Dataset Viewer
Auto-converted to Parquet Duplicate
bdmap_id
stringlengths
14
14
liver_label
stringclasses
3 values
pancreas_label
stringclasses
4 values
kidney_label
stringclasses
4 values
liver_max_mm
float64
0
200
pancreas_max_mm
float64
0
135
kidney_max_mm
float64
0
179
BDMAP_00000005
negative
negative
negative
0
0
0
BDMAP_00000006
small_only
negative
negative
8
0
0
BDMAP_00000010
negative
small_only
negative
0
8
0
BDMAP_00000056
negative
negative
positive
0
0
35
BDMAP_00000065
positive
negative
negative
13
0
0
BDMAP_00000078
negative
negative
negative
0
0
0
BDMAP_00000095
negative
negative
negative
0
0
0
BDMAP_00000099
negative
negative
negative
0
0
0
BDMAP_00000118
negative
positive
negative
0
27
0
BDMAP_00000126
negative
negative
negative
0
0
0
BDMAP_00000127
negative
negative
positive
0
0
21
BDMAP_00000134
negative
positive
negative
0
18
0
BDMAP_00000138
negative
negative
negative
0
0
0
BDMAP_00000155
negative
positive
negative
0
21
0
BDMAP_00000162
negative
negative
positive
0
0
30
BDMAP_00000190
negative
negative
negative
0
0
0
BDMAP_00000191
negative
negative
negative
0
0
0
BDMAP_00000198
negative
negative
negative
0
0
0
BDMAP_00000203
negative
negative
positive
0
0
29
BDMAP_00000207
positive
negative
negative
18
0
0
BDMAP_00000225
negative
positive
negative
0
76
0
BDMAP_00000226
negative
negative
negative
0
0
0
BDMAP_00000233
negative
negative
positive
0
0
93
BDMAP_00000236
negative
positive
negative
0
20
0
BDMAP_00000250
positive
positive
positive
11
18
15
BDMAP_00000256
negative
positive
positive
0
47
13
BDMAP_00000261
negative
negative
positive
0
0
31
BDMAP_00000263
negative
negative
positive
0
0
36
BDMAP_00000302
negative
negative
negative
0
0
0
BDMAP_00000303
negative
negative
negative
0
0
0
BDMAP_00000346
negative
negative
positive
0
0
20
BDMAP_00000365
negative
negative
negative
0
0
0
BDMAP_00000385
negative
negative
negative
0
0
0
BDMAP_00000392
negative
negative
negative
0
0
0
BDMAP_00000396
positive
negative
negative
24
0
0
BDMAP_00000405
negative
negative
positive
0
0
25
BDMAP_00000413
positive
negative
negative
117
0
0
BDMAP_00000418
negative
positive
positive
0
39
15
BDMAP_00000420
positive
negative
negative
88
0
0
BDMAP_00000467
negative
positive
negative
0
23
0
BDMAP_00000503
positive
positive
positive
10
30
19
BDMAP_00000505
negative
negative
positive
0
0
30
BDMAP_00000513
positive
negative
positive
162
0
12
BDMAP_00000524
positive
negative
negative
65
0
0
BDMAP_00000538
negative
negative
positive
0
0
49
BDMAP_00000540
negative
negative
positive
0
0
27
BDMAP_00000560
negative
negative
negative
0
0
0
BDMAP_00000570
negative
negative
negative
0
0
0
BDMAP_00000575
negative
negative
positive
0
0
11
BDMAP_00000581
negative
negative
negative
0
0
0
BDMAP_00000601
positive
positive
positive
21
35
40
BDMAP_00000607
positive
negative
negative
102
0
0
BDMAP_00000636
negative
negative
negative
0
0
0
BDMAP_00000639
positive
negative
negative
20
0
0
BDMAP_00000660
negative
negative
positive
0
0
35
BDMAP_00000664
negative
negative
negative
0
0
0
BDMAP_00000667
negative
negative
positive
0
0
52
BDMAP_00000668
small_only
negative
small_only
8
0
8
BDMAP_00000672
negative
negative
positive
0
0
10
BDMAP_00000699
negative
negative
small_only
0
0
9
BDMAP_00000700
negative
negative
negative
0
0
0
BDMAP_00000717
negative
positive
negative
0
15
0
BDMAP_00000718
negative
negative
negative
0
0
0
BDMAP_00000729
positive
positive
negative
89
21
0
BDMAP_00000761
negative
negative
negative
0
0
0
BDMAP_00000770
negative
negative
positive
0
0
25
BDMAP_00000771
negative
positive
negative
0
17
0
BDMAP_00000776
negative
negative
positive
0
0
33
BDMAP_00000785
negative
negative
negative
0
0
0
BDMAP_00000794
negative
negative
positive
0
0
109
BDMAP_00000815
positive
negative
positive
20
0
40
BDMAP_00000824
negative
negative
positive
0
0
14
BDMAP_00000826
negative
negative
positive
0
0
33
BDMAP_00000828
negative
negative
negative
0
0
0
BDMAP_00000831
positive
negative
negative
102
0
0
BDMAP_00000847
negative
negative
negative
0
0
0
BDMAP_00000852
negative
negative
negative
0
0
0
BDMAP_00000854
negative
negative
negative
0
0
0
BDMAP_00000855
negative
positive
negative
0
25
0
BDMAP_00000861
negative
negative
negative
0
0
0
BDMAP_00000864
negative
negative
negative
0
0
0
BDMAP_00000877
positive
negative
negative
55
0
0
BDMAP_00000879
negative
negative
negative
0
0
0
BDMAP_00000919
negative
positive
negative
0
13
0
BDMAP_00000923
negative
negative
positive
0
0
105
BDMAP_00000925
negative
negative
small_only
0
0
8
BDMAP_00000929
negative
negative
negative
0
0
0
BDMAP_00000951
positive
negative
positive
21
0
12
BDMAP_00000954
negative
negative
negative
0
0
0
BDMAP_00000957
negative
negative
positive
0
0
37
BDMAP_00000981
negative
negative
negative
0
0
0
BDMAP_00000985
positive
negative
positive
10
0
83
BDMAP_00000996
negative
negative
negative
0
0
0
BDMAP_00001016
positive
positive
negative
70
23
0
BDMAP_00001017
negative
negative
negative
0
0
0
BDMAP_00001022
negative
negative
negative
0
0
0
BDMAP_00001029
positive
negative
negative
14
0
0
BDMAP_00001046
negative
negative
negative
0
0
0
BDMAP_00001060
negative
negative
negative
0
0
0
BDMAP_00001061
negative
negative
negative
0
0
0
End of preview. Expand in Data Studio

