spatialmed/croissant_spatialmed

SpatialMed — Croissant Package

This folder packages the SpatialMed volumetric medical MCQ benchmark for upload to Hugging Face and submission to the NeurIPS 2025 Datasets & Benchmarks track.

Scope. 8,864 multiple-choice / yes-no questions (filtered from 10,487 raw questions by (a) dropping reviewer-score < 1 rows via spatial_med_scored_final.xlsx, (b) dropping rows that cal_splits.classify_metric labels as Unknown, and (c) dropping DIST rows) over 9 source anatomical datasets (BraTS, KiTS, AMOS, TotalSegmentator, Medical-Decathlon Liver / Pancreas / Prostate / Vessel / Lung) grouped into 4 spatial-reasoning tasks (DIR, EXT, VOL, COMP). Every question is grounded in both a 3-D NIfTI volume and three orthogonal 2-D RGB views (axial / sagittal / coronal).

Folder layout

croissant/
├── README.md                 ← this file (reviewer entry point)
├── NEURIPS_GUIDELINES.md     ← summary of neurips.cc/…/DataHostingGuidelines
├── croissant.json            ← Croissant 1.0 JSON-LD metadata
├── data/
│   ├── tasks.json            ← 5-task taxonomy (DIR / DIST / EXT / VOL / COMP)
│   ├── datasets.json         ← metadata for each of the 9 source datasets
│   └── annotations.jsonl     ← (generated) one MCQ per line, w/ task_type
├── scripts/
│   ├── build_annotations.py  ← produces data/annotations.jsonl (raw)
│   ├── filter_annotations.py ← drops reviewer-score<1 rows via xlsx
│   ├── visualize_sample.py   ← render a question + 3-view + GT + reasoning
│   └── validate_croissant.py ← mlcroissant validation check
└── examples/                 ← pre-rendered PNGs (one per task)
    ├── example_DIR_amos_amos_0044.png
    ├── example_EXT_brats_BraTS2021_01056.png
    ├── example_VOL_amos_amos_0033.png
    ├── example_COMP_amos_amos_0158.png
    └── volumes/              ← 3-D NIfTI for the same 4 cases (≈105 MB)
        ├── MANIFEST.json     ← cross-reference PNG ↔ volume ↔ MCQ
        ├── BraTS2021_01056/  ← 5 modalities (t1/t1ce/t2/flair/seg)
        ├── amos_0033/        ← image + label
        ├── amos_0044/
        └── amos_0158/

Tasks (4)

Code	Name	Example question
DIR	Direction	"Is the liver superior to the pancreas?"
EXT	Extent	"What is the largest diameter of the spleen (cm)?"
VOL	Volume	"Estimate the absolute volume of the liver."
COMP	Comparison	"Is the small bowel larger than the stomach?"

See data/tasks.json for the full taxonomy and cal_splits.py (at repo root) for the classifier.

Modalities

Modality	Path in the released archive	Format	Count
2-D	2d_rgb_multiview/<dataset>/<case>/*.png	PNG (RGB)	~12k
3-D	3d_volumes/<dataset>/<case>/*.nii.gz	NIfTI	~4k

Each case has exactly three 2-D views: axial, sagittal, coronal. 3-D volumes keep the native modalities of the source dataset (e.g. BraTS ships t1 / t1ce / t2 / flair / seg).

Reviewer quick-start

# 1. Build the annotations JSONL (adds 5-task labels).
python scripts/build_annotations.py \
    --in  ../dataset/spatialmed/spatial_med_v1.jsonl \
    --out data/annotations.jsonl

# 1b. Drop reviewer-score<1 rows (537 "Not relevant" questions).
python scripts/filter_annotations.py \
    --xlsx      ../spatial_med_scored_final.xlsx \
    --in_jsonl  data/annotations.jsonl \
    --out_jsonl data/annotations.jsonl

# 2. Render one example per task into examples/.
python scripts/visualize_sample.py \
    --annotations data/annotations.jsonl \
    --rgb_root    ../dataset/rgb_multiview \
    --mode        one-per-task \
    --out_dir     examples

# 3. Validate the Croissant metadata.
python scripts/validate_croissant.py croissant.json

License & attribution

• Annotations (data/annotations.jsonl): CC-BY-4.0.
• Source 3-D images / segmentations: keep the upstream licenses (BraTS 2021: CC-BY-SA-4.0; KiTS23: CC-BY-NC-SA-4.0; Medical-Decathlon: CC-BY-SA-4.0; AMOS22: CC-BY-4.0; TotalSegmentator: CC-BY-4.0). Users must agree to each upstream license before downloading. See data/datasets.json for URLs.