Scalable Biomedical Image Synthesis with GAN

Despite the fast-paced progress in imaging techniques made possible by ubiquitous applications of convolutional neural networks, biomedical imaging has yet to benefit from the full potential of deep learning. An unresolved bottleneck is the lack of training set data. Some experimentally obtained data are kept and preserved by individual research groups where they were produced, out of the reach of the public; more often, high cost and rare occurrences simply mean not enough such images have been made. We propose to develop deep learning based workflow to overcome this barrier. Leveraging the largest radiology data (chest X-Ray) recently published by the NIH, we train a generative adversarial network (GAN) and use it to produce photorealistic images that retain pathological quality. We also explore porting our models to a range of supercomputing platforms and systems that we have access to, including XSEDE, NERSC, OLCF, Blue Waters, NIH Biowulf etc., to investigate and compare their performance. In addition to the obvious benefits of biomedical research, our work will help understand how current supercomputing infrastructure embraces machine learning demands. Our code and enhanced data set are available through GitHub/Binder.