Generative Adversarial Networks for the Synthesis of Chest X-ray Images

Abstract

One way to diagnose COVID-19 is to use the Polymerase Chain Reaction (PCR) test. However, this test is rather invasive. An alternative would be to use chest images of the patients to diagnose if the patient has COVID-19. These chest X-ray images have to be manually annotated by a medical professional such as a radiologist, and due to privacy concerns, getting access to readily available and annotated COVID-19 chest X-ray images is difficult. In order to train a deep learning model to perform image classification tasks, it is prudent to train the deep-learning model on a large enough dataset to avoid the problem of overfitting. In this paper, we explore using Generative Adversarial Networks (GANs) as a form of data augmentation technique to enlarge the training data for deep learning models. We first explored how the synthetic data generated by GANs are affected by the training size. Following which, we compared the performance of the two different GAN architectures, namely the Deep Convolutional Generative Adversarial Networks (DCGAN) and the Wasserstein Generative Adversarial Networks with Gradient Penalty (WGAN-GP). We successfully used GANs to generate synthetic COVID-19 chest X-ray images with a Fréchet Inception Distance (FID) score that was below 2.