistent edits, raising concerns about data fidelity. Finally, MedGEN-Bench (Yang et al., 2025) introduced image-edit pairs using a mix of rule-based and model-based methods; however, the lack of specific architectural details hinders a thorough evaluation of their data quality. Moreover, existing benchmarks only focus on content generation evaluation, overlooking the critical aspect of medical image understanding.
+{2}------------------------------------------------
+<span id="page-2-0"></span>*Table 1.* Comparison of contemporary medical image editing benchmarks and datasets. In 'Perspective' column, P stands for Perception, M stands for Modification, and T stands for Transformation.
+| Benchmark | Size | Modalities | Targets | Perspectives | Source | <b>Human Inspection</b> |
+|------------------------------------|-------------|------------|---------|--------------|------------------|-------------------------|
+| MedE-Bench (Liu et al., 2025b) | $\sim 1 k$ | 4 | 13 | M | Real | ✓ |
+| Med-banana-50K (Chen & Feng, 2025) | $\sim$ 50k | 3 | 23 | M | Synthetic | X |
+| MedGEN-Bench (Yang et al., 2025) | $\sim$ 6k | 6 | 16 | M, T | Real & Synthetic | ✓ |
+| MieDB-100k (Ours) | $\sim$ 100k | 10 | 69 | P, M, T | Real & Synthetic | <b>✓</b> |
+<span id="page-2-1"></span>![](_page_2_Figure_3.jpeg)
+Figure 2. Modality distribution (a) and prompt word cloud (b).
+#### 2.2. Multimodal Generative Model
+129130
+131132
+Multimodal generative models (Liu et al., 2025c; Brooks et al., 2023) accept both images and natural language instructions as input, performing edits by translating semantic commands into precise visual manipulations. Recent studies (Wu et al., 2025a) often leverage vision-language model encoder and large-scale vision-language pretraining to align the semantic instruction with image modification. For instance, OmniGen2 (Wu et al., 2025b) utilizes Qwen2.5-VL (Bai et al., 2025b) to extract latent representations for semantic alignment, supported by a large-scale, multi-task training strategy. Furthermore, many recent studies (Deng et al., 2025) integrate image understanding and editing within a unified architecture. Exploiting these synergies is essential for creating robust models that are capable of performing both multimodal understanding and visual generation. On the commercial front, SOTA proprietary models like Gemini-3-Pro-Image (Nano Banana Pro) (DeepMind, 2025a) exhibit sophisticated image manipulation abilities, further realizing the real-world potential of multi-modal generative models. Despite these advancements, current models still struggle with the complexities of medical imaging(Liu et al., 2025b; Yang et al., 2025), highlighting the urgent need for comprehensive datasets to accelerate their adaptation to clinical domains.
+#### 3. MieDB-100k
+This section introduces MieDB-100k, a high-quality, rigorous, and highly diverse dataset for medical image editing with more than 69 associated medical targets. It contains
+112, 228 image-editing triplets. Figure 2(a) summarizes the distribution of samples across 10 imaging modalities.
+#### 3.1. Data Definition
+Each entry in MieDB-100k is a triplet (I, P, O), where I is the input medical image, P is the textual prompt that describes edit operation, and O is the target image.
+#### 3.2. Three Perspectives of MieDB-100k
+MieDB-100k is constructed under a novel categorization of three perspectives, considering both understanding and generation capabilities: (1) **Perception** tasks, which focus on model's intrinsic medical knowledge via pixel-wise identification of prompted clinical targets in the input image; (2) **Modification** tasks, which require the model to locate and alter specific medical features; and (3) **Transformation** tasks, involving medical image restoration, enhancement, and other low-level transformation. To ensure the rigor of the data triplets while maintaining scalability, we designed and implemented a specialized data construction pipeline for MieDB-100k (Fig 3), and we list all source datasets used for construction in App. A.
+# 3.2.1. Perception
+Perception tasks focus on medical image understanding, and we we formulate it as an editing task by instructing model to generate masks over regions of interest (ROIs), such as specific organs or lesions, through textual prompts. Notably, to align with image editing paradigm, the model is prompted to overlay the localization mask directly onto the source image rather than generating a standalone binary mask. This task serves two primary functions: First, since the mask-painting task only requires minimal pixel manipulation (typically modifying a single channel within a specific region), it serves as a direct assessment of the medical knowledge embedded in the generative model, isolating its perceptual accuracy from complex synthesis capabilities. Second, it introduces a promising application for multimodal generative models in the medical domain: assisted interpretation in multimodal manner. By allowing users to highlight specific targets in medical image through natural language prompts, this approach can assist
+{3}------------------------------------------------
+<span id="page-3-0"></span>![](_page_3_Figure_1.jpeg)
+patients in understanding their diagnostic images, aid medical students in their education, and reduce screening time for senior clinicians.
+The rule-based construction process for the Perception task's data triplets is illustrated in Fig. 3. Specifically, for a segmentation dataset, the original image serves as the input I. The output image O is synthesized by overlaying the ground-truth segmentation label, which is rendered in a randomly selected color (red, green, or blue), onto the input image with an alpha-blending transparency of 0.6. The ROIs of perception can be classified into three types: anatomical structure (organ, organism and so on), lesion area and holistic segmentation (segment all visible and clinically significant structures). We specifies the perception target and visualizing color scheme in the textual prompt P. Since this part of the data is constructed following a definite rule, it can be readily scaled up to a diverse set of medical knowledge assessments and to the associated training dataset by leveraging the extensive body of existing medical segmentation research. Finally, to ensure a high-quality final benchmark, we manually filtered the initial data pool to remove trivial, redundant, or incorrectly labeled samples.
+#### 3.2.2. MODIFICATION
+167168
+171172
+177178179
+The perspective of Modification is specifically designed for semantically modifying medical contents, so as to address the diverse requirements of editing beyond just locate them. However, constructing modification data triplets is challenging because counterfactual image pairs cannot be captured simultaneously in the real world. While one could theoretically leverage general-purpose generative models (e.g., Nano Banana Pro or Qwen-Image-Edit) to produce these edits, such models are not specialized for the medical domain, and therefore are prone to severe hallucinations, which is unacceptable in a healthcare context. To construct rigorous edit triplets and preserve scalability, we propose a four-stage process (Fig. 3) designed to bridge the gap between task complexity and model competence so as to fully utilize these automatic tools.
+Stage I: We develop a suite of modality-specific expert models for healthy tissue inpainting, built upon the FLUX.1-Fill-dev model. This strategy is based on the observation that generating healthy anatomical structures is more stable and predictable than generating lesions, as the former exhibits more tractable patterns and textures. For each modality, we curate a training dataset consisting exclusively of non-pathological samples from existing medical image repositories. Through parameter-efficient finetuning, these models learn to inpaint masked areas with high clinical accuracy. We further apply background restoration and edge blending to correct any unintended modifications made by the FLUX model outside the mask, ensuring the edited region blends seamlessly into the original image.
+Stage II: We leverage these expert models to modify lesion-bearing images (L) into their counterfactual 'healthy' results (H). Specifically, we fill the lesion area
+{4}------------------------------------------------
+in L using white pixels based on its ground-truth segmentation label. This masked image and its corresponding binary mask are then processed by the modality-matched expert model to synthesize H, where healthy tissue replaces lesion. Compared to distilling general-purpose generative models, our modality-specific approach not only restricts the high-variance generative process to a localized region to guarantee background consistency during the edit, but also ensures that tasks remain within the model's learned distribution, thereby significantly reducing hallucinations. Furthermore, unlike manual data collection from the internet, our approach provides superior scalability and efficiency.
+Stage III: We implement a rejection sampling mechanism for the generated 'healthy' images (H) to further enhance the data quality within the Modification tasks. For modalities that resemble natural images (e.g., endoscopy and dermoscopy), we prompt the Qwen3-VL-32B-Instruct model [\(Bai et al.,](#page-8-3) [2025a\)](#page-8-3) to filter out H that still contain lesions, exhibit artifacts, or are of low quality. For other modalities, we train separate nnUNet models [\(Isensee et al.,](#page-9-3) [2021\)](#page-9-3) for lesion segmentation and discard H where lesions remain detectable.
+Stage IV: Triplet combination. Using these high quality 'lesion-healthy' counterfactual pairs, we generate diverse Modification task data by swapping L and H from niche of input and output and varying the textual prompts P.
+#### 3.2.3. TRANSFORMATION
+Transformation tasks include a wide array of low-level medical image processing operations. Unlike the localized edits found in Perception and Modification categories, tasks in this category typically require a holistic transformation of the entire input image.
+The rule-based construction pipeline of Transformation tasks is shown in Fig. [3.](#page-3-0) From public repositories, we compile medical image pairs (I and O) representing 17 distinct transformation targets under four typical low-level vision categories. We then design specialized textual prompts P for each task to unify diverse medical image processing functions into a consistent image editing framework.
+#### 3.2.4. POST PROCESSING
+Prompt rephrasing. To enhance linguistic diversity, we utilize the Qwen-Max model to rephrase the prompts P for each data triplet. We also illustrate the linguistic diversity of our prompts via a word cloud in Figure [2\(](#page-2-1)b).
