Evidential Concept Embedding Models

Welcome👋! This repository provides the official implementation of our paper Evidential Concept Embedding Models: Towards Reliable Concept Explanations for Skin Disease Diagnosis (evi-CEM) [arXiv], which has been accepted by MICCAI 2024.

💡 TL;DR

Concept Bottleneck Models (CBM) incorporate human-interpretable concepts into decision-making. However, their concept predictions may lack reliability when applied to clinical diagnosis. To address this, we propose an evidential Concept Embedding Model (evi-CEM), which employs evidential learning to model the concept uncertainty. Additionally, we offer to leverage the concept uncertainty to rectify concept misalignments that arise when training CBMs using vision-language models without complete concept supervision.

📦 Get started

Environment Preparing

conda create -n evi-cem python=3.10
conda activate evi-cem
# please modify according to the CUDA version in your server
pip install torch==2.2.0 torchvision==0.17.0 --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt

Dataset Preparing

1. Run the following code to download fitzpatrick17k.csv

wget -P data/meta_data https://raw.githubusercontent.com/mattgroh/fitzpatrick17k/main/fitzpatrick17k.csv

2. Run the following code to download the SkinCon Fitzpatrick17k annotation

wget -O data/meta_data/skincon.csv https://skincon-dataset.github.io/files/annotations_fitzpatrick17k.csv

3. Run python data/raw_data_download.py to download the fitzpatrick17k images. If any image links in the fitzpatrick17k.csv become invalid, the raw images can be downloaded here
4. Run python data/generate_clip_concepts.py to generate soft concept labels with MONET

evi-CEM training

Under complete concept supervision:

python train.py --config configs/default.yaml

Label-efficient training:

python train.py --config configs/label_efficient.yaml

Label-efficient training with concept rectification:

python learn_cavs.py --configs/learn_cavs.yaml
python train.py --configs/train_rectified.yaml

🙋 Feedback and Contact

• ybgao22@m.fudan.edu.cn
• zxh@fudan.edu.cn

🛡️ License

This project is under the Apache-2.0 license. See LICENSE for details.

🙏 Acknowledgement

• We are grateful to cem for publicly sharing their code, which we have drawn upon in our work.
• We appreciate the open-source dermatology dataset Fitzpatrick17k and the SkinCon concept annotations for it.
• We appreciate the effort of MONET to train and release the dermatology vision-language model.

📝 Citation

If our work or code is helpful in your research, please star this repo and cite our paper as follows.

@inproceedings{Gao2024eviCEM,
    author={Yibo Gao, Zheyao Gao, Xin Gao, Yuanye Liu, Bomin Wang, Xiahai Zhuang},
    title={Evidential Concept Embedding Models: Towards Reliable Concept Explanations for Skin Disease Diagnosis},
    booktitle={International Conference on Medical Image Computing and Computer Assisted Intervention},
    year={2024}
}