This is the pytorch implementation of the paper submitted to MICCAI 2025:

Anonymous, "Enhancing Fundus Self-supervised Learning via CLIP-informed Lesion Attribute Learning", under review, 2025.

The datasets used in this work are listed as follows:

Pretraining:

• EyePACS [homepage].

Evalutation for classification:

• Messidor-2 [images] [labels].
• IChallenge-AMD [homepage]
• Retinal [homepage]
• ODIR-5k [homepage]
• MPOS [homepage]

To install the dependencies, download the repository and run:

conda create -n cila python=3.8.0
conda activate cila
pip install .

1. Organize the EyePACS in the following format:

├── dataset
    ├── class1
        ├── image1.jpg
        ├── image2.jpg
        ├── ...
    ├── class2
        ├── image3.jpg
        ├── image4.jpg
        ├── ...
    ├── class3
    ├── ...

Note: Image labels are not used in the pretraining stage.

2. Modify the train_path in /scripts/pretrain/eyepacs/[SSL method].yaml to point to your dataset. You can also adjust the training device by modifying devices in this YAML file.

3. Run pre-training:

python main_pretrain.py --config-path scripts/pretrain/eyepacs/ --config-name [SSL method].yaml

1. Ensure each dataset follows this structure:

├── dataset
    ├── train
        ├── class1
            ├── image1.jpg
            ├── image2.jpg
            ├── ...
        ├── class2
            ├── image3.jpg
            ├── image4.jpg
            ├── ...
        ├── class3
        ├── ...
    ├── val
    ├── test

2. Run fine-tuning on the downstream dataset with:

python finetuning.py --config-path ./scripts/finetune/fundus/ --config-name finetuning.yaml \
    ++finetuning.dataset={} ++finetuning.data_path={} ++finetuning.num_classes={} \
    ++finetuning.save_path={} ++finetuning.checkpoint={} ++finetuning.learning_rate={}

You can either update the arguments in {} directly or modify the configuration file ./scripts/finetune/fundus/finetuning.yaml.

3. Linear evaluation on downstream dataset by:

python finetuning.py --config-path ./scripts/finetune/fundus/ --config-name finetuning.yaml \
    ++finetuning.dataset={} ++finetuning.data_path={} ++finetuning.num_classes={} \
    ++finetuning.save_path={} ++finetuning.checkpoint={} ++finetuning.learning_rate={} \
    ++finetuning.linear=true

This repository is based on solo-learn, a wonderful self-supervised learning framework for PyTorch.