MedVerse

MedVerse, short for Medical Vision Universe, is a large-scale multi-modality medical vision dataset for foundation model pretraining. It contains 4,973,080 2D medical images/slices across 10 medical imaging modalities.

MedVerse aggregates public medical imaging datasets from radiology, ophthalmology, pathology, endoscopy, dermatology, and clinical photography. It is designed for self-supervised pretraining, masked image modeling, and multi-modality medical image representation learning.

Because MedVerse is built from many source datasets with different licenses, this repository does not assign one single license to all images. Each sample keeps its original source license and citation requirement.

Release Policy

This repository provides two types of data:

1. Hosted data: data sources whose licenses allow redistribution on Hugging Face, including Creative Commons or public-release datasets.
2. External data: data sources that cannot be redistributed here. For these datasets, we provide only metadata, official links, and preprocessing scripts. Users must download the original data themselves and then run our scripts.

Modality Composition

Modality	Images / Slices	Source datasets
CT	821,616	AutoPET-CT, ImageCAS, TopBrain-CT
PET	706,726	AutoPET-PET, AutoPET-SUV
MR / MRI	1,753,375	TotalSegmentator-MR, ACDC, MRNet, Duke-Breast-Cancer-MRI, Prostate-MRI-US-Biopsy, TopBrain-MR, BraTS-GLI, LLDMMR
X-ray	468,751	ChestX-ray8, CheXpert, Mammo-Bench, MURA, vzrad2, ADSD, ARCADE, XACD
Ultrasound	68,673	Thyroid Ultrasound Cine-clip, EchoNet-Dynamic, HC18, BUSI, UBPD
Fundus	96,692	MESSIDOR, REFUGE2, APTOS2019, EyePACS
OCT	110,738	ROSE, OCTA500-OCTA, OCTA500-OCT, ZhangLabData-OCT
Pathology	168,999	BBBC041v1, Lacuna Malaria Detection Challenge, Lacuna Malaria Datasets, SegPath
Endoscopy	376,451	HyperKvasir, Cataract-101, PitVis2023, Endoscapes-CVS, PSI-AVA
Clinical photography	401,059	ISIC Challenge 2024 Training

Total: 4,973,080 images/slices.

Hosted Sources

The following sources are included in the hosted subset, subject to their original licenses and attribution requirements.

Dataset	Modality	Anatomy / Domain	Images / Slices	License / Terms	Official Link
AutoPET-CT	CT	Whole body	558,473	CC BY-NC 4.0 if using the FDAT / Grand Challenge release; do not use the TCIA Restricted version for redistribution	AutoPET FDG-PET/CT
AutoPET-PET	PET	Whole body	354,826	CC BY-NC 4.0 if using the FDAT / Grand Challenge release; do not use the TCIA Restricted version for redistribution	AutoPET FDG-PET/CT
AutoPET-SUV	PET / SUV	Whole body	351,900	CC BY-NC 4.0 if using the FDAT / Grand Challenge release; do not use the TCIA Restricted version for redistribution	AutoPET FDG-PET/CT
TotalSegmentator-MR	MR	Whole body	16,059	CC BY-NC-SA 2.0	Zenodo
ChestX-ray8	X-ray	Chest	112,120	NIH public release; preserve original citation and terms	NIH ChestX-ray8
ARCADE	X-ray	Coronary artery	3,000	CC0 1.0	Zenodo
HC18	Ultrasound	Fetal head	1,334	CC BY 4.0	Zenodo
BUSI	Ultrasound	Breast	780	CC BY / open-access source terms	Kaggle mirror
ROSE	OCT / OCTA	Eye	229	Open Zenodo release; preserve source citation	Zenodo
ZhangLabData-OCT	OCT	Eye	109,309	CC BY 4.0	Mendeley Data
BBBC041v1	Pathology	Blood smears	1,328	CC BY-NC-SA 3.0	Broad Bioimage Benchmark Collection
Lacuna Malaria Detection Challenge	Pathology	Blood smears	3,925	CC BY-SA 4.0	Zindi
Lacuna Malaria Datasets	Pathology	Blood smears	5,059	Source-specific open terms; preserve attribution	Lacuna Fund
SegPath	Pathology	Cell / tissue microscopy	158,687	CC BY-NC-SA 4.0	SegPath
HyperKvasir	Endoscopy	Gastrointestinal tract	110,079	CC BY 4.0	HyperKvasir
Cataract-101	Endoscopy / ophthalmic surgery	Eye	42,157	CC BY-NC 4.0	Cataract-101
Endoscapes-CVS	Endoscopy	Gall bladder	55,783	CC BY-NC-SA 4.0	Endoscapes
ISIC Challenge 2024 Training	Clinical photography	Skin	401,059	CC BY-NC 4.0 / ISIC challenge terms; preserve attribution	ISIC Challenge Data

External Sources

The following sources are not hosted in this repository. We provide only metadata, official links, and preprocessing scripts. Users must obtain the data from the original providers.

