MedAD-38K

A multi-source benchmark for interpretable medical anomaly detection and diagnostic visual question answering.

Paper · Code · Dataset · Model

Release status: This is a project information page. No dataset images or annotations have been published here yet. The release files are being validated, and access requests will be reviewed manually when the files are ready. We plan open release after paper acceptance, subject to the original sources' licenses and redistribution terms. The linked arXiv record may currently show an earlier manuscript version.

What the benchmark covers

The revised manuscript reports 38,422 medical images and 130,146 VQA instances assembled from 16 public source datasets, across 10 imaging modalities and 10 anatomical regions. These are paper statistics, not the contents of this currently empty repository. Questions cover anatomy identification, modality classification, anomaly detection, pathology characterization, and lesion localization. The paper also describes quality-controlled diagnostic chain-of-thought annotations.

Imaging modality	Source datasets	Example anatomy
MRI / contrast-enhanced MRI	Br35H, BraTS2021, PediDemi, ATLAS	Brain, liver
CT / X-ray	MosMed, Chest	Lung, chest
OCT / fundus imaging	RESC, IDRiD	Retina
Dermoscopy	ISIC2018, PH2	Skin
Ultrasound	BUSI, DDTI	Breast, thyroid
Endoscopy	CVC-ClinicDB, Kvasir-SEG	Gastrointestinal tract
Microscopy	BACH, MLS	Breast, lymph node

The paper defines image-disjoint training and test splits and evaluates generalization on five additional source-disjoint external datasets. Please consult the paper and supplement for source-level counts, annotation procedures, and the evaluation protocol.

Access and use

The dataset files are not yet available. This repository has manual access review enabled, but an approved request cannot provide data until validated files are uploaded. We will update this card with the final file manifest, counts, source permissions, and loading instructions when release validation is complete.

MedAD-38K is intended for research into medical-image VQA, anomaly detection, and reasoning evaluation. It is not a clinical diagnostic dataset or a substitute for independent validation in patient care.

Project resources

Resource	Link	Status
Paper	arXiv:2602.01081	Public record; revised version may be pending.
Code	GitHub: zhtstar/MedAD-R1	Implementation in preparation.
Dataset	Hugging Face: zhtstar/MedAD-38K	Information page only; no data files yet.
Model	Hugging Face: zhtstar/MedAD-R1	Manual access review.

Citation

Please cite the arXiv paper. Citation details will be updated when the revised version is public.