+Benchmark curation. The training and test split of source dataset are strictly followed during the construction of MieDB-100k to preclude any data leakage. Furthermore, we recruit three people with clinical background to manually evaluate and curate 3, 485 of the most representative
+samples characterized by high clinical fidelity from raw data test split to serve as the benchmark of MieDB-100k, and we keep their original image size to minimize information loss.
+Train split construction. For train split, we establish three resolution bins (128, 256, and 512) and resize images to their nearest corresponding value. To check the fidelity of training split, we randomly select 6, 000 triplets for clinician evaluation, and over 95% are viewed as high quality.
+### 3.3. MieDB-100k Evaluation
+We evaluate MieDB-100k through two distinct approaches: (1) verifiable metrics for the Perception and Transformation tasks, amenable to reward design in prevailing reinforcement learning algorithms [\(Shao et al.,](#page-11-4) [2024;](#page-11-4) [Liu et al.,](#page-10-2) [2025a\)](#page-10-2); and (2) more subjective evaluations for the Modification tasks, reflecting their greater complexity.
+#### 3.3.1. VERIFIABLE EVALUATION
+Localization Accuracy Metric. We use the DICE Score for evaluating the spatial overlapping performance in Perception tasks. Notably, reconstructing a binary mask from the colored regions of an edited image is mathematically feasible when the background image and overlay color are known, and we detail this process in App. [D.1.](#page-15-0) This procedure is applied to both model's output O<sup>M</sup> and the ground truth images O to derive the mask of model's perceptual region and the ground truth region for DICE calculation.
+To differentiate between models that accurately identify specific medical targets and those that merely generate coarsegrained masks, we further propose Perception Accuracy. Under this metric, a result is considered successful only if the DICE score exceeds a threshold of τ = 0.8. This metric allows us to analyze whether a model possesses the specialized medical knowledge required for image understanding.
+Image Similarity Metrics. We utilize PSNR and SSIM [\(Wang et al.,](#page-11-5) [2004\)](#page-11-5) to evaluate the similarity between the ground-truth and edited images at both the pixel and structural levels. For evaluations within the Perception perspective, we mask out the pixels corresponding to the groundtruth segmentation in both images. This allows us to specifically assess the model's ability to preserve the background while performing the requested edit.
+### 3.3.2. EVALUATION FOR MODIFICATION TASKS
+Vision-Language Model Rubric Scoring. Automating reliable assessments in the Modification tasks is inherently challenging, as edits are defined semantically and cannot be evaluated via deterministic rules. Existing benchmarks often leverage Vision-Language Models (VLMs) for this purpose, and we standardize the process and mitigate potential critic hallucinations by implementing a rubric-based
+{5}------------------------------------------------
+<span id="page-5-0"></span>*Table 2.* Overall result on MieDB-100k benchmark. P-ACC means Perception Accuracy; B-PSNR and B-SSIM mean only calculate PSNR and SSIM on background pixels respectively; Rubric-S stands for the Rubric Score from VLM and Pref-Rank stands for human preference ranking. Best values are marked in red while second bests are in blue.
+| | | | | Perception | | Modification | | Trasnformation | | |
+|----------------------------------------|------|-------|-------|------------|--------|--------------|-----------|----------------|-------|--|
+| Model Name | Size | DICE | P-ACC | B-PSNR | B-SSIM | Rubric-S | Pref-Rank | PSNR | SSIM | |
+| Open-Source | | | | | | | | | | |
+| SDXL-turbo (Sauer et al., 2024) | 3.5B | 0.002 | 0.000 | 16.6 | 0.467 | 8.4 | 7.7 | 15.9 | 0.397 | |
+| Bagel (Deng et al., 2025) | 7B | 0.263 | 0.069 | 13.9 | 0.620 | 34.4 | 6.2 | 12.7 | 0.442 | |
+| OmniGen2 (Wu et al., 2025b) | 7B | 0.248 | 0.065 | 11.9 | 0.541 | 29.1 | 7.1 | 8.3 | 0.280 | |
+| Step1X-Edit (Liu et al., 2025c) | 21B | 0.332 | 0.126 | 15.5 | 0.727 | 35.6 | 4.5 | 16.6 | 0.539 | |
+| Qwen-Image-Edit (Wu et al., 2025a) | 27B | 0.387 | 0.153 | 15.4 | 0.722 | 32.2 | 5.5 | 18.9 | 0.606 | |
+| FLUX.1-Kontext-dev (Labs et al., 2025) | 12B | 0.341 | 0.126 | 15.4 | 0.701 | 37.8 | 6.2 | 17.9 | 0.543 | |
+| OmniGen2-MIE (Ours) | 7B | 0.831 | 0.737 | 28.1 | 0.917 | 65.9 | 1.4 | 22.6 | 0.685 | |
+| Closed-Source | | | | | | | | | | |
+| GPT-Image-1 (OpenAI, 2025) | | 0.467 | 0.221 | 16.3 | 0.510 | 42.8 | 4.8 | 14.4 | 0.451 | |
+| Nano Banana Pro (DeepMind, 2025a) | | 0.426 | 0.202 | 12.8 | 0.413 | 63.4 | 2.0 | 20.0 | 0.610 | |
+| Imagen4 (DeepMind, 2025b) | | 0.142 | 0.000 | 8.9 | 0.210 | 19.7 | 7.4 | 7.9 | 0.174 | |
+scoring system. Specifically, we provide the VLM with the input image I, edit instruction P, reference output O, and the model's generated result OM. Guided by the rubric, the VLM then performs a holistic evaluation of OM.
+We design a comprehensive scoring rubric (App. [D.2\)](#page-15-1) that assesses both the fulfillment of the editing intent and the model's ability to preserve the background. We utilize GPT-5.2 as an automated evaluator for this process, and map the final score to [0, 100].
+Human Preference Ranking. For each test case, we present the original triplet (I, P, O) and the outputs of all tested models simultaneously to evaluators, who are then asked to rank the various model-generated results according to their preference. By forcing this comparative ordering of all models, we are able to move beyond absolute quality scores and capture the relative strengths and weaknesses of current generative frameworks in a clinical setting. Specifically, we recruit 3 evaluators with clinical backgrounds to assess and rank the images edited by the benchmarked models, and compute the average ranking.
+# 4. Experiments
+# 4.1. Baselines
+We evaluate nine models on MieDB-100k, comprising six open-source models: Qwen-Image-Edit-2511 [\(Wu et al.,](#page-11-0) [2025a\)](#page-11-0), Bagel [\(Deng et al.,](#page-9-0) [2025\)](#page-9-0), OmniGen2 [\(Wu et al.,](#page-11-1) [2025b\)](#page-11-1), Step1X-Edit-v1p2 [\(Liu et al.,](#page-10-0) [2025c\)](#page-10-0), FLUX.1- Kontext-dev [\(Labs et al.,](#page-10-4) [2025\)](#page-10-4) and SDXL-turbo [\(Sauer](#page-10-3) [et al.,](#page-10-3) [2024\)](#page-10-3), plus three closed-source models: Nano Banana Pro [\(DeepMind,](#page-9-2) [2025a\)](#page-9-2), GPT-Image-1 [\(OpenAI,](#page-10-5) [2025\)](#page-10-5), and Imagen4 [\(DeepMind,](#page-9-4) [2025b\)](#page-9-4). We implement open-source models following their official inference settings.
+To validate the effectiveness of MieDB-100k, we finetune
+the OmniGen2 baseline on the training split and subject it to the same evaluation protocol as the other models. Specifically, we train the Diffusion Transformer (DiT) component for 20,000 iterations, employing a global batch size of 64 and a learning rate of 1e-4.
+## 4.2. Quantitative Results
+We report the benchmarking results of MieDB-100k in Tab. [2.](#page-5-0) First, the extremely low perception accuracy indicates that all tested models except ours fail to accurately comprehend and localize the specified anatomical targets under our evaluation protocol. Consequently, in Modification tasks, most of them are unable to generate clinically meaningful edits. Although a few models, such as Nano Banana Pro, achieve competitive results, we are indeed observing the 'right-for-the-wrong-reason' phenomenon, a risk that must be strictly avoided in clinical settings. Since the poor performance in Perception tasks expose their intrinsic lack of necessary medical knowledge, their edits cannot be justified. Notably in Transformation tasks, Nano Banana Pro also presents competitive results in certain cases. This may be attributed to the similarity between tasks like denoising or artifact removal and general-purpose low-level vision tasks, for which the model already possesses some capability [\(Zuo et al.,](#page-12-0) [2025\)](#page-12-0). Alternatively, it is possible that similar medical image processing tasks were included in its training set. Regardless, its absolute performance remains insufficient for practical clinical deployment. In summary, the benchmark result demonstrates that current multimodal generative model cannot meet the requirement of medical imaging editing.
+Conversely, after training on MieDB-100k, a standard baseline model can achieve superior medical editing capabilities. As shown in Tab. [2,](#page-5-0) the OmniGen2-MIE model delivers the best performance across all three edit-
+{6}------------------------------------------------
+<span id="page-6-0"></span>![](_page_6_Figure_1.jpeg)
+**[Edit Instruction]** Process the T1 MRI to mitigate motion effects without changing the appearance of brain tissue, tumors, or other clinical features.
+*Figure 4.* Qualitative editing result comparison.
