MedIAnomaly: A comparative study of anomaly detection in medical images

This is the repository for our benchmark paper MedIAnomaly: A comparative study of anomaly detection in medical images.

Environment

• Python 3.10
• PyTorch 2.1.2

Data Preparation

We provide the pre-processed seven datasets.

1. Download our pre-processed datasets from: MedIAnomaly-Data
2. Download the ISIC2018_Task3_Training, ISIC2018_Task3_Test, and their ground truth from ISIC2018
3. Unzip the datasets via:

tar -zxvf RSNA.tar.gz
tar -zxvf VinCXR.tar.gz
tar -zxvf BrainTumor.tar.gz
tar -zxvf LAG.tar.gz
tar -zxvf ISIC2018_Task3.tar.gz
tar -zxvf Camelyon16.tar.gz
tar -zxvf BraTS2021.tar.gz

4. Place the MedIAnomaly-Data directory in the user's home directory, i.e., ~/MedIAnomaly-Data/. (Otherwise, you need to modify the data root in your code.)

Finally, the data path should have the following structure:

~/MedIAnomaly-Data
├─RSNA
│  ├─images
│  └─data.json
├─VinCXR
│  ├─images
│  └─data.json
├─BrainTumor
│  ├─images
│  └─data.json
├─LAG
│  ├─images
│  └─data.json
├─ISIC2018_Task3
│  ├─ISIC2018_Task3_Training_Input
│  ├─ISIC2018_Task3_Training_GroundTruth
│  ├─ISIC2018_Task3_Test_Input
│  └─ISIC2018_Task3_Test_GroundTruth
├─Camelyon16
│  ├─train
│  │  ├─good
│  ├─test
│  │  ├─good
│  └─ └─Ungood
├─BraTS2021
│  ├─train
│  ├─test
│  │  ├─normal
│  │  ├─tumor
└─ └─ └─annotation

Train & Evaluate

Reconstruction-based methods

• AE ($\ell_2$, $\ell_1$, SSIM, Perceptual Loss)

• AE-Spatial

• VAE

• Constrained AE

• MemAE

• CeAE

• GANomaly

• AE-U

• DAE

• AE-Grad

• VAE-Grad ($Grad_{ELBO}$, $Grad_{KL}$, $Grad_{rec}$, $Grad_{Combi}$)

Train and evaluate these methods via:

cd reconstruction/;
./train_eval.sh

[Reproduce the results in AE4AD] Train and evaluate AE with different latent size via:

cd reconstruction/;
./train_eval_latent_size.sh

SSL-based methods

one-stage

• CutPaste
• FPI
• PII
• NSA

Train and evaluate these methods via:

cd ssl/one_stage/;
./train_eval.sh

two-stage

• CutPaste
• AnatPaste
• ResNet18-ImageNet

Train and evaluate these methods via:

cd ssl/two_stage/;
./train_eval.sh

Visualization

Acknowledgement

Some datasets and codes in this repository are based on DDAD-ASR, BMAD, NSA, CutPaste, AnatPaste. We thank the original authors for their excellent work.

Contact

If any questions, feel free to raise an issue or contact Yu Cai: yu.cai@connect.ust.hk.

Citation

@article{cai2024medianomaly,
  title={MedIAnomaly: A comparative study of anomaly detection in medical images},
  author={Cai, Yu and Zhang, Weiwen and Chen, Hao and Cheng, Kwang-Ting},
  journal={arXiv preprint arXiv:2404.04518},
  year={2024}
}
@inproceedings{cai2024rethinking,
  title={Rethinking autoencoders for medical anomaly detection from a theoretical perspective},
  author={Cai, Yu and Chen, Hao and Cheng, Kwang-Ting},
  booktitle={International Conference on Medical Image Computing and Computer-Assisted Intervention},
  pages={544--554},
  year={2024},
  organization={Springer}
}