Cross-domain Collaborative Learning (CdCL)

Code for our BIBM2023 paper Supervised Domain Adaptation for Recognizing Retinal Diseases from Wide-Field Fundus Images

Get started

Environment

Install packages by pip install -r requirements.txt. This step is suggested to be done in your docker container or virtual environment or things like that.

Datasets

• RFMiD: https://ieee-dataport.org/open-access/retinal-fundus-multi-disease-image-dataset-rfmid
• TOP: https://github.com/DateCazuki/Fundus_Diagnosis

Data split

The data split is provided in Annotations. For RFMiD dataset, we follow the official data split. For TOP dataset, with the absence of official data split, we have randomly divided it into training, validation and test sets with ratio 6:2:2. Furthermore, as mentioned in our paper, we have built a subset of TOP training set with one tenth of its images used.

Dataset	Images	DR	RVO	AMD
RFMiD_train	1,920	376	101	100
RFMiD_val	640	132	31	38
RFMiD_test	640	124	32	31
TOP_train	7,885	2,035	452	227
TOP_train_10percent	788	226	38	19
TOP_val	2,542	611	151	85
TOP_test	2,620	677	175	101

Data preparation

It is recommended to crop the RFMiD images into squares by using

python tools/RFMiD_dataset/square.py

Data Organization

The images are organized as follows:

imagedata/  
├── RFMiD/
│   ├── train_1.jpg
│   ├── train_2.jpg
│   ├── ...
│   ├── val_1.jpg
│   ├── val_2.jpg
│   ├── ...
│   ├── test_1.jpg
│   ├── test_2.jpg
│   └── ...
└── TOP/
    ├── 000000_00.jpg
    ├── 000000_01.jpg
    ├── 000001_00.jpg
    └── ...

The annotations are organized as follows:

Annotations/
├── RFMiD_test/
│   └── Annotations/
│       └── anno.txt
├── RFMiD_train/
├── RFMiD_val/
├── TOP_test/
├── TOP_train/
├── TOP_train_10percent/
└── TOP_val/

Model checkpoint

You can download the model checkpoint in Google drive.

Codes

Prediction and evaluation

By using python checkpoint_eval.py, you can evaluate the downloaded checkpoint. If everything goes well, you may get the following result:

Disease: DR, AP: 0.8478
Disease: RVO, AP: 0.6796
Disease: AMD, AP: 0.5056
Mean AP: 0.6777

Training

Try sh main.sh which combines training, predicting and evaluating together.

Citation

If you find this our work useful, please consider citing:

@inproceedings{wei2023cdcl,
  title={Supervised Domain Adaptation for Recognizing Retinal Diseases from Wide-Field Fundus Images},
  author={Wei, Qijie and Yang, Jingyuan and Wang, Bo and Wang, Jinrui and Zhao, Jianchun and Zhao, Xinyu and Yang, Sheng and Manivannan, Niranchana and Chen, Youxin and Ding, Dayong and Zhou, Jing and Li, Xirong},
  booktitle={2023 IEEE International Conference on Bioinformatics and Biomedicine (BIBM)},
  year={2023}
}

Contact

If you encounter any issue when running the code, please feel free to reach us either by creating a new issue in the GitHub or by emailing

• Qijie Wei ([email protected])