+ing perspectives. The most significant improvements are observed in the Perception perspective, which demonstrate that MieDB-100k can effectively inject essential medical knowledge, thereby enhancing the interpretability of downstream editing tasks. Furthermore, in the Modification and Transformation tasks, where general-purpose editing abilities transfer more readily, our enhanced model still yields superior editing results compared to Nano Banana Pro, the SOTA multi-modal generative models. These findings highlight the pivotal role of our dataset in domain adaptation and
+establish a foundation for the development of understandinggeneration unified medical models.
+#### 4.3. Qualitative Results
+Fig. [4](#page-6-0) presents qualitative editing results for several baseline models across the diverse modalities and tasks in MieDB-100k. These results demonstrate that the finetuned model exhibits an enhanced capability in both understanding and generation, allowing it to navigate the inherent complexities
+{7}------------------------------------------------
+<span id="page-7-0"></span>Table 3. Ablation study result on MieDB-100k. P stands for Perception, M stands for Modification, and T stands for Transformation. Best values are marked in red, second bests are in blue.
+| | Perce | ption | Modification | Trasnformation | | | |
+|---------------------|----------|-------|--------------|----------------|-------|--|--|
+| Training Data | DICE ACC | | RubricScore | PSNR | SSIM | | |
+| Baseline (No train) | 0.248 | 0.065 | 29.1 | 8.3 | 0.280 | | |
+| P-only | 0.833 | 0.740 | 37.8 | 19.7 | 0.631 | | |
+| M-only | 0.001 | 0.000 | 57.5 | 19.8 | 0.631 | | |
+| T-only | 0.034 | 0.000 | 15.0 | 23.7 | 0.702 | | |
+| MieDB-100k | 0.831 | 0.737 | 65.9 | 22.6 | 0.685 | | |
+of medical image editing. Moreover, despite being explicitly prompted, even sophisticated closed-source models such as Nano Banana Pro fail to maintain background consistency in certain tasks. While their instruction-following proficiency stems from large-scale pre-training on natural image pairs, these capabilities tend to degrade when the distribution of medical modalities deviates significantly from the natural images seen during pre-training. To further study the impact of modality deviation, we conduct a modality-wise analysis in App. E.1, and the results prove our judgment. This observation underscores the necessity of a highly diverse dataset like MieDB-100k to equip models with the capacity to handle a vast range of medical imaging modalities.
+#### 4.4. Ablation Study
+To investigate the contribution of each task category, we conduct an ablation study by training models on individual perspective of MieDB-100k. We again utilize OmniGen2 as baseline model, following the training recipe described above while varying only the training data. As shown in Tab 3, each specialized model significantly outperforms the original baseline in its respective domain, validating the high information density and clinical relevance of our data. For the model trained on the full dataset, it achieves comparable or even better performance on all three perspectives, showing the effectiveness of the joint training. More importantly, we observe significant performance improvement in the Modification perspective, demonstrating visual understanding ability has the potential to enhance visual generation ability. In summary, the ablation study shows that MieDB-100k can provide a synergistic training signal, enabling the development of a versatile model capable of handling diverse medical editing tasks simultaneously.
+### <span id="page-7-2"></span>4.5. Generalization Test
+To further investigate the cross-task synergy and the resulting generalization capabilities, we conduct an out-ofdistribution (OOD) editing experiment. Specifically, we target 'bone metastasis', a medical target included in Perception tasks but strictly excluded from the Modification training data. We then prompt the OmniGen2-MIE model
+<span id="page-7-1"></span>![](_page_7_Figure_10.jpeg)
+Figure 5. Generalization test assessment. (a) and (b): Edit samples output by different models on bone metastasis addition (a) and removal (b) tasks. Red bounding boxes are added post-hoc to highlight the edited regions for visualization; (c): Quantitative assessments following the recipe of Modification task evaluation.
+to perform metastasis addition and removal in CT scans.
+As shown in Fig. 5, OmniGen2-MIE significantly outperforms OmniGen2 on this unseen task, demonstrating that our unified training on MieDB-100k can enhance the model's generalization capabilities across editing tasks. We also observe that Nano Banana Pro achieves the best OOD editing performance, marginally surpassing OmniGen2-MIE. We attribute this performance to the utilization of massive-scale general and medical editing data, which further underscores the necessity of scaling up medical editing data.
+#### 5. Conclusion
+In this paper, we introduce MieDB-100k, a large-scale and diverse dataset for text-guided medical image editing. By unifying Perception, Modification, and Transformation tasks into the paradigm of editing, our dataset bridges the gap between medical image understanding and generation. We develop a robust curation pipeline, integrating modalityspecific expert models with rule-based synthesis, and enforce rigorous manual quality control to ensure clinical fidelity across all data. Extensive benchmarking demonstrates that model trained on MieDB-100k consistently outperform both SOTA open-source and proprietary multimodal models while exhibiting exceptional generalization to unseen clinical tasks. Our work thus provides the data foundation to support the development and evaluation of multimodal generative models for clinical applications.
+{8}------------------------------------------------
+# Impact Statement
+This paper presents work whose goal is to advance the field of Machine Learning. There are many potential societal consequences of our work, none which we feel must be specifically highlighted here.
+# References
+- <span id="page-8-9"></span>AAPM. Low dose ct grand challenge, 2016. URL [https:]( [// [ect/#trainingData]( Accessed: 2026-01-14.
+- <span id="page-8-14"></span>Abdi, A. H., Kasaei, S., and Mehdizadeh, M. Automatic segmentation of mandible in panoramic x-ray. *Journal of Medical Imaging*, 2(4):044003–044003, 2015.
+- <span id="page-8-10"></span>Ahmed, Z., Panhwar, S. Q., Baqai, A., Umrani, F. A., Ahmed, M., and Khan, A. Deep learning based automated detection of intraretinal cystoid fluid. *International Journal of Imaging Systems and Technology*, 32(3):902–917, 2022.
+- <span id="page-8-4"></span>Al-Dhabyani, W., Gomaa, M., Khaled, H., and Fahmy, A. Dataset of breast ultrasound images. *Data in brief*, 28: 104863, 2020.
+- <span id="page-8-15"></span>Ali, S., Ghatwary, N., Jha, D., Isik-Polat, E., Polat, G., Yang, C., Li, W., Galdran, A., Ballester, M.-A. G., Thambawita, ´ V., et al. Assessing generalisability of deep learningbased polyp detection and segmentation methods through a computer vision challenge. *Scientific Reports*, 14(1): 2032, 2024.
+- <span id="page-8-7"></span>Angiographics. Chuac dataset. URL [ [re.com/s/4d24cf3d14bc901a94bf]( Accessed: 2026-01-14.
+- <span id="page-8-13"></span>Antonelli, M., Reinke, A., Bakas, S., Farahani, K., Kopp-Schneider, A., Landman, B. A., Litjens, G., Menze, B., Ronneberger, O., Summers, R. M., et al. The medical segmentation decathlon. *Nature communications*, 13(1): 4128, 2022.
+- <span id="page-8-16"></span>Aysen, D., Muhammad, C., Tawsifur, R., Amith, K., Yazan, Q., Mete, A., Anas M., T., and serkan, k. Qata-cov19 dataset, 2024. URL [ [atasets/aysendegerli/qatacov19-datas]( [et]( Accessed: 2026-01-14.
+- <span id="page-8-3"></span>Bai, S., Cai, Y., Chen, R., Chen, K., Chen, X., Cheng, Z., Deng, L., Ding, W., Gao, C., Ge, C., Ge, W., Guo, Z., Huang, Q., Huang, J., Huang, F., Hui, B., Jiang, S., Li, Z., Li, M., Li, M., Li, K., Lin, Z., Lin, J., Liu, X., Liu, J., Liu, C., Liu, Y., Liu, D., Liu, S., Lu, D., Luo, R., Lv, C., Men, R., Meng, L., Ren, X., Ren, X., Song, S., Sun, Y., Tang, J., Tu, J., Wan, J., Wang, P., Wang, P.,
+- Wang, Q., Wang, Y., Xie, T., Xu, Y., Xu, H., Xu, J., Yang, Z., Yang, M., Yang, J., Yang, A., Yu, B., Zhang, F., Zhang, H., Zhang, X., Zheng, B., Zhong, H., Zhou, J., Zhou, F., Zhou, J., Zhu, Y., and Zhu, K. Qwen3-vl technical report. *Preprint at arXiv*, 2025a. URL [https:]( [//arxiv.org/abs/2511.21631](
+- <span id="page-8-2"></span>Bai, S., Chen, K., Liu, X., Wang, J., Ge, W., Song, S., Dang, K., Wang, P., Wang, S., Tang, J., Zhong, H., Zhu, Y., Yang, M., Li, Z., Wan, J., Wang, P., Ding, W., Fu, Z., Xu, Y., Ye, J., Zhang, X., Xie, T., Cheng, Z., Zhang, H., Yang, Z., Xu, H., and Lin, J. Qwen2.5-vl technical report. *Preprint at arXiv*, 2025b. URL [ [g/abs/2502.13923](
+- <span id="page-8-1"></span>Brooks, T., Holynski, A., and Efros, A. A. Instructpix2pix: Learning to follow image editing instructions. In *Proceedings of the IEEE/CVF conference on computer vision and pattern recognition*, pp. 18392–18402, 2023.
