FairDisCo: Fairer AI in Dermatology via Disentanglement Contrastive Learning

Diagram of 3 skin disease classifiers: (a) Baseline; (b) Attribute-aware method; (c) Our proposed disentanglement network with contrastive learning (FairDisCo).

This is a PyTorch implementation for FairDisCo: Fairer AI in Dermatology via Disentanglement Contrastive Learning, ECCV ISIC Workshop 2022.

• Requirements
• Data
• Training
• Evaluation
• Acknowledgements

If you use this code in your research, please consider citing:

@inproceedings{du2023fairdisco,
  title={{FairDisCo}: Fairer {ai} in dermatology via disentanglement contrastive learning},
  author={Du, Siyi and Hers, Ben and Bayasi, Nourhan and Hamarneh, Ghassan and Garbi, Rafeef},
  booktitle={Computer Vision--ECCV 2022 Workshops: Tel Aviv, Israel, October 23--27, 2022, Proceedings, Part IV},
  pages={185--202},
  year={2023},
  organization={Springer}
}

Requirements

This code is implemented using Python 3.8.1, PyTorch v1.8.0, CUDA 11.1 and CuDNN 7.

conda create -n skinlesion python=3.8
conda activate skinlesion  # activate the environment and install all dependencies
cd FairDisCo/
conda install pytorch==1.11.0 torchvision==0.12.0 torchaudio==0.11.0 cudatoolkit=11.3 -c pytorch
# or go to https://pytorch.org/get-started/previous-versions/ to find a right command to install pytorch
pip install -r requirements.txt

Data

1. Download Fitzpatrick17k dataset by filling the form here

2. Download Diverse Dermatology Images (DDI) from here

3. Use data_play_fitz.ipynb and data_play_ddi.ipynb to remove unknown skin types, encode disease labels, and generate the weights of reweighting and resampling methods.

Training

We have 6 models: baseline (BASE), attribute-aware (ATRB), resampling (RESM), reweighting (REWT), FairDisCo, FairDisCo without contrastive loss. Train one of those models as

python -u train_BASE.py 20 full fitzpatrick BASE
# or
python -u train_BASE.py 15 full ddi BASE

Evaluation

Use multi_evaluate.ipynb

Acknowledgements

• This code began with mattgroh/fitzpatrick17k. We thank the developers for building the Fitzpatrick17k dataset and providing the baseline.