Dataset	Modality	Anatomy / Domain	Images / Slices	Reason not hosted	Official Link
ImageCAS	CT	Heart	257,496	Redistribution license not clearly verified	ImageCAS GitHub
TopBrain-CT	CT	Brain	5,647	Grand Challenge access terms	TopBrain
ACDC	MR	Heart	1,902	Challenge license; redistribution of dataset or modified versions is not allowed	ACDC
MRNet	MR	Knee	116,624	Stanford Research Use Agreement; no redistribution	MRNet
Duke-Breast-Cancer-MRI	MR	Breast	515,747	TCIA Data Usage Policy; users should download from TCIA	TCIA
Prostate-MRI-US-Biopsy	MR	Prostate	97,180	TCIA Data Usage Policy; users should download from TCIA	TCIA
TopBrain-MR	MR	Brain	4,604	Grand Challenge access terms	TopBrain
BraTS-GLI	MR	Brain	802,341	BraTS / Synapse challenge terms	BraTS / Synapse
LLDMMR	MR	Abdomen	198,918	Redistribution license not clearly verified	LLD-MMRI GitHub
CheXpert	X-ray	Chest	223,648	Stanford Research Use Agreement; no redistribution	CheXpert
Mammo-Bench	X-ray	Breast	71,844	Aggregated benchmark; upstream licenses need source-by-source verification	Mammo-Bench
MURA	X-ray	Bone	40,005	Stanford Research Use Agreement; no redistribution	MURA
vzrad2	X-ray	Tooth	8,188	Source license needs version-level verification	Roboflow Universe
ADSD	X-ray	Coronary artery	8,325	Redistribution license not clearly verified	TODO
XACD	X-ray	Coronary artery	1,621	Redistribution license not clearly verified	TODO
Thyroid Ultrasound Cine-clip	Ultrasound	Thyroid	1,737	Stanford AIMI access terms; no redistribution by default	Stanford AIMI
EchoNet-Dynamic	Ultrasound	Heart	63,867	Stanford Research Use Agreement; redistribution is not allowed	EchoNet-Dynamic
UBPD	Ultrasound	Brachial plexus	955	Redistribution license not clearly verified	TODO
MESSIDOR	Fundus	Eye	1,200	Source-specific research/education terms; redistribution not allowed by default	MESSIDOR
REFUGE2	Fundus	Eye	1,200	Grand Challenge access terms	REFUGE2
APTOS2019	Fundus	Eye	5,590	Kaggle competition terms	APTOS2019
EyePACS	Fundus	Eye	88,702	Kaggle competition terms	EyePACS / Kaggle DR
OCTA500-OCTA	OCT / OCTA	Eye	600	IEEE DataPort terms; redistribution not hosted here	OCTA500
OCTA500-OCT	OCT	Eye	600	IEEE DataPort terms; redistribution not hosted here	OCTA500
PitVis2023	Endoscopy	Pituitary	96,272	CC BY-NC-ND 4.0; processed frames are derivatives, so not hosted	PitVis2023
PSI-AVA	Endoscopy	Prostate	72,160	Redistribution license not clearly verified	PSI-AVA / TAPIR

Repository Structure

MedVerse/
  README.md
  data/
    hosted/
      CT_AutoPET_Wholebody/
      ...
  scripts/
    download_external_sources.py
    preprocess_medverse.py

Loading the Dataset

Load the hosted subset:

from datasets import load_dataset

dataset = load_dataset("YutingHe-list/MedVerse", "hosted", split="train")

Load the full source index:

from datasets import load_dataset

index = load_dataset("YutingHe-list/MedVerse", "index", split="train")

For external datasets, download the original data from the provider and run:

python scripts/preprocess_medverse.py \
  --source CheXpert \
  --input_dir /path/to/original/CheXpert \
  --output_dir /path/to/processed/MedVerse

Preprocessing

For volumetric data such as CT, PET, and MR, volumes are converted into 2D slices. Low-quality, corrupted, abnormally small, near-empty, highly repetitive, and near-duplicate samples are removed.

Typical preprocessing:

Modality	Preprocessing
CT	Clip HU to [-200, 400], then normalize to [0, 1].
PET	Clip by 1st and 99th percentiles, z-score normalize, then min-max normalize to [0, 1].
PET/SUV	Clip SUV values to [0, 20], then normalize to [0, 1].
MR / MRI	Clip by 1st and 99th percentiles, then min-max normalize to [0, 1].
X-ray	Divide pixel values by 255.
Ultrasound	Divide pixel values by 255. For videos, sample one frame every 30 frames.
Fundus	Divide RGB pixel values by 255.
OCT	Divide pixel values by 255.
Pathology	Divide RGB pixel values by 255.
Endoscopy	Divide RGB pixel values by 255. For videos, sample one frame every 30 frames.
Clinical photography	Divide RGB pixel values by 255.

Intended Use

MedVerse is intended for:

• self-supervised medical image pretraining,
• masked image modeling,
• multi-modality medical image representation learning,
• medical vision foundation model research,
• cross-modality transfer learning.

Out-of-Scope Use

MedVerse is not intended for direct clinical diagnosis, treatment planning, patient triage, or deployment in safety-critical systems without clinical validation and regulatory approval.

Users must not attempt to identify patients, recover protected health information, or bypass the access terms of the original source datasets.

License

MedVerse uses source-specific licenses.

The Hugging Face repository license is set to:

Users must follow the original license or access agreement of each source dataset.

Citation

If you use MedVerse, please cite:

@inproceedings{he2026dex,
  title = {Learning Emergent Modular Representations in Multi-modality Medical Vision Foundation Models},
  author = {He, Yuting and You, Chenyu and Li, Shuo},
  booktitle = {Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
  year = {2026},
  publisher = {ACM},
  doi = {10.1145/3770855.3817805}
}

Please also cite the original source datasets used in your experiments.