+- <span id="page-8-12"></span>Buda, M., Saha, A., and Mazurowski, M. A. Association of genomic subtypes of lower-grade gliomas with shape features automatically extracted by a deep learning algorithm. *Computers in biology and medicine*, 109:218–225, 2019.
+- <span id="page-8-5"></span>Caicedo, J. C., Goodman, A., Karhohs, K. W., Cimini, B. A., Ackerman, J., Haghighi, M., Heng, C., Becker, T., Doan, M., McQuin, C., et al. Nucleus segmentation across imaging experiments: the 2018 data science bowl. *Nature methods*, 16(12):1247–1253, 2019.
+- <span id="page-8-6"></span>Carballal, A., Novoa, F. J., Fernandez-Lozano, C., Garc´ıa-Guimaraes, M., Aldama-Lopez, G., Calvi ´ no-Santos, R., ˜ Vazquez-Rodriguez, J. M., and Pazos, A. Automatic multiscale vascular image segmentation algorithm for coronary angiography. *Biomedical Signal Processing and Control*, 46:1–9, 2018.
+- <span id="page-8-11"></span>Cervantes-Sanchez, F., Cruz-Aceves, I., Hernandez-Aguirre, A., Hernandez-Gonzalez, M. A., and Solorio-Meza, S. E. Automatic segmentation of coronary arteries in x-ray angiograms using multiscale analysis and artificial neural networks. *Applied Sciences*, 9(24):5507, 2019.
+- <span id="page-8-0"></span>Chen, Z. and Feng, M. Med-banana-50k: A cross-modality large-scale dataset for text-guided medical image editing. *Preprint at arXiv*, 2025. URL [ [g/abs/2511.00801](
+- <span id="page-8-8"></span>Chowdhury, M. E., Rahman, T., Khandakar, A., Mazhar, R., Kadir, M. A., Mahbub, Z. B., Islam, K. R., Khan, M. S., Iqbal, A., Al Emadi, N., et al. Can ai help in screening viral and covid-19 pneumonia? *Ieee Access*, 8: 132665–132676, 2020.
+{9}------------------------------------------------
+<span id="page-9-8"></span>495 496 497 498 499 500 Codella, N., Rotemberg, V., Tschandl, P., Celebi, M. E., Dusza, S., Gutman, D., Helba, B., Kalloo, A., Liopyris, K., Marchetti, M., et al. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic). *arXiv preprint arXiv:1902.03368*, 2019.
+<span id="page-9-13"></span>501
+<span id="page-9-4"></span>536
+- Dao, B. T., Nguyen, T. V., Pham, H. H., and Nguyen, H. Q. Phase recognition in contrast-enhanced ct scans based on deep learning and random sampling. *Medical Physics*, 49 (7):4518–4528, 2022.
+- <span id="page-9-5"></span>de Verdier, M. C., Saluja, R., Gagnon, L., LaBella, D., Baid, U., Tahon, N. H., Foltyn-Dumitru, M., Zhang, J., Alafif, M., Baig, S., Chang, K., D'Anna, G., Deptula, L., Gupta, D., Haider, M. A., Hussain, A., Iv, M., Kontzialis, M., Manning, P., Moodi, F., Nunes, T., Simon, A., Sollmann, N., Vu, D., Adewole, M., Albrecht, J., Anazodo, U., Chai, R., Chung, V., Faghani, S., Farahani, K., Kazerooni, A. F., Iglesias, E., Kofler, F., Li, H., Linguraru, M. G., Menze, B., Moawad, A. W., Velichko, Y., Wiestler, B., Altes, T., Basavasagar, P., Bendszus, M., Brugnara, G., Cho, J., Dhemesh, Y., Fields, B. K. K., Garrett, F., Gass, J., Hadjiiski, L., Hattangadi-Gluth, J., Hess, C., Houk, J. L., Isufi, E., Layfield, L. J., Mastorakos, G., Mongan, J., Nedelec, P., Nguyen, U., Oliva, S., Pease, M. W., Rastogi, A., Sinclair, J., Smith, R. X., Sugrue, L. P., Thacker, J., Vidic, I., Villanueva-Meyer, J., White, N. S., Aboian, M., Conte, G. M., Dale, A., Sabuncu, M. R., Seibert, T. M., Weinberg, B., Abayazeed, A., Huang, R., Turk, S., Rauschecker, A. M., Farid, N., Vollmuth, P., Nada, A., Bakas, S., Calabrese, E., and Rudie, J. D. The 2024 brain tumor segmentation (brats) challenge: Glioma segmentation on post-treatment mri. *Preprint at arXiv*, 2024. URL [ [8368](
+- <span id="page-9-2"></span>DeepMind, G. Gemini 3 pro: High-precision multimodal reasoning, 2025a. URL [ [gle/models/gemini-image/pro/]( Accessed: 2026-01-14.
+- DeepMind, G. Imagen 4.0: Model documentation and generation guide, 2025b. URL [ [loud.google.com/vertex-ai/generative]( [-ai/docs/models/imagen/4-0-generate]( Accessed: 2026-01-14.
+- <span id="page-9-0"></span>Deng, C., Zhu, D., Li, K., Gou, C., Li, F., Wang, Z., Zhong, S., Yu, W., Nie, X., Song, Z., Shi, G., and Fan, H. Emerging properties in unified multimodal pretraining. *Preprint at arXiv*, 2025. URL [https:]( [//arxiv.org/abs/2505.14683](
+- <span id="page-9-14"></span>Dietler, N., Minder, M., Gligorovski, V., Economou, A. M., Joly, D. A. H. L., Sadeghi, A., Chan, C. H. M., Kozinski, ´
+- M., Weigert, M., Bitbol, A.-F., et al. A convolutional neural network segments yeast microscopy images with high accuracy. *Nature communications*, 11(1):5723, 2020.
+- <span id="page-9-12"></span>Duong, V. H., Vu, H., Phan, H. D., Nguyen, D. Q., Pham, D. H., Le, Q. T., Nguyen, B. S., Do, T. D., Dinh, V. S., Nguyen, T. C., et al. Thyroidxl: Advancing thyroid nodule diagnosis with an expert-labeled, pathology-validated dataset. In *International Conference on Medical Image Computing and Computer-Assisted Intervention*, pp. 616– 626. Springer, 2025.
+- <span id="page-9-10"></span>Fang, H., Li, F., Wu, J., Fu, H., Sun, X., Son, J., Yu, S., Zhang, M., Yuan, C., Bian, C., Lei, B., Zhao, B., Xu, X., Li, S., Fumero, F., Sigut, J., Almubarak, H., Bazi, Y., Guo, Y., Zhou, Y., Baid, U., Innani, S., Guo, T., Yang, J., Orlando, J. I., Bogunovic, H., Zhang, X., and Xu, Y. ´ Refuge2 challenge: A treasure trove for multi-dimension analysis and evaluation in glaucoma screening. *Preprint at arXiv*, 2022. URL [ [2202.08994](
+- <span id="page-9-11"></span>Garcia-Peraza-Herrera, L. C., Fidon, L., D'Ettorre, C., Stoyanov, D., Vercauteren, T., and Ourselin, S. Image compositing for segmentation of surgical tools without manual annotations. *IEEE transactions on medical imaging*, 40(5):1450–1460, 2021.
+- <span id="page-9-7"></span>Gutman, D., Codella, N. C., Celebi, E., Helba, B., Marchetti, M., Mishra, N., and Halpern, A. Skin lesion analysis toward melanoma detection: A challenge at the international symposium on biomedical imaging (isbi) 2016, hosted by the international skin imaging collaboration (isic). *arXiv preprint arXiv:1605.01397*, 2016.
+- <span id="page-9-6"></span>Haneda, E., Peters, N., Zhang, J., Karageorgos, G., Xia, W., Paganetti, H., Wang, G., Guo, Y., Ma, J., Park, H. S., et al. Aapm ct metal artifact reduction grand challenge. *Medical physics*, 52(10):e70050, 2025.
+- <span id="page-9-1"></span>Huang, Z., Zheng, D., Zou, C., Liu, R., Wang, X., Ji, K., Chai, W., Sun, J., Wang, L., Lv, Y., Huang, T., Liu, J., Guo, Q., Yang, M., Chen, J., and Zhou, J. Ming-univision: Joint image understanding and generation with a unified continuous tokenizer. *Preprint at arXiv*, 2025. URL <
+- <span id="page-9-3"></span>Isensee, F., Jaeger, P. F., Kohl, S. A., Petersen, J., and Maier-Hein, K. H. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. *Nature methods*, 18(2):203–211, 2021.
+- <span id="page-9-9"></span>Jha, D., Smedsrud, P. H., Riegler, M. A., Halvorsen, P., De Lange, T., Johansen, D., and Johansen, H. D. Kvasirseg: A segmented polyp dataset. In *International conference on multimedia modeling*, pp. 451–462. Springer, 2019.
+{10}------------------------------------------------
+<span id="page-10-20"></span>550 Kus¸, Z. and Aydin, M. Medsegbench: A comprehensive benchmark for medical image segmentation in diverse data modalities. *Scientific Data*, 11(1):1283, 2024.
