Disentangled Representations for Domain-generalized Cardiac Segmentation [Paper]. In M&Ms Challenge of STACOM 2020.

The repository is created by Xiao Liu, Spyridon Thermos, Agisilaos Chartsias, Alison O'Neil, and Sotirios A. Tsaftaris, as a result of the collaboration between The University of Edinburgh and Canon Medical Systems Europe.

This repository contains the official PyTorch implementation of the Resolution Augmentation (RA) and Factor-based Augmentation (FA) methods proposed in the paper.

• Pytorch 1.5.1 or higher with GPU support
• Python 3.7.2 or higher
• SciPy 1.5.2 or higher
• tqdm
• logging
• CUDA toolkit 10 or newer

In this repository, we train a SDNet [code], [paper] with our proposed Resolution Augmentation and Factor-based Augmentation in a semi-supervised manner.

We propose to use random resampling to augment the original dataset such that the resolutions of all the data are equally distributed in a certain range.

We first pre-train a SDNet model to extract the anatomy and modality factors. Then mix the anatomy and modality factors to generate new images.

To train the model, run the following command:

python train.py -e 50 -bs 4 -g 0

If you find our method useful please cite the following paper:

@article{liu2020disentangled,
  title={Disentangled Representations for Domain-generalized Cardiac Segmentation},
  author={Liu, Xiao and Thermos, Spyridon and Chartsias, Agisilaos and O'Neil, Alison and Tsaftaris, Sotirios A},
  journal={arXiv preprint arXiv:2008.11514},
  year={2020}
}

All scripts are released under the MIT License.