{ "benchmark": "ORGANIC-3D", "protocol_version": "1.6.0", "findings": [ { "name": "ct_medical_material", "ctrate_label_column": "Medical material", "target_organs": [], "control_organs": [ "lung_left", "lung_right", "heart", "aorta", "esophagus", "trachea", "liver", "gallbladder", "spleen", "pancreas", "kidney_left", "kidney_right", "stomach", "small_bowel", "colon", "pelvic_organs" ], "phrase": "medical material", "size_finding": false, "prompt_positive": "There is medical material.", "prompt_negative": "No medical material.", "vlm_question": "Is there medical material in this CT? Answer Yes or No.", "radchest_mapping": { "abnormalities": [ "catheter_or_port", "pacemaker_or_defib", "stent", "chest_tube", "gi_tube", "tracheal_tube", "heart_valve_replacement", "hardware", "clip", "staple", "suture" ], "locations": null, "approximate": false } }, { "name": "ct_arterial_calcification", "ctrate_label_column": "Arterial wall calcification", "target_organs": [ "aorta" ], "control_organs": [ "lung_left", "lung_right", "trachea", "liver", "gallbladder", "spleen", "pancreas", "kidney_left", "kidney_right", "stomach", "small_bowel", "colon", "pelvic_organs" ], "phrase": "arterial wall calcification", "size_finding": false, "prompt_positive": "There is arterial wall calcification.", "prompt_negative": "No arterial wall calcification.", "vlm_question": "Is there arterial wall calcification in this CT? Answer Yes or No.", "radchest_mapping": { "abnormalities": [ "atherosclerosis", "calcification", "plaque" ], "locations": [ "aorta", "pulmonary_artery", "svc" ], "approximate": false } }, { "name": "ct_cardiomegaly", "ctrate_label_column": "Cardiomegaly", "target_organs": [ "heart" ], "control_organs": [ "esophagus", "trachea", "liver", "gallbladder", "spleen", "pancreas", "kidney_left", "kidney_right", "stomach", "small_bowel", "colon", "pelvic_organs" ], "phrase": "cardiomegaly", "size_finding": true, "prompt_positive": "There is cardiomegaly.", "prompt_negative": "No cardiomegaly.", "vlm_question": "Is there cardiomegaly in this CT? Answer Yes or No.", "radchest_mapping": { "abnormalities": [ "cardiomegaly" ], "locations": null, "approximate": false } }, { "name": "ct_pericardial_effusion", "ctrate_label_column": "Pericardial effusion", "target_organs": [ "heart" ], "control_organs": [ "esophagus", "trachea", "liver", "gallbladder", "spleen", "pancreas", "kidney_left", "kidney_right", "stomach", "small_bowel", "colon", "pelvic_organs" ], "phrase": "pericardial effusion", "size_finding": false, "prompt_positive": "There is pericardial effusion.", "prompt_negative": "No pericardial effusion.", "vlm_question": "Is there pericardial effusion in this CT? Answer Yes or No.", "radchest_mapping": { "abnormalities": [ "pericardial_effusion" ], "locations": null, "approximate": false } }, { "name": "ct_coronary_calcification", "ctrate_label_column": "Coronary artery wall calcification", "target_organs": [ "heart" ], "control_organs": [ "esophagus", "trachea", "liver", "gallbladder", "spleen", "pancreas", "kidney_left", "kidney_right", "stomach", "small_bowel", "colon", "pelvic_organs" ], "phrase": "coronary artery calcification", "size_finding": false, "prompt_positive": "There is coronary artery calcification.", "prompt_negative": "No coronary artery calcification.", "vlm_question": "Is there coronary artery calcification in this CT? Answer Yes or No.", "radchest_mapping": { "abnormalities": [ "coronary_artery_disease" ], "locations": null, "approximate": false } }, { "name": "ct_hiatal_hernia", "ctrate_label_column": "Hiatal