An empirical evaluation of generative adversarial nets in synthesizing X-ray chest images

In the last decade, Generative Adversarial Nets (GAN) have become a subject of growing interest in multiple research fields. In this paper, we focus on applications in the medical field by attempting to generate realistic X-ray chest images. A heuristic approach is adopted to perform an extensive evaluation of different architecture configurations and optimization algorithms and we propose an optimal configuration of the baseline Deep Convolutional GAN (DCGAN) based on empirical findings. Additionally, we highlight the technical limitations of GAN and provide an analysis of the high memory requirements, which we reduce by a range of 1.2-7 percent by removing unnecessary layers. Finally, we verify that the loss of the discriminator can be used as an assessment metric.