CLAT

Official implementation of CLAT: Concept-based Lesion Aware Transformer for Interpretable Retinal Disease Diagnosis (TMI, 2024)

Usage

Installation

Recommended environment:

• python 3.9.7
• pytorch 2.0.1
• torchvision 0.15.2
• lightning 2.1.0

To install the dependencies, run:

git clone https://github.com/Sorades/CLAT.git
cd CLAT
# if you are using pixi
pixi i && pixi shell
# or conda:
conda env create -f environment && conda activate CLAT
# or pip:
pip install -r requirements.txt

Dataset

• FGADR Dataset
• DDR Dataset

The annotation files are placed at ./data

Preprocess

Refer to src/preprocess for the implementation for image preprocessing.

!!Please note that the lesion label simply indicates the presence or absence of the corresponding lesion. In our experiments, lesion labels for DDR and FGADR are determined by the existence of a mask.

Training and Testing

Modify the settings in ./configs/default.yaml, and then run the commands below to train and test the model:

python src/main.py fit_and_test --config configs/default.yaml --data configs/data/FGADDR.yaml

# test with automatic intervention
python src/main.py exp_int --config configs/default.yaml --data configs/data/FGADDR.yaml

Log

By default, Tensorboard is used to log metrics and heatmaps.

tensorboard --logdir <INPUT_YOUR_LOG_DIR> --bind_all

Run the command and open the url (usually http://<IP>:6006) to access the dashboard.

Citation

@article{wen2024concept,
  title={Concept-based Lesion Aware Transformer for Interpretable Retinal Disease Diagnosis},
  author={Wen, Chi and Ye, Mang and Li, He and Chen, Ting and Xiao, Xuan},
  journal={IEEE Transactions on Medical Imaging},
  year={2024},
  publisher={IEEE},
  doi={10.1109/TMI.2024.3429148}
}