hernia", "target_organs": [ "esophagus", "stomach" ], "control_organs": [ "lung_left", "lung_right", "heart", "trachea", "liver", "gallbladder", "spleen", "pancreas", "kidney_left", "kidney_right", "small_bowel", "colon", "pelvic_organs" ], "phrase": "hiatal hernia", "size_finding": true, "prompt_positive": "There is hiatal hernia.", "prompt_negative": "No hiatal hernia.", "vlm_question": "Is there hiatal hernia in this CT? Answer Yes or No.", "radchest_mapping": { "abnormalities": [ "hernia" ], "locations": [ "esophagus", "stomach", "diaphragm" ], "approximate": false } }, { "name": "ct_lymphadenopathy", "ctrate_label_column": "Lymphadenopathy", "target_organs": [], "control_organs": [ "lung_left", "lung_right", "heart", "aorta", "esophagus", "trachea", "liver", "gallbladder", "spleen", "pancreas", "kidney_left", "kidney_right", "stomach", "small_bowel", "colon", "pelvic_organs" ], "phrase": "mediastinal lymphadenopathy", "size_finding": false, "prompt_positive": "There is mediastinal lymphadenopathy.", "prompt_negative": "No mediastinal lymphadenopathy.", "vlm_question": "Is there mediastinal lymphadenopathy in this CT? Answer Yes or No.", "radchest_mapping": { "abnormalities": [ "lymphadenopathy" ], "locations": null, "approximate": false } }, { "name": "ct_emphysema", "ctrate_label_column": "Emphysema", "target_organs": [ "lung_left", "lung_right" ], "control_organs": [ "aorta", "esophagus", "trachea", "liver", "gallbladder", "spleen", "pancreas", "kidney_left", "kidney_right", "stomach", "small_bowel", "colon", "pelvic_organs" ], "phrase": "emphysema", "size_finding": false, "prompt_positive": "There is emphysema.", "prompt_negative": "No emphysema.", "vlm_question": "Is there emphysema in this CT? Answer Yes or No.", "radchest_mapping": { "abnormalities": [ "emphysema" ], "locations": null, "approximate": false } }, { "name": "ct_atelectasis", "ctrate_label_column": "Atelectasis", "target_organs": [ "lung_left", "lung_right" ], "control_organs": [ "aorta", "esophagus", "trachea", "liver", "gallbladder", "spleen", "pancreas", "kidney_left", "kidney_right", "stomach", "small_bowel", "colon", "pelvic_organs" ], "phrase": "atelectasis", "size_finding": false, "prompt_positive": "There is atelectasis.", "prompt_negative": "No atelectasis.", "vlm_question": "Is there atelectasis in this CT? Answer Yes or No.", "radchest_mapping": { "abnormalities": [ "atelectasis" ], "locations": null, "approximate": false } }, { "name": "ct_lung_nodule", "ctrate_label_column": "Lung nodule", "target_organs": [ "lung_left", "lung_right" ], "control_organs": [ "aorta", "esophagus", "trachea", "liver", "gallbladder", "spleen", "pancreas", "kidney_left", "kidney_right", "stomach", "small_bowel", "colon", "pelvic_organs" ], "phrase": "lung nodule", "size_finding": false, "prompt_positive": "There is lung nodule.", "prompt_negative": "No lung nodule.", "vlm_question": "Is there lung nodule in this CT? Answer Yes or No.", "radchest_mapping": { "abnormalities": [ "nodule", "nodulegr1cm" ], "locations": [ "left_upper", "left_mid", "left_lower", "right_upper", "right_mid", "right_lower", "right_lung", "left_lung", "lung", "interstitial", "subpleural", "centrilobular" ], "approximate": false } }, { "name": "ct_lung_opacity", "ctrate_label_column": "Lung opacity", "target_organs": [ "lung_left", "lung_right" ], "control_organs": [ "aorta", "esophagus", "trachea", "liver", "gallbladder", "spleen", "pancreas", "kidney_left", "kidney_right", "stomach", "small_bowel", "colon", "pelvic_organs" ], "phrase": "lung