OmniTumorData

A curated multi-source benchmark for text-prompted 3D tumor segmentation across CT and MRI, accompanying the OmniTumor paper.

11,326 subjects · 12 public cohorts · CT + MRI · 21 sub-region prompts

---

Data access

Imaging data is not redistributed from this page. Several of the constituent cohorts (e.g., AbdomenCT-1K, ULS23, COVID-19 CT) are released under non-redistribution licenses, and a few originate from clinical sites with patient-privacy restrictions on derivative works. We therefore host only the dataset documentation here.

For access to the curated PNG dataset and the consolidated metadata file (dataset_metadata_v2.json):

1. Request via Google Drive — a curated copy with the unified PNG layout, ontology, and split files is staged at: https://drive.google.com/drive/folders/1Kd7NgrMCbzE0vidIj0SHKA7pSfRk_8n0?usp=sharing Access is granted on a per-request basis after a brief usage statement.
2. Or contact the authors directly via the GitHub repository linked at the bottom of this page; we will share download instructions once the request is reviewed.

We additionally provide the original sources below so that users with the appropriate licenses can rebuild the curated layout from scratch.

Composition

#	Cohort	Anatomy	Modality	# Subjects	License	Source
1	BraTS 2023	Brain	MRI (T1c)	2,350	CC-BY-NC-SA 4.0	synapse.org/brats2023
2	MSD Task01 BrainTumour	Brain	MRI	484	CC-BY-SA 4.0	medicaldecathlon.com
3	LGG Segmentation (Buda 2019)	Brain	MRI	110	CC-BY-SA 3.0	kaggle.com/.../lgg-mri-segmentation
4	AbdomenCT-1K (tumor subset)	Abdomen	CT	715	CC-BY-NC-ND 4.0	github.com/JunMa11/AbdomenCT-1K
5	MSD Task03 Liver	Liver	CT	131	CC-BY-SA 4.0	medicaldecathlon.com
6	MSD Task08 (tumor only)	Liver	CT	303	CC-BY-SA 4.0	medicaldecathlon.com
7	ULS23 Part 1	Multi-organ	CT	1,560	CC-BY-NC 4.0	uls23.grand-challenge.org
8	ULS23 Part 2	Multi-organ	CT	3,400	CC-BY-NC 4.0	uls23.grand-challenge.org
9	ULS23 Part 3	Multi-organ	CT	1,416	CC-BY-NC 4.0	uls23.grand-challenge.org
10	LUNA16	Lung	CT	601	CC-BY 3.0	luna16.grand-challenge.org
11	LNDb	Lung	CT	236	CC-BY 4.0	lndb.grand-challenge.org
12	COVID-19 CT	Lung	CT	20	CC-BY-NC-SA 4.0	zenodo.org/record/3757476
	Total			11,326

ULS23 Part 1–3 internally redistribute the following sources, also covered by the ULS23 license: DeepLesion3D, Radboudumc Bone, Radboudumc Pancreas, KiTS21, LIDC-IDRI, LiTS, MSD Lung/Pancreas/Colon, NIH Lymph Node. We load these only via ULS23 to avoid duplicate ingestion.

Ontology

The 12 cohorts are mapped to 21 unique specific-object prompts following the BiomedParse ontology, with sub-region distinctions preserved (e.g., BraTS produces three prompts: necrotic tumor core in brain MRI, peritumoral edema in brain MRI, enhancing tumor in brain MRI). Each canonical prompt is expanded into 7 synonymous variations, yielding 147 unique training strings total.

The complete ontology and augmented prompt pool ship together with the curated dataset behind the access link above.

Splits

Case-level random splits (80% train / 10% val / 10% test, fixed seed = 42) prevent volumetric leakage by assigning all slices from the same 3D volume to the same split. For cohorts with multiple lesions per patient (DeepLesion3D, BraTS, ULS23), a stricter patient-level grouping is additionally enforced so that no patient appears across splits.

Split	# Subjects
Train	9,057
Val	1,130
Test	1,139

The split files (splits/train.txt, splits/val.txt, splits/test.txt) are included in this repository so that any user who reconstructs the layout from raw cohorts can reproduce the exact partitioning.

Reproducing the layout from raw cohorts

After downloading each cohort to RAW_ROOT/<cohort>/, run:

python scripts/convert_all.sh           # 12 cohorts -> unified PNG layout (1024x1024)
python scripts/build_metadata_v2.py     # consolidate ontology + ULS23 routing -> dataset_metadata_v2.json

Preprocessing applied during conversion:

• Resize axial slices to 1024×1024 (Lanczos for image, nearest for mask)
• CT: window center 40 HU, width 400 HU
• MRI: per-volume 1st–99th percentile min-max scaling
• Body mask threshold: −500 HU for CT, 5th-percentile intensity for MRI
• Mask encoding: label k → pixel value 50k; recover via division by 50

Citation

Please cite both OmniTumor and the original source datasets.

@article{omnitumor2025,
  title  = {A Spatial Vision-Language Foundation Model for Universal Volumetric Tumor Segmentation},
  author = {Zhao, Songlin and Sun, Lichao and Liu, Wei},
  year   = {2025},
}

Contact

Songlin Zhao — see github.com/soz223/OmniTumor or open an issue there for data-access requests.