ORGANIC-3D benchmark

Does a 3D CT model's prediction for a finding follow the image evidence for that finding? This benchmark answers it with anatomical interventions on the CT volume, for contrastive encoders and generative VLMs alike, on five public datasets: CT-RATE, RAD-ChestCT, Merlin, ReXGroundingCT and AbdomenAtlas 3.0. Paper: ORGANIC-3D: Grounding 3D CT Vision Language Models in Organ-Level Image–Text Coupling. Code: https://github.com/lzl199704/organic3d · Models: https://huggingface.co/zllhhhh/organic3d-encoder, https://huggingface.co/zllhhhh/organic3d-vlm

Test Metric What it checks
Classification MLC-AUROC; lesion AUROC; zero-shot tumour detection accuracy
Clinical finding independence matched-pair AUROC (MP) patients who differ only in the finding are ranked correctly
Organ dependence organ-erasure ΔAUROC (OE); for VLMs the organ-erasure hit rate (OEH) erasing the finding's organ, not a control organ, changes the prediction
Laterality Side; lung lobe erasure hit rate (LL) the sentence that names the correct side or lobe wins
Lesion erasure lesion-erasure hit rate (LE) replacing an expert-segmented lesion by normal tissue lowers its sentence's score more than the same edit of a healthy region

Contents

protocol/      protocol.json (findings, organ groups, controls, prompts, interventions), builders/ that derive every
               scan and finding list from your own copy of each dataset, files/ with the shipped labels:
               merlin_test_labels.csv (the paper's Merlin test labels) and the AbdomenAtlas 3.0 tumour labels
metric/        organic3d_benchmark.py: the metrics with patient-level bootstrap CIs (CPU)
leaderboard/   the paper's results for 11 released encoders, 7 released VLMs and ORGANIC-3D (results.csv), and a
               Gradio app
harness/       the evaluation harness and metric scripts (snapshot of the code repository at commit d545a61)

Merlin test labels (protocol/files/merlin_test_labels.csv): 5,109 test studies × 21 abdominal findings, 1 / 0, empty = uncertain, re-extracted from the reports with LLMs (paper Sec. 4.1). They contain no report text; use them under Merlin's terms (protocol/files/LICENSE_merlin_labels.md).

Evaluate a model

  1. Get the datasets from their official sources and prepare them (harness/docs/data_preparation.md; pass the Merlin labels with scripts/data/prepare_manifests.py --merlin_test_labels protocol/files/merlin_test_labels.csv).
  2. Write an adapter for your model (harness/docs/adapter_guide.md).
  3. Run every test and metric of the paper with one command:
    cd harness
    python scripts/evaluation/evaluate.py encoder --model <adapter> --name <MyModel> [--ckpt ...] [--model_python <env>/bin/python]
    python scripts/evaluation/evaluate.py vlm --model <adapter> --name <MyVLM> ...
    
    --dry_run prints the steps; without arguments the script asks for each option. Results go to {work_root}/eval/.

harness/demo/organic3d_demo.py runs the tests on a single CT and draws them (see the code repository's README).

Leaderboard

leaderboard/results.csv (one row per model × dataset × task × metric, with 95% CIs); leaderboard/app.py shows it with the paper's metric names. Run ids: enc4_organic3d (ORGANIC-3D encoder), vlm4_organic3d (ORGANIC-3D VLM), vlm1_plain (plain fine-tuning).

Citation

@article{liu2026organic3d,
  title   = {ORGANIC3D: Grounding 3D CT Vision Language Models in Organ Level Image Text Coupling},
  author  = {Liu, Zelong and Ahmed, Hamza and Lee, Wynn Wingyi and Torjman, Mickael and Dimitrakakis, Alexander and
             Lu, Jiahua and Deyer, Timothy and Taouli, Bachir and Fayad, Zahi A. and Mei, Xueyan},
  journal = {arXiv preprint arXiv:2611.23529},
  year    = {2026}
}

Please also cite the datasets: CT-RATE (Hamamci et al., 2024), RAD-ChestCT (Draelos et al., 2021), Merlin (Blankemeier et al., 2024), ReXGroundingCT (Baharoon et al., 2026) and AbdomenAtlas 3.0 (Bassi et al., 2025).

Downloads last month
2

Paper for zllhhhh/organic3d-benchmark