+<span id="page-10-4"></span>554
+<span id="page-10-2"></span>574
+<span id="page-10-6"></span>594
+- Labs, B. F., Batifol, S., Blattmann, A., Boesel, F., Consul, S., Diagne, C., Dockhorn, T., English, J., English, Z., Esser, P., Kulal, S., Lacey, K., Levi, Y., Li, C., Lorenz, D., Muller, J., Podell, D., Rombach, R., Saini, H., Sauer, ¨ A., and Smith, L. Flux.1 kontext: Flow matching for in-context image generation and editing in latent space. *Preprint at arXiv*, 2025. URL [ [g/abs/2506.15742](
+- <span id="page-10-15"></span>Lei, W., Chen, H., Zhang, Z., Luo, L., Xiao, Q., Gu, Y., Gao, P., Jiang, Y., Wang, C., Wu, G., Xu, T., Zhang, Y., Rajpurkar, P., Zhang, X., Zhang, S., and Wang, Z. A synthetic data-driven radiology foundation model for pantumor clinical diagnosis. *Preprint at arXiv*, 2025. URL <
+- <span id="page-10-16"></span>Litjens, G., Toth, R., Van De Ven, W., Hoeks, C., Kerkstra, S., Van Ginneken, B., Vincent, G., Guillard, G., Birbeck, N., Zhang, J., et al. Evaluation of prostate segmentation algorithms for mri: the promise12 challenge. *Medical image analysis*, 18(2):359–373, 2014.
+- Liu, J., Liu, G., Liang, J., Li, Y., Liu, J., Wang, X., Wan, P., Zhang, D., and Ouyang, W. Flow-grpo: Training flow matching models via online rl. *arXiv preprint at arXiv*, 2025a. URL [ [5470](
+- <span id="page-10-1"></span>Liu, M., He, Z., Fan, Z., Wang, Q., and Fung, Y. R. Medebench: Diagnosing reliability in text-guided medical image editing. In *Findings of the Association for Computational Linguistics: EMNLP 2025*, pp. 767–791, 2025b.
+- <span id="page-10-0"></span>Liu, S., Han, Y., Xing, P., Yin, F., Wang, R., Cheng, W., Liao, J., Wang, Y., Fu, H., Han, C., Li, G., Peng, Y., Sun, Q., Wu, J., Cai, Y., Ge, Z., Ming, R., Xia, L., Zeng, X., Zhu, Y., Jiao, B., Zhang, X., Yu, G., and Jiang, D. Step1xedit: A practical framework for general image editing. *Preprint at arXiv*, 2025c. URL [ [g/abs/2504.17761](
+- Ljosa, V., Sokolnicki, K. L., and Carpenter, A. E. Annotated high-throughput microscopy image sets for validation. *Nature methods*, 9(7):637, 2012.
+- <span id="page-10-12"></span>Lu, Y., Zhou, M., Zhi, D., Zhou, M., Jiang, X., Qiu, R., Ou, Z., Wang, H., Qiu, D., Zhong, M., et al. The jnu-ifm dataset for segmenting pubic symphysis-fetal head. *Data in brief*, 41:107904, 2022.
+- <span id="page-10-8"></span>Mathieu, G., Bachir, E. D., et al. Brifiseg: a deep learningbased method for semantic and instance segmentation
+- of nuclei in brightfield images. *Preprint at arXiv*, 2022. URL <
+- <span id="page-10-18"></span>Montoya, A. Ultrasound nerve segmentation, 2026. URL [ [ultrasound-nerve-segmentation]( Accessed: 2026-01-14.
+- <span id="page-10-11"></span>Morozov, S. P., Andreychenko, A. E., Pavlov, N. A., Vladzymyrskyy, A., Ledikhova, N. V., Gombolevskiy, V. A., Blokhin, I. A., Gelezhe, P. B., Gonchar, A., and Chernina, V. Y. Mosmeddata: Chest ct scans with covid-19 related findings dataset. *arXiv preprint arXiv:2005.06465*, 2020.
+- <span id="page-10-14"></span>Narai, ´ A., Hermann, P., Auer, T., Kemenczky, P., Szalma, ´ J., Homolya, I., Somogyi, E., Vakli, P., Weiss, B., and Vidnyanszky, Z. Movement-related artefacts (mr-art) ´ dataset of matched motion-corrupted and clean structural mri brain scans. *Scientific data*, 9(1):630, 2022.
+- <span id="page-10-17"></span>Naylor, P., Lae, M., Reyal, F., and Walter, T. Segmentation ´ of nuclei in histopathology images by deep regression of the distance map. *IEEE transactions on medical imaging*, 38(2):448–459, 2018.
+- <span id="page-10-7"></span>Ngoc Lan, P., An, N. S., Hang, D. V., Long, D. V., Trung, T. Q., Thuy, N. T., and Sang, D. V. Neounet: Towards accurate colon polyp segmentation and neoplasm detection. In *International Symposium on Visual Computing*, pp. 15–28. Springer, 2021.
+- <span id="page-10-5"></span>OpenAI. Gpt image 1 model documentation, 2025. URL [ [els/gpt-image-1]( Accessed: 2026-01-14.
+- <span id="page-10-10"></span>Pandey, N. Chest xray masks and labels, 2025. URL [ht]( [tps:// [pandey360/chest-xray-masks-and-labels]( Accessed: 2026-01-14.
+- <span id="page-10-19"></span>pazhoulab. Low-dose ct reconstruction contest, 2024. URL [ [ontestdetail?id=667e0c687ff47da8cc82]( [7679&award=1,000,000]( Accessed: 2026-01-14.
+- <span id="page-10-9"></span>Polo, M. Chest ct segmentation, 2025. URL [https:]( [// [chest-ct-segmentation]( Accessed: 2026-01-14.
+- <span id="page-10-13"></span>Porwal, P., Pachade, S., Kamble, R., Kokare, M., Deshmukh, G., Sahasrabuddhe, V., and Meriaudeau, F. Indian diabetic retinopathy image dataset (idrid): a database for diabetic retinopathy screening research. *Data*, 3(3):25, 2018.
+- <span id="page-10-3"></span>Sauer, A., Lorenz, D., Blattmann, A., and Rombach, R. Adversarial diffusion distillation. In *European Conference on Computer Vision*, pp. 87–103. Springer, 2024.
+{11}------------------------------------------------
+<span id="page-11-4"></span>Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., Bi, X., Zhang, H., Zhang, M., Li, Y. K., Wu, Y., and Guo, D. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. *Preprint at arXiv*, 2024. URL <
+- <span id="page-11-16"></span>Song, Y., Zheng, J., Lei, L., Ni, Z., Zhao, B., and Hu, Y. Ct2us: Cross-modal transfer learning for kidney segmentation in ultrasound images with synthesized data. *Ultrasonics*, 122:106706, 2022.
+- <span id="page-11-8"></span>Spahn, C., Gomez-de Mariscal, E., Laine, R. F., Pereira, ´ P. M., von Chamier, L., Conduit, M., Pinho, M. G., Jacquemet, G., Holden, S., Heilemann, M., et al. Deepbacs for multi-task bacterial image analysis using opensource deep learning approaches. *Communications Biology*, 5(1):688, 2022.
+- <span id="page-11-9"></span>Staal, J., Abramoff, M. D., Niemeijer, M., Viergever, M. A., ` and Van Ginneken, B. Ridge-based vessel segmentation in color images of the retina. *IEEE transactions on medical imaging*, 23(4):501–509, 2004.
+- <span id="page-11-7"></span>Tahir, A. M., Chowdhury, M. E., Khandakar, A., Rahman, T., Qiblawey, Y., Khurshid, U., Kiranyaz, S., Ibtehaz, N., Rahman, M. S., Al-Maadeed, S., et al. Covid-19 infection localization and severity grading from chest x-ray images. *Computers in biology and medicine*, 139:105002, 2021.
+- <span id="page-11-2"></span>Tong, S., Fan, D., Li, J., Xiong, Y., Chen, X., Sinha, K., Rabbat, M., LeCun, Y., Xie, S., and Liu, Z. Metamorph: Multimodal understanding and generation via instruction tuning. In *Proceedings of the IEEE/CVF International Conference on Computer Vision*, pp. 17001–17012, 2025.
+- <span id="page-11-10"></span>Van Valen, D. A., Kudo, T., Lane, K. M., Macklin, D. N., Quach, N. T., DeFelice, M. M., Maayan, I., Tanouchi, Y., Ashley, E. A., and Covert, M. W. Deep learning automates the quantitative analysis of individual cells in live-cell imaging experiments. *PLoS computational biology*, 12(11):e1005177, 2016.
+- <span id="page-11-12"></span>Verma, R., Kumar, N., Patil, A., Kurian, N. C., Rane, S., Graham, S., Vu, Q. D., Zwager, M., Raza, S. E. A., Rajpoot, N., et al. Monusac2020: A multi-organ nuclei segmentation and classification challenge. *IEEE Transactions on Medical Imaging*, 40(12):3413–3423, 2021.
+- <span id="page-11-17"></span>Vision and Lab, I. P. Skin cancer detection, 2024. URL [ [r-detection/]( Accessed: 2026-01-14.
+- <span id="page-11-6"></span>Vitale, S., Orlando, J. I., Iarussi, E., and Larrabide, I. Improving realism in patient-specific abdominal ultrasound simulation using cyclegans. *International journal of computer assisted radiology and surgery*, 15(2):183–192, 2020.
+- <span id="page-11-11"></span>Wang, Y. and Shi, F. KMAR-50K. Mendeley Data, V6, 2025. URL [ [mrg7ntg.6]( Accessed: 2026-01-14.
