Rethinking Transfer Learning for Medical Image Classification

This is the offical implemention of paper Rethinking Transfer Learning for Medical Image Classification [BMVC'23 oral]

Overview | example of TTL on resent50 :-------------------------:|:-------------------------: <img src="figures/overview.png" width="500" /> | <img src="figures/svcca.png" width="450" />

Usage

Setup

pip

Requires python>=3.10+

See the requirements.txt for environment configuration

pip install -r requirements.txt

Dataset

<figure><img src="figures/examples.png"></figure>

BIMCV

• Please download our pre-processed datasets TBA, put under data/ directory and perform following commands:

  cd ./data
  unzip digit_dataset.zip
  

<!-- ### **MIDOG 2022** - Please download the dataset [here](https://midog2022.grand-challenge.org/), put under `data/MIDOG/` -->

HAM10000

• Please download the dataset here, put under data/HAM10000/

PENet Dataset

• Please download the dataset here, put under data/PENet/ directory and perform following commands:

Train

2D experiment (BIMCV & HAM1000)

block-wise TTL

Please using following commands to train a model with federated learning strategy.

• --model specify model archicture: resnet50 | densenet201
• --pretrained specify source domain: imagenet | chexpert
• --dataset specify target dataset: BIMCV | HAM10000
• --trunc specify truncation point: {-1, 1, 2, 3}

python main.py --model resnet50 --bs 64 --data_parallel --num_workers 12 --max_epoch 200 --pretrained imagenet --dataset BIMCV --trunc -1 --exp 1 --sub 100

layer-wise TTL

--trunc specify truncation point: {-1, 1, 2, ..., 16}

python main.py --model layerttl_resnet50 --bs 64 --data_parallel --num_workers 12 --max_epoch 200 --pretrained imagenet --dataset BIMCV --trunc -1 --exp 1 --sub 100

Test

block-wise TTL

python main.py --model resnet50 --bs 64 --data_parallel --num_workers 12 --max_epoch 200 --pretrained imagenet --dataset BIMCV --trunc -1 --exp 1 --sub 100

layer-wise TTL

python main.py --model layerttl_resnet50 --bs 64 --data_parallel --num_workers 12 --max_epoch 200 --pretrained imagenet --dataset BIMCV --trunc -1 --exp 1 --sub 100

<figure><img src="figures/BIMVC.png"></figure>

If you use this code or dataset in you research, please consider citing our paper with the following Bibtex code:

@article{peng2022rethinking,
  title={Rethinking Transfer Learning for Medical Image Classification},
  author={Peng, Le and Liang, Hengyue and Luo, Gaoxiang and Li, Taihui and Sun, Ju},
  journal={medRxiv},
  pages={2022--11},
  year={2022},
  publisher={Cold Spring Harbor Laboratory Press}
}