Jiancheng Yang, Rui Shi, Bingbing Ni, Bilian Ke

We present MedMNIST, a collection of 10 pre-processed medical open datasets. MedMNIST is standardized to perform classification tasks on lightweight 28 * 28 images, which requires no background knowledge. Covering the primary data modalities in medical image analysis, it is diverse on data scale (from 100 to 100,000) and tasks (binary/multi-class, ordinal regression and multi-label). MedMNIST could be used for educational purpose, rapid prototyping, multi-modal machine learning or AutoML in medical image analysis. Moreover, MedMNIST Classification Decathlon is designed to benchmark AutoML algorithms on all 10 datasets.

For more details, please refer to our paper:

MedMNIST Classification Decathlon: A Lightweight AutoML Benchmark for Medical Image Analysis (arXiv)

• Educational: Our multi-modal data, from multiple open medical image datasets with Creative Commons (CC) Licenses, is easy to use for educational purpose.
• Standardized: Data is pre-processed into same format, which requires no background knowledge for users.
• Diverse: The multi-modal datasets covers diverse data scales (from 100 to 100,000) and tasks (binary/multiclass, ordinal regression and multi-label).
• Lightweight: The small size of 28 × 28 is friendly for rapid prototyping and experimenting multi-modal machine learning and AutoML algorithms.

Please note that this dataset is NOT intended for clinical use.

• medmnist/:
  • dataset.py: dataloaders of medmnist.
  • models.py: ResNet-18 and ResNet-50 models.
  • evaluator.py: evaluate metrics.
• train.py: the training script.
• getting_started.ipynb: Explore the MedMNIST dataset with jupyter notebook.

The code requires only common Python environments for machine learning; Basicially, it was tested with

• Python 3 (Anaconda 3.6.3 specifically)
• PyTorch==0.3.1
• numpy==1.18.5, pandas==0.25.3, scikit-learn==0.22.2, tqdm

Higher (or lower) versions should also work (perhaps with minor modifications).

You could download the dataset(s) via the following free accesses:

• zenodo.org (recommended): You could also use our code to download the datasets from zenodo.org automatically.
• Google Drive
• 百度网盘 (code: gx6i)

The dataset contains ten subsets, and each subset (e.g., pathmnist.npz) is comprised of train_images, train_labels, val_images, val_labels, test_images and test_labels.

• Download Dataset.

• Run the demo code train.py script in terminal.

  First, change directory to where train.py locates. Then, use command python train.py --data_name xxxmnist --input_root input --output_root output --num_epoch 100 --download True to run the experiments, where xxxmnist is subset of our MedMNIST (e.g., pathmnist), input is the path of the data files, output is the folder to save the results, num_epoch is the number of epochs of training, and download is the bool value whether download the dataset.

  To run PathMNIST

  python train.py --data_name pathmnist --input_root input --output_root output --num_epoch 100 --download True
  

If you find this project useful, please cite our paper as:

  Jiancheng Yang, Rui Shi, Bingbing Ni. "MedMNIST Classification Decathlon: A Lightweight AutoML Benchmark for Medical Image Analysis," arXiv preprint arXiv:2010.14925, 2020.

or using bibtex:

 @article{medmnist,
 title={MedMNIST Classification Decathlon: A Lightweight AutoML Benchmark for Medical Image Analysis},
 author={Yang, Jiancheng and Shi, Rui and Ni, Bingbing},
 journal={arXiv preprint arXiv:2010.14925},
 year={2020}
 }

The code is under Apache-2.0 License.