opacity", "size_finding": false, "prompt_positive": "There is lung opacity.", "prompt_negative": "No lung opacity.", "vlm_question": "Is there lung opacity in this CT? Answer Yes or No.", "radchest_mapping": { "abnormalities": [ "opacity", "groundglass" ], "locations": [ "left_upper", "left_mid", "left_lower", "right_upper", "right_mid", "right_lower", "right_lung", "left_lung", "lung", "interstitial", "subpleural", "centrilobular" ], "approximate": false } }, { "name": "ct_fibrotic_sequela", "ctrate_label_column": "Pulmonary fibrotic sequela", "target_organs": [ "lung_left", "lung_right" ], "control_organs": [ "aorta", "esophagus", "trachea", "liver", "gallbladder", "spleen", "pancreas", "kidney_left", "kidney_right", "stomach", "small_bowel", "colon", "pelvic_organs" ], "phrase": "pulmonary fibrotic sequela", "size_finding": false, "prompt_positive": "There is pulmonary fibrotic sequela.", "prompt_negative": "No pulmonary fibrotic sequela.", "vlm_question": "Is there pulmonary fibrotic sequela in this CT? Answer Yes or No.", "radchest_mapping": { "abnormalities": [ "fibrosis", "scarring" ], "locations": [ "left_upper", "left_mid", "left_lower", "right_upper", "right_mid", "right_lower", "right_lung", "left_lung", "lung", "interstitial", "subpleural", "centrilobular" ], "approximate": false } }, { "name": "ct_pleural_effusion", "ctrate_label_column": "Pleural effusion", "target_organs": [], "control_organs": [ "lung_left", "lung_right", "heart", "aorta", "esophagus", "trachea", "liver", "gallbladder", "spleen", "pancreas", "kidney_left", "kidney_right", "stomach", "small_bowel", "colon", "pelvic_organs" ], "phrase": "pleural effusion", "size_finding": false, "prompt_positive": "There is pleural effusion.", "prompt_negative": "No pleural effusion.", "vlm_question": "Is there pleural effusion in this CT? Answer Yes or No.", "radchest_mapping": { "abnormalities": [ "pleural_effusion" ], "locations": null, "approximate": false } }, { "name": "ct_mosaic_attenuation", "ctrate_label_column": "Mosaic attenuation pattern", "target_organs": [ "lung_left", "lung_right" ], "control_organs": [ "aorta", "esophagus", "trachea", "liver", "gallbladder", "spleen", "pancreas", "kidney_left", "kidney_right", "stomach", "small_bowel", "colon", "pelvic_organs" ], "phrase": "mosaic attenuation pattern", "size_finding": false, "prompt_positive": "There is mosaic attenuation pattern.", "prompt_negative": "No mosaic attenuation pattern.", "vlm_question": "Is there mosaic attenuation pattern in this CT? Answer Yes or No.", "radchest_mapping": { "abnormalities": [ "air_trapping" ], "locations": null, "approximate": true } }, { "name": "ct_peribronchial_thickening", "ctrate_label_column": "Peribronchial thickening", "target_organs": [ "lung_left", "lung_right" ], "control_organs": [ "aorta", "esophagus", "trachea", "liver", "gallbladder", "spleen", "pancreas", "kidney_left", "kidney_right", "stomach", "small_bowel", "colon", "pelvic_organs" ], "phrase": "peribronchial thickening", "size_finding": false, "prompt_positive": "There is peribronchial thickening.", "prompt_negative": "No peribronchial thickening.", "vlm_question": "Is there peribronchial thickening in this CT? Answer Yes or No.", "radchest_mapping": { "abnormalities": [ "bronchial_wall_thickening" ], "locations": null, "approximate": false } }, { "name": "ct_consolidation", "ctrate_label_column": "Consolidation", "target_organs": [ "lung_left", "lung_right" ], "control_organs": [ "aorta", "esophagus", "trachea", "liver", "gallbladder", "spleen", "pancreas", "kidney_left", "kidney_right", "stomach", "small_bowel", "colon", "pelvic_organs" ], "phrase": "consolidation", "size_finding": false, "prompt_positive": "There is consolidation.", "prompt_negative": "No consolidation.", "vlm_question": "Is there consolidation in this CT? Answer Yes or No.", "radchest_mapping": { "abnormalities": [ "consolidation" ], "locations": null, "approximate": false } }, { "name": "ct_bronchiectasis", "ctrate_label_column": "Bronchiectasis", "target_organs": [ "lung_left", "lung_right" ], "control_organs": [ "aorta", "esophagus", "trachea", "liver", "gallbladder", "spleen", "pancreas", "kidney_left", "kidney_right", "stomach", "small_bowel", "colon", "pelvic_organs" ], "phrase": "bronchiectasis", "size_finding": false, "prompt_positive": "There is bronchiectasis.", "prompt_negative": "No bronchiectasis.", "vlm_question": "Is there bronchiectasis in this CT? Answer Yes or No.", "radchest_mapping": { "abnormalities": [ "bronchiectasis" ], "locations": null, "approximate": false } }, { "name": "ct_septal_thickening", "ctrate_label_column": "Interlobular septal thickening", "target_organs": [ "lung_left", "lung_right" ], "control_organs": [ "aorta", "esophagus", "trachea", "liver", "gallbladder", "spleen", "pancreas", "kidney_left", "kidney_right", "stomach", "small_bowel", "colon", "pelvic_organs" ], "phrase": "interlobular septal thickening", "size_finding": false, "prompt_positive": "There is interlobular septal thickening.", "prompt_negative": "No interlobular septal thickening.", "vlm_question": "Is there interlobular septal thickening in this CT? Answer Yes or No.", "radchest_mapping": { "abnormalities": [ "septal_thickening" ], "locations": null, "approximate": false } } ], "organ_groups_totalsegmentator_v2": { "lung_left": [ 10, 11 ], "lung_right": [ 12, 13, 14 ], "heart": [ 51, 61 ], "aorta": [ 52 ], "esophagus": [ 15 ], "trachea": [ 16 ], "liver": [ 5 ], "gallbladder": [ 4 ], "spleen": [ 1 ], "pancreas": [ 7 ], "kidney_left": [ 3, 23 ], "kidney_right": [ 2, 24 ], "stomach": [ 6 ], "small_bowel": [ 18, 19 ], "colon": [ 20 ], "pelvic_organs": [ 21, 22 ] }, "related_organs": { "liver": [ "gallbladder" ], "gallbladder": [ "liver" ], "stomach": [ "esophagus" ], "esophagus": [ "aorta", "stomach" ], "small_bowel": [ "colon" ], "colon": [ "small_bowel" ], "lung_left": [ "heart", "lung_right" ], "lung_right": [ "heart", "lung_left" ], "heart": [ "aorta", "lung_left", "lung_right" ], "aorta": [ "esophagus", "heart" ], "kidney_left": [ "kidney_right" ], "kidney_right": [ "kidney_left" ] }, "control_rule": "control organs of a finding = every organ group that is neither a target nor related to a target; an erased group counts as target if it shares an organ with the targets, as control if all its organs are controls", "eval_organ_groups": { "ctrate": [ [ "lung_left", "lung_right" ], [ "heart" ], [ "aorta" ], [ "esophagus", "stomach" ], [ "trachea" ] ], "radchest": [ [ "lung_left", "lung_right" ], [ "heart" ], [ "aorta" ], [ "esophagus", "stomach" ], [ "trachea" ] ] }, "interventions": { "common": "every intervention edits the 1 mm RAS volume in HU before the model's own preprocessing; masks are TotalSegmentator v2 \"total\" (fast) masks on the same grid", "organ_erasure": { "mask_dilation_mm": 2, "fills_hu": { "fat": -100.0, "soft": 40.0 }, "noise_sd_hu": 15.0, "conditions": "one per eval organ group and fill (encoders: fat and soft; VLMs: fat)", "skip_rule": "a scan is skipped for a condition when the group has 