MedMNIST+: 18 standardized biomedical image datasets
Yang, Shi, Wei, et al. (Shanghai Jiao Tong University)
Eighteen medical imaging classification tasks shrunk to MNIST-size, in several resolutions.
A collection of 12 two-dimensional and 6 three-dimensional biomedical image datasets, preprocessed into small standardized images with classification labels. Sizes range from about 100 to 100,000 samples, with binary, multiclass, ordinal, and multi-label tasks. The MedMNIST+ release adds larger image size options.
Access
- Cost
- FreeFree download from Zenodo; no account needed.
- License
- Open licenseCC BY 4.0 (source data sets may carry their own licenses)
- Sign-up
- No sign-up
- How you get it
- Bulk downloadNPZ
What's in it
- One row is
- One image (2D or 3D) with a label
- Covers
- Source images from multiple hospitals and public sources
- Years
- 2021-2024 releases
- Updated
- Irregular
- Size
- 18 datasets; from hundreds to about 100,000 samples each
Quality
Standardized, easy to use, with a published Scientific Data paper; designed for benchmarking, not clinical use.
- Downsampled images lose clinically relevant detail
- Underlying source datasets have their own terms
Connects to
- MNIST: Handwriting benchmark whose format and spirit it copies
- NCT-CRC-HE-100K: Source of the pathology task
- CheXpert: Large chest X-ray set for full-resolution work
Questions it could answer
- Do model rankings on MedMNIST agree with rankings on full-resolution data?
- How does accuracy change with image size?