+- <span id="page-11-5"></span>Wang, Z., Bovik, A. C., Sheikh, H. R., and Simoncelli, E. P. Image quality assessment: from error visibility to structural similarity. *IEEE transactions on image processing*, 13(4):600–612, 2004.
+- <span id="page-11-15"></span>Wasserthal, J., Breit, H.-C., Meyer, M. T., Pradella, M., Hinck, D., Sauter, A. W., Heye, T., Boll, D. T., Cyriac, J., Yang, S., et al. Totalsegmentator: robust segmentation of 104 anatomic structures in ct images. *Radiology: Artificial Intelligence*, 5(5):e230024, 2023.
+- <span id="page-11-0"></span>Wu, C., Li, J., Zhou, J., Lin, J., Gao, K., Yan, K., ming Yin, S., Bai, S., Xu, X., Chen, Y., Chen, Y., Tang, Z., Zhang, Z., Wang, Z., Yang, A., Yu, B., Cheng, C., Liu, D., Li, D., Zhang, H., Meng, H., Wei, H., Ni, J., Chen, K., Cao, K., Peng, L., Qu, L., Wu, M., Wang, P., Yu, S., Wen, T., Feng, W., Xu, X., Wang, Y., Zhang, Y., Zhu, Y., Wu, Y., Cai, Y., and Liu, Z. Qwen-image technical report. *Preprint at arXiv*, 2025a. URL [ [08.02324](
+- <span id="page-11-1"></span>Wu, C., Zheng, P., Yan, R., Xiao, S., Luo, X., Wang, Y., Li, W., Jiang, X., Liu, Y., Zhou, J., et al. Omnigen2: Exploration to advanced multimodal generation. *Preprint at arXiv*, 2025b. URL [ [2506.18871](
+- <span id="page-11-3"></span>Yang, J., Yan, Y., Wu, G., Wang, Y., Liang, R., Jiang, X., Wan, X., Fan, F., Zhang, Y., Qin, F., and Wang, C. Medgen-bench: Contextually entangled benchmark for open-ended multimodal medical generation. *Preprint at arXiv*, 2025. URL [ [11.13135](
+- <span id="page-11-13"></span>Yang, L., Ghosh, R. P., Franklin, J. M., Chen, S., You, C., Narayan, R. R., Melcher, M. L., and Liphardt, J. T. Nuset: A deep learning tool for reliably separating and analyzing crowded cells. *PLoS computational biology*, 16(9):e1008193, 2020.
+- <span id="page-11-14"></span>Younus Akon, Z. S. S. Paired ct and mri dataset for medical applications, 2025. URL [ [om/datasets/29c3607295965ebb030f2d15]( [8fec487412d84c82528dd44f8ef956aef355]( [41aa]( Accessed: 2026-01-14.
+- <span id="page-11-18"></span>Zhang, S., Zhang, Q., Zhang, S., Liu, X., Yue, J., Lu, M., Xu, H., Yao, J., Wei, X., Cao, J., et al. A generalist foundation model and database for open-world medical image segmentation. *Nature Biomedical Engineering*, pp. 1–16, 2025.
+{12}------------------------------------------------
+<span id="page-12-1"></span>Zheng, X., Wang, Y., Wang, G., and Liu, J. Fast and robust segmentation of white blood cell images by selfsupervised learning. *Micron*, 107:55–71, 2018.
+<span id="page-12-0"></span>Zuo, J., Deng, H., Zhou, H., Zhu, J., Zhang, Y., Zhang, Y., Yan, Y., Huang, K., Chen, W., Deng, Y., Jin, R., Sang, N., and Gao, C. Is nano banana pro a low-level vision all-rounder? a comprehensive evaluation on 14 tasks and 40 datasets. *Preprint at arXiv*, 2025. URL <
+{13}------------------------------------------------
+# <span id="page-13-0"></span>A. Data Sources
+Our work is compiled based on following public medical image repositories:
+*Table 4.* Summary of public medical datasets utilized in the construction of MieDB-100k. The columns #Train and #Benchmark denote the number of samples allocated to our training and benchmark splits respectively from each source dataset.
+| DatasetName | #Train | #Benchmark | Modality |
+|-------------------------------------------------------|--------|------------|----------------|
+| AbdomenUS (Vitale et al., 2020) | 569 | 62 | Ultrasound |
+| Bbbc010 (Ljosa et al., 2012) | 70 | 20 | Microscopy |
+| Bkai-Igh (Ngoc Lan et al., 2021) | 700 | 81 | Endoscopy |
+| Brats-gli (de Verdier et al., 2024) | 1529 | 80 | MRI |
+| BriFiSeg (Mathieu et al., 2022) | 1005 | 40 | Microscopy |
+| BUSI (Al-Dhabyani et al., 2020) | 452 | 80 | Ultrasound |
+| CellNuclei (Caicedo et al., 2019) | 469 | 51 | Microscopy |
+| ChaseDB1 (Carballal et al., 2018) | 19 | 7 | Fundus |
+| Chest-ct-segmentation (Polo, 2025) | 278 | 19 | CT |
+| Chest-xray-masks-and-labels (Pandey, 2025) | 666 | 32 | Xray |
+| CHUAC (Angiographics) | 17 | 5 | Fundus |
+| COVID-19 Radiography Dataset (Chowdhury et al., 2020) | 2010 | 95 | Xray |
+| COVID-19-CT-SCAN-Lesion (Morozov et al., 2020) | 255 | 15 | CT |
+| CovidQU (Tahir et al., 2021) | 5684 | 122 | Xray |
+| CT MAR (Haneda et al., 2025) | 1595 | 82 | CT |
+| CT-Low-Dose-Reconstruction (AAPM, 2016) | 867 | 51 | CT |
+| CystoidFluid (Ahmed et al., 2022) | 703 | 59 | OCT |
+| Dca1 (Cervantes-Sanchez et al., 2019) | 93 | 28 | Fundus |
+| Deepbacs (Spahn et al., 2022) | 17 | 10 | Microscopy |
+| Drive (Staal et al., 2004) | 18 | 20 | Fundus |
+| DynamicNuclear (Van Valen et al., 2016) | 50 | 17 | Microscopy |
+| FHPsAOP (Lu et al., 2022) | 2800 | 80 | Ultrasound |
+| IDRiD (Porwal et al., 2018) | 47 | 27 | Fundus |
+| ISIC2016 (Gutman et al., 2016) | 810 | 80 | Dermoscopy |
+| ISIC2018 (Codella et al., 2019) | 9973 | 115 | Dermoscopy |
+| KMAR-50K (Wang & Shi, 2025) | 651 | 47 | MRI |
+| Kvasir (Jha et al., 2019) | 4429 | 139 | Endoscopy |
+| Lgg-mri-segmentation (Buda et al., 2019) | 1669 | 55 | MRI |
+| MoNuSAC (Verma et al., 2021) | 0 | 21 | Microscopy |
+| MR-ART (Narai et al. ´ , 2022) | 820 | 18 | MRI |
+| MSD (Antonelli et al., 2022) | 797 | 3912 | MRI |
+| NuSeT (Yang et al., 2020) | 2383 | 40 | Microscopy |
+| Paried MRI CT (Younus Akon, 2025) | 1974 | 72 | CT, MRI |
+| Pandental (Abdi et al., 2015) | 81 | 24 | Xray |
+| Pasta-GEN (Lei et al., 2025) | 32299 | 731 | CT |
+| PolypGen (Ali et al., 2024) | 984 | 75 | Endoscopy |
+| PROMISE12 (Litjens et al., 2014) | 1031 | 80 | MRI |
+| QaTa-COV19 (Aysen et al., 2024) | 3573 | 85 | Xray |
+| Refuge (Fang et al., 2022) | 80 | 80 | Fundus |
+| RoboTool (Garcia-Peraza-Herrera et al., 2021) | 350 | 76 | Surgical Photo |
+| ThyroidXL (Duong et al., 2025) | 7029 | 138 | Ultrasound |
+| Tnbcnuclei (Naylor et al., 2018) | 35 | 10 | Microscopy |
+| TotalSegmentator (Wasserthal et al., 2023) | 5206 | 154 | CT, MRI |
+| UltrasoundNerve (Montoya, 2026) | 1651 | 50 | Ultrasound |
+| USforKidney (Song et al., 2022) | 4351 | 50 | Ultrasound |
+| UWSkinCancer (Vision & Lab, 2024) | 143 | 44 | Dermoscopy |
+| VinDr-Multiphase (Dao et al., 2022) | 3486 | 44 | CT |
+| WBC (Zheng et al., 2018) | 280 | 40 | Microscopy |
+| YeaZ (Dietler et al., 2020) | 358 | 51 | Microscopy |
+| YGA low dose ct (pazhoulab, 2024) | 4387 | 44 | CT |
+We also appreciate MedSegBench(Kus¸ [& Aydin,](#page-10-20) [2024\)](#page-10-20) and MedSegDB[\(Zhang et al.,](#page-11-18) [2025\)](#page-11-18) for collecting and pre-processing some of these datasets.
+{14}------------------------------------------------
+# B. Construction Details
+![](_page_14_Figure_2.jpeg)
+*Figure 6.* Construction details of three perspective. We manually curate the benchmark split to uphold high clinical standards. The remaining training data is validated through sampling-based quality checks, establishing a high-quality data proportion exceeding 95%.
