AsynDGAN in pyTorch

Synthetic Learning: Learn From Distributed Asynchronized Discriminator GAN Without Sharing Medical Image Data

Qi Chang^1*, Hui Qu^1*, Yikai Zhang^1*, Mert Sabuncu², Chao Chen³, Tong Zhang⁴ and Dimitris Metaxas¹ {qc58,hq43,yz422,dnm}@cs.rutgers.edu , msabuncu@cornell.edu, chao.chen.cchen@gmail.com, tongzhang@tongzhang-ml.org

Abstract

In this paper, we propose a data privacy-preserving and communication efficient distributed GAN learning frame- work named Distributed Asynchronized Discriminator GAN (AsynDGAN). Our proposed framework aims to train a cen- tral generator learns from distributed discriminator, and use the generated synthetic image solely to train the seg- mentation model. We validate the proposed framework on the application of health entities learning problem which is known to be privacy sensitive. Our experiments show that our approach: 1) could learn the real image’s distri- bution from multiple datasets without sharing the patient’s raw data. 2) more efficient and requires lower bandwidth than other distributed deep learning methods. 3) achieve higher performance compared to the model trained by one real dataset, and almost the same performance compared to the model trained by all real datasets. 4) proofs that the generator could learn the distributed distribution in an all important fashion thus has an unbiasedness guarantee.

In CVPR 2020 [bibtex]

@inproceedings{chang2020synthetic,
  title={Synthetic Learning: Learn From Distributed Asynchronized Discriminator GAN Without Sharing Medical Image Data},
  author={Chang, Qi and Qu, Hui and Zhang, Yikai and Sabuncu, Mert and Chen, Chao and Zhang, Tong and Metaxas, Dimitris N},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={13856--13866},
  year={2020}
}

Requirements:

pytorch 1.1+
python 3.6.8
visdom 0.1.8.8
pillow 5.4.1

Dataset build:

Download the BraTS 2018 dataset: https://www.med.upenn.edu/sbia/brats2018/data.html
Split dataset into training/validation/testing as the following script: https://github.com/tommy-qichang/AsynDGAN/files/5703602/brats_dataset_build.py.txt
Split the training dataset into 10 subsets based on the segmentation volume size.

Acknowledgments