The repository contains code used for the "Red-GAN: Attacking class imbalance via conditioned generation. Yet another medical imaging perspective" accepted at MIDL2020 (http://arxiv.org/abs/2004.10734 ). The paper proposes one more data augmentation protocol based on generative adversarial networks. It is based on the SPADE (that is capable to learn a mapping between label maps and images), and is equiped with two additional components: (a) network conditioning at a global-level information (e.g. acquisition environment or lesion type), (b) a passive player in a form of segmentor introduced into the the adversarial game. The method was validated on two medical datasets: BraTS, ISIC.

To use the three player design (SPADE design equiped with the segmentor) introduced in the paper, one has to first train the segmentor. For that, please step to the unet folder and check out the readme there. Once the segmentor is trained, step to the gan folder.

To use only the two-player design (vanila SPADE architecture), step to the gan folder right away (also, in this case, --segmentator option should be specified as None).

Both GAN and U-Net parts are written in Pytorch. For further installation requirements, please check the corresponding folders.

To cite the repository/paper please use:

@article{qasim2020red, title={Red-GAN: Attacking class imbalance via conditioned generation. Yet another medical imaging perspective}, author={Qasim, Ahmad B and Ezhov, Ivan and Shit, Suprosanna and Schoppe, Oliver and Paetzold, Johannes C and Sekuboyina, Anjany and Kofler, Florian and Lipkova, Jana and Li, Hongwei and Menze, Bjoern}, journal={arXiv preprint arXiv:2004.10734}, year={2020} }