+# C. Implementation Details of OmniGen2-MIE
+| Hyper-Parameter | Value |
+|-------------------------------|---------------------------|
+| Finetuning method | Full-Parameter Finetuning |
+| snr type | lognorm |
+| do shift | True |
+| dynamic time shift | True |
+| Steps | 20, 000 |
+| #GPUs | 8 |
+| Per-device batch size | 8 |
+| Gradient accumulation | 1 |
+| Global batch size (effective) | 64 |
+| Learning rate | 1 × 10−4 |
+| LR scheduler | timm constant with warmup |
+| Warm-up t | 500 |
+| Precision | BF16 |
+| Random seed | 2233 |
+*Table 5.* Training hyper-parameters used for finetuning OmniGen2-MIE on our dataset.
+{15}------------------------------------------------
+# <span id="page-15-0"></span>D. Evaluation Details
+#### D.1. Mask Reconstruction via Alpha De-blending
+### D.1.1. MATHEMATICS
+To recover the segmentation mask from the visualized output, we model the edited image O as a linear interpolation between the original background image B (a.k.a. the input image I) and a known overlay color C (red, green or blue). This relationship is governed by the per-pixel alpha channel α ∈ [0, 1], according to the standard alpha blending equation:
+$$\mathbf{O} = (1 - \alpha)\mathbf{B} + \alpha\mathbf{C} \tag{1}$$
+By rearranging the terms as O − B = α(C − B), the scalar value α can be interpreted as the projection of the observed color shift onto the vector representing the maximum possible color change. To account for potential noise in the RGB space, we solve for α at each pixel using the least-squares solution:
+$$\alpha = \frac{(\mathbf{O} - \mathbf{B}) \cdot (\mathbf{C} - \mathbf{B})}{|\mathbf{C} - \mathbf{B}|^2}$$
+(2)
+The continuous alpha map is subsequently binarized to produce the final segmentation mask M. This is achieved by applying a global threshold τ , such that:
+$$M_{i,j} = \begin{cases} 1 & \text{if } \alpha_{i,j} > \tau \\ 0 & \text{otherwise} \end{cases}$$
+(3)
+In our implementation, a threshold of τ = 0.5 is utilized to effectively separate the predicted regions from the background.
+#### D.1.2. CASE OF MASK RECONSTRUCTION
+![](_page_15_Picture_12.jpeg)
+*Figure 7.* Case of perception mask reconstruction.
+#### <span id="page-15-1"></span>D.2. VLM Automatic Scoring
+#### D.2.1. VLM SCORING RUBRIC
+# Scoring Rubric for Modification Tasks
+You are a helpful assistant in evaluating medical image editing result.
+You will be provided with an edit instruction and a collage image where the leftmost is origin image, center is edited image and rightmost is the reference ground truth image.
+You should score how well an edited image matches the intended edit while preserving clinical realism and image integrity based on following scoring rubrics:
+#### 1) Edit Goal Fulfillment (Edit Correctness): Assesses whether the intended lesion change is achieved.
+Scoring reference:
+- *5: Lesion added/removed exactly as intended; no residuals or unintended remnants.*
+- *4: Mostly correct; slight residual signal after removal or slight under/over-addition.*
+- *3: Partial success; lesion still partially present (removal) or incomplete/incorrect lesion (addition).*
+{16}------------------------------------------------
+> 932 933 934
+- *2: Wrong area or wrong type of change; target lesion largely unchanged.*
+- *1: No effective edit or opposite edit performed.*
+- 2) Edit Area Morphology (Shape, Margins, Internal Structure): Evaluates whether edit area matches expected morphology and/or reference.
+### Scoring reference:
+- *5: Shape, border characteristics, and internal texture are highly consistent.*
+- *4: Minor border/shape irregularities; still plausible.*
+- *3: Morphology is generic/unconvincing; borders/texture inconsistent.*
+- *2: Clearly artificial morphology (blocky, repeated patterns, unnatural contours).*
+- *1: Morphology nonsensical or misleading (e.g., appears like different pathology).*
+- 3) Intensity / Signal / Attenuation Consistency: Checks whether edited region match modality-specific intensities.
+#### Scoring reference:
+- *5: Intensities match local tissue statistics; no intensity discontinuities.*
+- *4: Slight intensity mismatch detectable with careful viewing.*
+- *3: Obvious mismatch (too bright/dark), inconsistent with modality or anatomy.*
+- *2: Strong intensity discontinuity; clearly edited.*
+- *1: Severe intensity errors that invalidate the image (e.g., saturation/clipping, inverted contrast).*
+- 4) Boundary Blending & Transition Naturalness: Rates edge blending and transitions between edited and unedited regions.
+#### Scoring reference:
+- *5: Seamless blending; no halos, ringing, cut-paste edges.*
+- *4: Minor halo/edge artifacts only on close inspection.*
+- *3: Visible seams; boundary looks edited.*
+- *2: Strong cutout appearance or blur patches.*
+- *1: Boundary artifacts dominate the image.*
+- 5) Background / Non-target Preservation: Measures unintended changes outside the lesion edit region. Scoring reference:
+- *5: Non-target anatomy and background unchanged (within expected noise).*
+- *4: Small unintended changes but not clinically meaningful.*
+- *3: Noticeable unintended alterations in nearby structures.*
+- *2: Large unintended modifications to anatomy or overall image.*
+- *1: Global corruption or major anatomical distortions.*
+- 6) Anatomical Plausibility & Clinical Coherence: Assesses whether result respects anatomy and pathology logic (e.g., lesion doesn't cross impossible boundaries). Scoring reference:
+- *5: Fully plausible; consistent with organ boundaries and expected presentation.*
+- *4: Mostly plausible; minor oddity but acceptable.*
+- *3: Questionable plausibility (e.g., lesion overlaps structures unnaturally).*
+- *2: Clearly implausible anatomy/pathology relationship.*
+- *1: Clinically nonsensical or misleading.*
+- 7) Artifact Introduction (Noise, Texture, Aliasing, Compression, Repetition): Evaluates new artifacts introduced by editing. Scoring reference:
+{17}------------------------------------------------
+- *5: No new artifacts; noise texture consistent with original.*
+- *4: Minor artifacts (subtle smoothing/grain mismatch).*
+- *3: Artifacts visible and distracting.*
+- *2: Strong artifacts (banding, checkerboard, repeated texture).*
+- *1: Severe artifacts preventing clinical use.*
+- 8) Image Quality & Acquisition Consistency: Checks consistency with scanner characteristics (resolution, blur, point spread, slice thickness cues, motion). Scoring reference:
+- *5: Matches acquisition characteristics; sharpness/noise consistent.*
+- *4: Slight mismatch in sharpness or noise level.*
+- *3: Clear mismatch (over-smoothed or over-sharpened region).*
+- *2: Strong mismatch; edited region appears from different source.*
+- *1: Completely inconsistent with acquisition; unusable.*
+#### Return Format: Return a JSON dictionary with two fields:
+- "conclusion": A brief conclusion to the edited image.
+- "score list": The scores of the eight aspects in a JSON list.
+#### D.2.2. CASE
+![](_page_17_Picture_26.jpeg)
+# **Input collage VLM Output**
+**conclusion**: The edited image successfully removes the renal cyst and replaces it with plausible parenchyma closely matching the reference, with only mild residual smoothing/texture mismatch in the filled region and minimal impact on surrounding anatomy.
+**score\_list**: [5, 4, 4, 4, 5, 5, 4, 4]
+![](_page_17_Picture_31.jpeg)
+**conclusion**: The edited image does not introduce a clinically plausible gallstone. Instead, it contains multiple obvious synthetic circular overlays and line artifacts across the abdomen, with major non-target corruption and mismatch to CT texture; it does not match the reference which shows no such artifacts.
+**score\_list**: [1, 1, 1, 1, 1, 1, 1, 1]
+*Figure 8.* Cases of VLM rubric scoring.
+{18}------------------------------------------------
+# **E. Supplementary Experiments**
+#### <span id="page-18-1"></span><span id="page-18-0"></span>E.1. Modality-Wise Performance Analysis
+![](_page_18_Figure_3.jpeg)
+Figure 9. Modality-wise performance analysis result within perception perspective. Left: DICE score; right: PSNR score.
+To investigate the impact of modality deviation, we conduct a modality-wise analysis of the benchmarking results within the Perception perspective. Specifically, we report the DICE and PSNR scores of six representative models across all medical imaging modalities included in MieDB-100k. As illustrated in Fig. 9, the experimental results are consistent with our hypotheses. For the baseline models, performance is unevenly distributed across the various modalities: They achieve relatively strong results on modalities that resemble natural images, such as Endoscopy, Dermoscopy, and Surgical Photo. However, on non-optical modalities (e.g., CT, MRI, Ultrasound), their performance degrades drastically. In contrast, the model trained on our dataset exhibits balanced and superior performance across all imaging types. Collectively, these results demonstrate that a diverse dataset like MieDB-100k is essential for successfully adapting multi-modal generative models to the medical domain.