0 voxels (nvox = 0)" }, "mirror": { "axis": "left-right (RAS axis 0)", "masks": "mask-guided models receive the mirrored mask" }, "side_sentences": { "source": "CT-RATE report = \"Findings: \" + Findings_EN + \" Impression: \" + Impressions_EN", "rule": "sentences (split on [.!?] followed by a capital; > 8 characters; first 40) naming exactly one side (left|right) and not matching the negation pattern; at most 6 per scan; swapped = left<->right", "negation_pattern": "\\b(no|not|without|negative for|resolved|absent)\\b", "correct": "S(V, s) > S(V, s_swapped)", "follows_image": "correct AND S(mirror(V), s_swapped) > S(mirror(V), s)" }, "lesion_heal_rexgroundingct": { "fill": "median HU of the scan's lung (TotalSegmentator lobes 10-14) outside all lesions dilated 3 mm (-850 HU if empty)", "noise_sd_hu": 15.0, "control": "the lesion mask reflected left-right about the midline of the lungs' x extent", "control_valid": "overlap with any lesion (dilated 3 mm) < 5% of the lesion size AND >= 70% of the control in lung (category 2e pleural exempt)", "subsets": "held_out = CT-RATE validation scans, volume ids valid_* (held_out_ctrate_test) and scans of the ReXGroundingCT val split (held_out_rex_val); training_scans = the rest (CT-RATE train scans); ReXGroundingCT test masks are not public and not used", "category_to_finding": { "1a": "ct_peribronchial_thickening", "1b": "ct_bronchiectasis", "1c": "ct_emphysema", "1d": "ct_septal_thickening", "2a": "ct_atelectasis", "2b": "ct_consolidation", "2c": "ct_lung_opacity", "2d": "ct_lung_nodule", "2e": "ct_pleural_effusion", "2f": "ct_fibrotic_sequela" }, "categories": { "1a": "bronchial wall thickening", "1b": "bronchiectasis", "1c": "emphysema", "1d": "septal / reticular", "1e": "small airway / micronodules", "1f": "air trapping / other airway", "2a": "linear atelectasis", "2b": "consolidation / atelectasis", "2c": "ground glass / mosaic", "2d": "nodules", "2e": "pleural", "2f": "honeycombing", "2g": "bleb / pneumothorax", "2h": "other patterns" } }, "density_edits": { "emph": "lung voxels < -950 HU not connected to the airway set to -860 HU + N(0, 15)", "ca_ao": "aortic calcium (>= 130 HU after 3x3x3 mean filter, < 2000 HU raw, >= 3 voxels, + 1 mm rim) set to 40 HU + N(0, 15)", "ca_cor": "the same in a 3 mm shell around the heart outside heart, aorta (dilated 3 mm) and other labelled structures", "skip": "calcium edits skipped on contrast-enhanced scans (median aorta > 120 HU)", "own_finding": { "emph": "ct_emphysema", "ca_ao": "ct_arterial_calcification", "ca_cor": "ct_coronary_calcification" } }, "lobe_test": { "lobes": [ "right upper lobe", "right middle lobe", "right lower lobe", "left upper lobe", "left lower lobe" ], "totalsegmentator_ids": [ 12, 13, 14, 10, 11 ], "erase": "each lobe dilated 2 mm, fat fill + N(0, 15)", "sentences": "report sentences naming exactly one lobe once (with its side), not negated, not both/bilateral; at most 8 per scan; the lobe phrase is replaced by each of the five lobe names", "top1": "the named lobe causes the largest drop of the true-lobe sentence after subtracting each lobe's mean drop for sentences naming other lobes (chance 0.2)" }, "tumour_detection": { "abdomenatlas3": { "unit": "scan", "split": "IID test (926 scans)", "positives_negatives": { "liver": [ 120, 792 ], "pancreas": [ 109, 812 ], "kidney": [ 190, 693 ] }, "labels": "files/abdomenatlas3_iid_test_tumour_labels.csv", "field_of_view": "480 mm head-foot window centred on liver + pancreas", "harness_task": "aa3_det", "download": "scripts/data/download_aa3_test.sh (GitHub repo); preparation scripts/data/prepare_aa3.py" } }, "lesion_erasure": { "definition": "an expert-segmented lesion (>= 10 mm) is replaced by its organ's normal tissue: median HU of the organ outside every lesion dilated 3 mm, + N(0, 15)", "control": "the lesion's own shape translated inside the same organ (4 mm grid search), erased the same way", "control_valid": ">= 70% inside the parenchyma and < 5% overlap with any lesion dilated 10 mm", "sets": "ReXGroundingCT lung findings (sentence = the report finding); AbdomenAtlas 3.0 liver, pancreas and kidney tumours (sentence 'There is a {d} cm mass in the {organ}.')", "metric": "LE = share of lesions with s_clean - s_lesion > s_clean - s_control (chance 0.5)", "code": "organic3d repository: scripts/data/prepare_heal.py, extaudit/lesion_erasure.py, scripts/metrics/heal_metrics.py" } }, "metrics": { "min_positives_encoder": 10, "min_positives_and_negatives_vlm": 10, "probe": "StandardScaler + LogisticRegression(C=0.1, max_iter=2000, class_weight=balanced); StratifiedGroupKFold(5, shuffle, random_state=0) by patient; out-of-fold decision function; AUROC per finding; macro mean", "adr": "(mean ZS - 0.5) / (mean probe - 0.5) (ratio of macro means); VLM use likewise with the probe on the LLM-input visual tokens", "erasure_effect": "per finding and fill: mean over target groups of dAUROC minus mean over control groups of dAUROC (each on the scans that contain the group); reported as the mean over texture findings (size findings excluded)", "matched_pair_auroc": "AUROC over (positive, negative) pairs identical on every other labelled finding; prior gap = ZS - MP", "wrong_patient_auroc": "answers of another patient (the row half the list away in patient order, as finding_reports) scored against this scan's labels (control, expect 0.5)", "bootstrap": "patient-level percentile 95% CI; numpy default_rng(0); 2,000 resamples; AbdomenAtlas 3.0 unit = scan", "aggregation": "every metric is computed per finding and reported as the mean over findings (each finding counts equally)", "tumour_detection": "AUROC of s_pos - s_neg with each model's own prompt template and score rule (the MLC rule) for the findings liver, pancreatic and renal mass (AbdomenAtlas 3.0, per scan); positive = a lesion with largest extent >= 10 mm, negative = no lesion; value = mean of the three organ AUROCs; the free pair 'There is (no) mass in the [organ].' and the positive prompt alone are checks", "finding_reports": "5 patient folds (default_rng(0)); per fold and finding the F1-maximising threshold (49 quantiles 0.02-0.98) of the other folds; F1 macro over findings with >= 1 positive; wrong-patient partner = the row half the list away in patient order (next different patient); retraction = share of correctly stated findings withdrawn after erasing a target organ minus after a control organ (>= 5 scans per term)" }, "datasets": { "ctrate": { "split": "CT-RATE v2 test = official validation split (multi_abnormality_labels/valid_predicted_labels.csv) minus dataset/metadata/no_chest_valid.txt", "n_scans": 3002, "n_patients": 1304, "labels": "18 released multi-abnormality labels (valid_predicted_labels.csv)", "order": "sorted by \"//\"", "id_lists_shipped": false, "vlm_set": "the CT-RATE v2 test set (3,002 scans, protocol order; the same list as the encoders)" }, "radchest": { "split": "all 3,630 public RAD-ChestCT scans (external test)", "n_patients": 3072, "order": "img{train,valid,test}_Abnormality_and_Location_Labels.csv concatenated in that order, file