Source-Free Domain Adaptive Fundus Image Segmentation with Denoised Pseudo-Labeling

Pytorch implementation of our source-free unsupervised domain adaptation method with denoised pseudo-labeling.

Paper

Source-Free Domain Adaptive Fundus Image Segmentation with Denoised Pseudo-Labeling MICCAI 2021

Installation

• Install Pytorch 0.4.1 and CUDA 9.0 (Note that the results reported in the paper are obtained by running the code on this Pytorch version. As raised by the issue, using higher version of Pytorch may seem to have a performance decrease on optic cup segmentation.)
• Clone this repo

git clone https://github.com/cchen-cc/SFDA-DPL
cd SFDA-DPL

Train

• Download datasets from here.
• Download source domain model from here or specify the data path in ./train_source.py and then train ./train_source.py.
• Save source domain model into folder ./logs/source.
• Download generated pseudo labels from here or specify the model path and data path in ./generate_pseudo.py and then train ./generate_pseudo.py.
• Save generated pseudo labels into folder ./generate_pseudo.
• Run ./train_target.py to start the target domain training process.

Acknowledgement

The code for source domain training is modified from BEAL.

Note

• Contact: Cheng Chen (chencheng236@gmail.com)