+#### <span id="page-18-2"></span>E.2. Multi-Round Generation
+Table 6. Multi-round generation result. Best values are marked in Bold
+| | Perception | | | | | | | | Trasnformation | | | | |
+|---------------------|------------|--------|--------|--------|--------|--------|--------|--------|----------------|--------|--------|--------|--|
+| | DICE | | P-A | P-ACC | | B-PSNR | | B-SSIM | | PSNR | | SSIM | |
+| | Pass@1 | Pass@3 | Pass@1 | Pass@3 | Pass@1 | Pass@3 | Pass@1 | Pass@3 | Pass@1 | Pass@3 | Pass@1 | Pass@3 | |
+| Open-Source | | | | | | | | | | | | | |
+| SDXL-turbo | 0.002 | 0.003 | 0.000 | 0.000 | 16.6 | 17.0 | 0.467 | 0.484 | 15.9 | 16.1 | 0.397 | 0.427 | |
+| Bagel | 0.263 | 0.383 | 0.069 | 0.137 | 13.9 | 16.1 | 0.620 | 0.703 | 12.7 | 15.2 | 0.442 | 0.548 | |
+| OmniGen2 | 0.248 | 0.357 | 0.065 | 0.125 | 11.9 | 14.4 | 0.541 | 0.628 | 8.3 | 16.0 | 0.280 | 0.551 | |
+| Step1X-Edit | 0.332 | 0.369 | 0.126 | 0.143 | 15.5 | 16.4 | 0.727 | 0.748 | 16.6 | 17.1 | 0.539 | 0.558 | |
+| FLUX.1-Kontext-dev | 0.347 | 0.41 | 0.126 | 0.174 | 15.4 | 16.5 | 0.701 | 0.761 | 17.9 | 19.5 | 0.543 | 0.602 | |
+| Qwen-Image-Edit | 0.387 | 0.493 | 0.153 | 0.249 | 15.4 | 17.4 | 0.722 | 0.795 | 18.9 | 20.3 | 0.606 | 0.652 | |
+| OmniGen2-MIE (Ours) | 0.831 | 0.856 | 0.737 | 0.789 | 28.1 | 28.8 | 0.917 | 0.921 | 22.6 | 23.3 | 0.685 | 0.711 | |
+To mitigate the inherent variance of the generative process, we report Pass@3 scores for the open-source models on Perception tasks. Specifically, we generate three independent outputs for each editing task and select the highest-performing sample to represent the task's score. These results are then averaged across all tasks to provide a robust assessment of overall performance.
+The results of the multi-round generation tests are summarized in Table 6. While multi-round generation improves the absolute scores for baseline models, it does not alter the underlying fact that these models lack essential medical knowledge. Furthermore, the significant fluctuations across rounds expose the high-variance nature of these baselines, undermining their reliability under clinical applications. In contrast, our model exhibits remarkable stability across all three trials. This
+{19}------------------------------------------------
+consistency suggests that model trained on MieDB-100k has developed a deterministic understanding of medical concepts rather than relying on fortuitous generation.
+#### E.3. Out-Of-Distribution Image Edit
+While Section [4.5](#page-7-2) demonstrates that the model trained on MieDB-100k generalizes effectively to OOD editing targets, we further evaluate its robustness by performing edits on 'in-the-wild' medical images sourced from the internet (Fig. [10\)](#page-19-0).
+<span id="page-19-0"></span>![](_page_19_Figure_4.jpeg)
+*Figure 10.* Examples of Out-Of-Distribution Editing.
+The results indicate that our model is capable of readily adapting to medical images outside of datasets. This suggests that the diversity of MieDB-100k has successfully decoupled the model from specific data distribution, allowing it to internalize generalizable edit operations that are applicable to real-world clinical scenarios.
+# F. Data Gallery and More Qualitative Result
+#### F.1. Examples of Healthy Tissue Inpainting
+![](_page_19_Figure_9.jpeg)
+*Figure 11.* Examples of Inpainting. We train different inpainting models on each medical modalities. H: the Healthy image; L: the Lesion-bearing image.
+#### F.2. Extended Examples of Qualitative result
+{20}------------------------------------------------
+Input Image Bagel FLUX.1 KontextOmniGen2OmniGen2-MIE Reference Nano Banana ProGPT Qwen Image Edit -Image **[Edit Instruction]** Illustrate all nuclei instances by painting GREEN masks directly onto the microscopy image, with zero modification to the background. Edit **[Edit Instruction]** Use a RED mask to highlight the polyp(s) by drawing it directly onto the given image, preserving the background exactly as is. Edit **[Edit Instruction]** Overlay unique RED masks on each organ in the ultrasound image by painting them directly, keeping the original background intact. Edit **[Edit Instruction]** Visually segment Caenorhabditis elegans by applying BLUE masks directly onto the microscope image without affecting the background. **[Edit Instruction]** Mark the region corresponding to BLADDER TUMOR in the Enhanced CT image using a solid RED mask applied directly, while ensuring the background stays unmodified. **[Edit Instruction]** Encode the lesion location with a GREEN mask drawn directly on the image, maintaining original background appearance. **[Edit Instruction]** Cover the infected zones with a RED mask by painting it directly onto the chest X-ray, ensuring background fidelity. Edit Edit Edit Edit
+{21}------------------------------------------------
+![](_page_21_Picture_2.jpeg)
+**[Edit Instruction]** Delete all visible manifestations of COVID pneumonia from the X-ray image and replace them with normal lung appearance, maintaining vascular and bronchial structures.
+![](_page_21_Picture_4.jpeg)
+**[Edit Instruction]** Eliminate the Tumor in the Stomach while ensuring the reconstructed zone blends naturally with adjacent healthy tissue in terms of Hounsfield units and spatial patterns.
+![](_page_21_Picture_6.jpeg)
+**[Edit Instruction]** Simulate a focal mucosal lesion in the white masked region that could represent a common colonic or gastric polyp, based on realistic endoscopic criteria.
+![](_page_21_Picture_8.jpeg)
+**[Edit Instruction]** Reconstruct the affected skin area by removing the lesion and generating realistic, symmetryconsistent healthy tissue based on surrounding context.
+![](_page_21_Picture_10.jpeg)
+**[Edit Instruction]** Convert the white masked area into a realistic representation of healthy skin, avoiding artificial smoothness or color mismatches.
+![](_page_21_Picture_12.jpeg)
+**[Edit Instruction]** Synthetically insert a thyroid nodule into the given ultrasound scan, ensuring it exhibits clinically plausible features such as shape, margin, echogenicity, and vascularity.
+![](_page_21_Picture_14.jpeg)
+**[Edit Instruction]** Add a convincing brain tumor to the FLAIR MRI while maintaining correct contrast dynamics—such as suppressed CSF and bright pathological signal.
+{22}------------------------------------------------
+![](_page_22_Picture_2.jpeg)
+![](_page_22_Picture_3.jpeg)
+**[Edit Instruction]** Improve diagnostic clarity of the CT by removing metal degradation while ensuring zero change to medically relevant content outside the artifact zone.
+![](_page_22_Picture_5.jpeg)
+**[Edit Instruction]** Filter out motion artifacts from the PD MRI scan without smoothing, warping, or otherwise altering true tissue signals.
+![](_page_22_Picture_7.jpeg)
+**[Edit Instruction]** Render the CT with Abdominal window parameters to mimic how it would appear on a PACS viewer configured for that tissue type.
+![](_page_22_Picture_9.jpeg)
+**[Edit Instruction]** Synthesize a Arterial phase-equivalent CT image from the provided Non-Contrast phase input, ensuring realistic vascular and tissue enhancement patterns.
+![](_page_22_Picture_11.jpeg)
+**[Edit Instruction]** Apply post-processing to remove metal artifacts from the CT image without smoothing, interpolating, or altering real anatomical details unnecessarily.
+![](_page_22_Picture_13.jpeg)
+**[Edit Instruction]** Convert the given T2-MRI image to resemble a CT image, maintaining all underlying medical information intact.
+![](_page_22_Picture_15.jpeg)
+**[Edit Instruction]** Process the low-dose CT using a SOFT kernel to suppress graininess while preserving low-contrast diagnostic details.
+{23}------------------------------------------------
+# <span id="page-23-0"></span>G. Failure Cases
+![](_page_23_Figure_2.jpeg)
+*Figure 12.* Failure Cases.
+Fig [12](#page-23-0) illustrates several representative failure cases of OmniGen2-MIE. The most frequent failure modes include: (1) semantic confusion between targeted anatomical features and morphologically similar background tissues; (2) intensity inconsistency, where the brightness of the edited region deviates from the surrounding context in a physically implausible manner; and (3) background inconsistency, especially after holistic transformations. These limitations underscore the need for more sophisticated multimodal architectures capable of preserving fine-grained details, as well as even more comprehensive training datasets to satisfy the requirements of rigorous clinical applications.
+# H. Limitations
+Despite MieDB-100k provides a large-scale and diverse dataset for medical image editing, the primary limitation lies in the inherent difficulty of capturing ALL possible medical imaging modalities, and the relative scarcity of data for rare clinical cases. Continuous efforts to enrich these underrepresented categories will be vital for enhancing the dataset's diversity and effectivity. Furthermore, while our work establishes a foundation for unified understanding and generation, it focuses exclusively on editing tasks. Integrating medical VQA and text-to-image datasets represents a natural progression of this research direction, resulting in a more comprehensive resource for the development of holistic medical models.