Evaluation Metrics and Multi-level GAN Approach for Medical Images

Abstract

This article examined methods for using GANs in medicine, their prospects, as well as problems with training generative adversarial networks associated with the increasing use of generated images for training other networks. The analysis of single-layer and multi-layer GANs concluded that although multi-layer GANs perform better statistically, they do not exactly match the distribution of the original dataset and, without medical supervision, such synthetic data should not be used when training new networks. Problems associated with the phenomenon of recursive learning, biased assessments of image realism, and non-optimized structures are considered. Approach is described in context of integrating generative adversarial network models into the OSTIS Technology based hybrid computer systems.