Portable-Skin-Lesion-Diagnosis

This repository provides our codes to train a portable skin lesion diagnosis model based on our proposed knowledge distillation framework. The portable student model is trained by distilling, integrating, and transferring diverse knowledge (D-KD) extracted from a pretrained teacher model. A self-supervised variant (SSD-KD) is also employed to guide the student to capture richer informative knowledge for skin lesion diagnosis.

Environment configuration

Our KD implementations were tested under a linux-centos7-x86_64 system with GPU cards as NVIDIA Tesla V100 (16GB/32GB memory). Please refer to python dependencies in the requirements.txt file. You can install these required dependencies via,

pip3 install -r requirements.txt

Dataset preparation

Please download ISIC 2019, a large-scale dermoscopic image dataset, from this link. After downloading ISIC 2019, please put it in the data/ folder, then preprocess this dataset using the provided script,

bash ./run_preprocess.sh

Train a portable diagnosis model

Please go to the experiment/isic/ directory, download the teacher model ResNet-50 pretrained by us in Google Drive, and put it in the results/ folder. You can also train it from scratch using isic.py which was borrowed from the MetaBlock project.

To train the portable student MobileNet-V2 model, you can choose to use D-KD or SSD-KD:

• For the D-KD method, please perform training following a demo,

python3 isic_d_kd.py with '_kd_method="d_kd"' '_lambd_drkd=1' '_lambd_crkd=1000'

You can also download our trained student model by D-KD from Google Drive.

• For the SSD-KD method, please perform training following a demo,

python3 isic_ssd_kd.py with '_kd_method="ssd_kd"' '_lambd_drkd=1' '_lambd_crkd=1000'

You can also download our trained student model by SSD-KD from Google Drive.

If you encounter possible issues regarding this code, please do not hesitate to contact me.

If you find our code useful in your research, please consider citing our work:

@article{wang2023ssd,
  title={Ssd-kd: A self-supervised diverse knowledge distillation method for lightweight skin lesion classification using dermoscopic images},
  author={Wang, Yongwei and Wang, Yuheng and Cai, Jiayue and Lee, Tim K and Miao, Chunyan and Wang, Z Jane},
  journal={Medical Image Analysis},
  volume={84},
  pages={102693},
  year={2023},
  publisher={Elsevier}
}

Acknowledgements

Many thanks to the following open-sourced repositories that we borrowed which have greatly facilitated our implementations:

• [https://github.com/paaatcha/MetaBlock]

• [https://github.com/paaatcha/raug]

• [https://raw.githubusercontent.com/AberHu/Knowledge-Distillation-Zoo/]

• [https://github.com/xuguodong03/SSKD]