Our significant extension version of DDAD, named DDAD-ASR, will be soon available at https://github.com/caiyu6666/DDAD-ASR
This is the PyTorch implementation of our paper:
Dual-Distribution Discrepancy for Anomaly Detection in Chest X-Rays
Yu Cai, Hao Chen, Xin Yang, Yu Zhou, Kwang-Ting Cheng.
International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI), 2022, Early Accept.
- Python 3.6
- Pytorch 1.7.0
- tensorboard 2.5.0
- pillow 6.1.0
- pydicom 2.3.0 (for data preprocessing)
- Download the training dataset of RSNA Pneumonia Detection Challenge and VinBigData Chest X-ray Abnormalities Detection challenge. Note that we only use their training set as labels of testing set are not available.
- Use
data/preprocess.pyto preprocess the two datasets respectively. The output files should be*.png. - Move the repartition files
rsna_data.jsonandvin_data.jsonto corresponding data roots and rename todata.json.
The final structure of datasets should be as following:
├─DATA_PATH
│ ├─rsna-pneumonia-detection-challenge # data root of RSNA dataset
│ │ ├─train_png_512 # preprocessed images of rsna dataset
│ │ │ ├─xxx.png
│ │ │ ├─ ......
│ │ ├─data.json # repartition file of rsna dataset (renamed from "rsna_data.json")
│ ├─VinCXR # data root of VinBigData dataset
│ │ ├─train_png_512 # preprocessed images of VinBigData dataset
│ │ │ ├─xxx.png
│ │ │ ├─ ......
│ │ ├─data.json # repartition file of VinBigData dataset (renamed from "vin_data.json")The data.json is a dictionary that storing the data repartition information:
{
"train": {
"0": ["*.png", ], // The known normal images for one-class training
"unlabeled": {
"0": ["*.png", ], // normal images used to build the unlabeled dataset
"1": ["*.png", ] // abnormal images used to build the unlabeled dataset
}
},
"test": {
"0": ["*.png", ], // normal testing images
"1": ["*.png", ] // abnormal testing images
}
}Train the reconstruction network for module A. (Repeat K=3 times to obtain an ensemble of K networks.)
python main.py --config cfgs/RSNA_AE.yaml --mode a
Train the reconstruction network for module B. (Repeat K=3 times to obtain an ensemble of K networks.)
python main.py --config cfgs/RSNA_AE.yaml --mode b
Evaluation
python main.py --config cfgs/RSNA_AE.yaml --mode eval
Each single reconstruction network can also be tested for comparison.
python main.py --config cfgs/RSNA_AE.yaml --mode test
In total, you can directly excute ./train_eval.sh to train and evaluate the DDAD (AE) on RSNA dataset.
Different configuration files can be used for experiments on different datasets and different basic networks.
Experiments on RSNA dataset. AE is the basic network.
If this work is helpful for you, please cite our paper:
@article{cai2022dual,
title={Dual-Distribution Discrepancy for Anomaly Detection in Chest X-Rays},
author={Cai, Yu and Chen, Hao and Yang, Xin and Zhou, Yu and Cheng, Kwang-Ting},
journal={arXiv preprint arXiv:2206.03935},
year={2022}
}
If you have any question, feel free to email Yu Cai.