Source Free Domain Adaptation for Medical Image Segmentation with Fourier Style Mining (MIA 2022)

This is the official PyTorch implementation of FSM (Fourier Style Mining) (MIA 2022).

Source Free Domain Adaptation for Medical Image Segmentation with Fourier Style Mining[Paper]

Chen Yang, Xiaoqing Guo, Zhen Chen, Yixuan Yuan

Get Started

Environment

Install dependencies

pip install -r requirements.txt

Datasets Preparation

EndoScene and ETIS-Larib

(1) Download the EndoScene and ETIS-Larib dataset.

(2) Put the data in the corresponding folders. The dataset files are organized as follows.

SFDA-FSM
├── data
│   ├── EndoScene
│   │   ├── images
│   │   │   ├── [case_id].png
│   │   ├── labels
│   │   │   ├── [case_id].png
│   ├── ETIS-Larib
│   │   ├── images
│   │   │   ├── [case_id].png
│   │   ├── labels
│   │   │   ├── [case_id].png

(3) Split dataset into training set and test set as follows.

python preprocess.py

Training

Generation Stage

(1) Generate source-like images with pretrained source model as follows.

python tools/domain_inversion.py 

(2) Visualization of source-like images.

Adaptation Stage

python tools/train_adapt.py 

Testing

python tools/test.py 

Citation

If you find this project useful, please consider citing:

@article{yang2022source,
  title={Source free domain adaptation for medical image segmentation with fourier style mining},
  author={Yang, Chen and Guo, Xiaoqing and Chen, Zhen and Yuan, Yixuan},
  journal={Medical Image Analysis},
  volume={79},
  pages={102457},
  year={2022},
  publisher={Elsevier}
}