row order", "labels": "abnormality x location labels mapped to the 18 findings (radchest_mapping; positive if any listed abnormality at any listed location)", "patient": "MRN_DEID of Summary_3630.csv", "id_lists_shipped": false }, "rexgroundingct": { "source": "official ReXGroundingCT release (https://huggingface.co/datasets/rajpurkarlab/ReXGroundingCT: dataset.json, segmentations/) on CT-RATE volumes (https://huggingface.co/datasets/ibrahimhamamci/CT-RATE: dataset/{train,valid}_fixed); nothing of either dataset is shipped", "scans": "every scan of the ReXGroundingCT train and val splits (all released masks; the test split is not used: its masks are not public)", "mask_mapping": "finding k = channel k of the scan mask on the original CT-RATE grid, mapped to the harness 1 mm RAS grid by world coordinates (nearest neighbour, sampled backwards inside the lesion bounding box + 1 voxel)", "finding_rule": "a released train/val finding is evaluated iff its mapped mask has >= 1 voxel on the harness grid (7 of 7,802 drop on the reference grid)", "order": "sorted by (scan_id, k)", "subset_rule": "held_out = CT-RATE validation scan (valid_*) or ReXGroundingCT val scan; training_scans = the rest", "control_validity": "computed by the harness at evaluation time from the control geometry (interventions.lesion_heal_rexgroundingct.control_valid); not an input", "n_released_findings": 7802, "n_findings": 7795, "n_scans": 3041, "n_held_out": 1267, "n_held_out_ctrate_test": 1152, "n_held_out_rex_val": 115, "finding_list_sha256": "98130f40e654ca107abec5f1e7d65142f3d86e981062f4171d49f9b8d421dc56", "finding_list_sha256_input": "one line \"scan_id,k,subset\" per finding in canonical order, each ending in a newline, UTF-8", "n_valid_findings_reference_run": 3846, "n_valid_held_out": 614, "n_valid_held_out_ctrate_test": 558, "n_valid_held_out_rex_val": 56, "builder": "protocol/builders/build_rexgroundingct.py", "id_lists_shipped": false }, "abdomenatlas3": { "role": "liver / pancreas / kidney tumour detection (zero-shot)", "n_scans": 926, "licence": "CC BY-NC-SA 4.0", "source": "https://huggingface.co/datasets/AbdomenAtlas/AbdomenAtlas3.0Mini" } }, "tasks": { "classification": { "encoder_metrics": [ "MLC-AUROC (summary key zero_shot)", "lesion AUROC: ReXGroundingCT findings; tumour detection on AbdomenAtlas 3.0" ], "vlm_metrics": [ "answer AUROC", "balanced accuracy and F1 of the stated answer (Yes when P(Yes) >= 0.5)" ] }, "clinical_finding_independence": { "metric": "MP: matched-pair AUROC over pairs identical on every other labelled finding (summary key matched_pair_auroc)" }, "organ_dependence": { "encoder_metric": "OE: organ-erasure delta-AUROC, AUROC lost after target-organ erasure minus after control-organ erasure, fat fill, non-size findings (= minus summary key effect_fat_texture)", "vlm_metric": "OEH (organ-erasure hit rate): share of correctly stated findings withdrawn after target- minus control-organ erasure (finding_reports retraction)" }, "laterality": { "encoder_metrics": [ "Side: laterality accuracy, a side-naming report sentence scores above its side-swapped version on the original scan, balanced over left and right (chance 0.50)", "LE: lesion-erasure hit rate, an expert-segmented lesion replaced by normal tissue lowers its sentence score more than the same edit of a same-shaped healthy region of the same organ (chance 0.50)", "LL: lung lobe erasure hit rate, the named lobe causes the largest drop when erased (chance 0.20)" ], "vlm": "not a VLM task (side-free label list)" } } }