OmniTumorData

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

The paper is the authoritative description of the benchmark. Where this page and the manuscript disagree on naming or numbers, the manuscript is correct and this page will be updated to match.

10,241 annotated cases · 17 public cohorts · CT + MRI · 20 targets · 20 canonical prompts · 400 prompt variants

Data access

Imaging data is not redistributed publicly 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. Access is therefore gated.

To request access:

1. Click "Acknowledge and request access" above and submit the usage form. Requests are reviewed manually.
2. Once approved, the curated PNG layout, consolidated metadata (dataset_metadata_v2.json), and the pretrained OmniTumor checkpoint (omnitumor_v2_final.pth) become downloadable from this repository.
3. Alternatively, contact the authors via the GitHub repository linked at the bottom of this page; we will share download instructions once the request is reviewed.

If you have already secured licenses to the original sources, the Composition table below lists each cohort so that the curated layout can be rebuilt locally.

Summary

Annotated cases	10,241
Axial slices	approximately 2.55 million
Modalities	CT, MRI
Anatomical systems	8 (Brain/CNS, Thoracic, Hepatic, Renal, Pancreatic, Colorectal, Lymphatic, Musculoskeletal)
Segmentation targets	20 tumor and lesion types
Canonical prompts	20
Prompt variants	400 (1 canonical + 19 paraphrases per prompt)

Cohort composition

#	Cohort	Anatomy	Modality	Cases	Source
1	BraTS 2023	Brain	MRI (T1c)	2,350	synapse.org/syn51156910
2	LGG Segmentation (Buda 2019)	Brain	MRI	110	kaggle.com/datasets/mateuszbuda/lgg-mri-segmentation
3	AbdomenCT-1K	Abdomen	CT	715	github.com/JunMa11/AbdomenCT-1K
4	MSD Liver (Task03)	Liver	CT	131	medicaldecathlon.com
5	MSD Hepatic Vessel (Task08)	Liver	CT	303	medicaldecathlon.com
6	LiTS*	Liver	CT	330	uls23.grand-challenge.org
7	KiTS21*	Kidney	CT	877	uls23.grand-challenge.org
8	MSD Pancreas*	Pancreas	CT	617	uls23.grand-challenge.org
9	Radboudumc Pancreas*	Pancreas	CT	183	uls23.grand-challenge.org
10	MSD Colon*	Colon	CT	277	uls23.grand-challenge.org
11	Radboudumc Bone*	Bone	CT	233	uls23.grand-challenge.org
12	NIH-LN*	Lymph Node	CT	384	uls23.grand-challenge.org
13	DeepLesion3D*	Multi-organ	CT	1,144	uls23.grand-challenge.org
14	LIDC-IDRI*	Lung	CT	2,193	uls23.grand-challenge.org
15	MSD Lung*	Lung	CT	138	uls23.grand-challenge.org
16	LNDb	Lung	CT	236	lndb.grand-challenge.org
17	COVID-19 CT	Lung	CT	20	zenodo.org/record/3757476
	Total			10,241

* Loaded via ULS23 re-distribution. The ULS23 license also covers the redistributed sources, and ingestion is performed exclusively through the ULS23 packaging to avoid duplicate cases.

Text prompts

Each label is mapped via a three-level ontology (Category > Meta-object > Specific object) to a canonical prompt of the form [OBJECT TYPE] in [ANATOMIC SITE] [MODALITY]. Sub-region distinctions are preserved end-to-end: BraTS produces three prompts (necrotic core, peritumoral edema, enhancing tumor); abdominal vs. mediastinal lymph nodes are separate prompts; and so on.

Each of the 20 canonical prompts is expanded into 20 variants (1 canonical + 19 paraphrases), yielding 400 unique prompt strings. One variant is sampled per (image, mask) pair per epoch during training.

Pretrained checkpoint

The OmniTumor checkpoint trained on this dataset is released as omnitumor_v2_final.pth (1.8 GB) in this repository, accessible after the gated-access request above is approved. See the OmniTumor codebase at github.com/soz223/OmniTumor for the training recipe.

Repository contents

File	Description
ontology.json	Three-level target ontology and the 20 canonical prompts
per_case_index.csv	Every case identifier with its source dataset, patient and split
splits/patient_level_splits.csv	Patient-level 80/10/10 train/val/test assignment
dataset_metadata_v2.json	Consolidated metadata consumed by the training pipeline. Not shipped here; regenerate locally with scripts/build_metadata_v2.py
scripts/	Conversion and metadata-building scripts
omnitumor_v2_final.pth	Pretrained OmniTumor checkpoint (1.8 GB)

The ontology, per-case index and splits will be released after publication.

Reproducing the local layout from your own downloads

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

# Convert all cohorts to the unified PNG layout
bash scripts/convert_all.sh

# Build dataset_metadata_v2.json (ULS23 routing + augmented prompts)
python scripts/build_metadata_v2.py

License

Each constituent dataset retains its original license. Refer to each cohort's source page for per-cohort license terms; access to the ULS23-redistributed cohorts is additionally governed by the ULS23 license. The pretrained checkpoint is released under CC BY-NC 4.0, non-commercial research use only, the same terms as the dataset compilation.

Contact

Songlin Zhao, Zhao.Songlin@mayo.edu. Code at github.com/soz223/OmniTumor.