Evolution of Generative Adversarial Networks Using PSO for Synthesis of COVID-19 Chest X-ray Images

The use of biomedical images for the training of various Deep Learning (DL) systems oriented to health has reported a competitive performance. However, DL needs a large number of images for a correct generalization and, particularly in the case of biomedical images, these can be scarce. Generative Adversarial Networks (GANs) as Data Augmenting tools have reaped significant results to improve performance in tasks that involve the use of this kind of image. However, the architectural design of these generative models in the biomedical image area has been usually relegated to the expertise of researchers. Moreover, GANs are affected by training instability that may lead to poor quality results. This paper presents a neuroevolution algorithm based on Particle Swarm Optimization for the design and training of GANs for the generation of biomedical Chest X-Ray (CXR) images of pneumonia caused by COVID-19. The proposed approach allows having a swarm of GANs topologies, where each one of them grows progressively while being trained at the same time. The fitness value is based on the Frechet Inception Distance (FID). The proposed algorithm is able to obtain better FID results compared to handcrafted GANs for the synthesis of CXR images.