The Influence of Hyperparameters on GANs Performance for Medical Image Transformation

Generative Adversarial Networks (GANs) have accomplished compelling performance in many fields such as image-to-image transformation, image data generation, translation of image information into text information, and many more. In GANs, two neural networks, a generator and discriminator, compete with each other in an adversarial manner, and due to their robust generating ability, GANs can produce high-quality images for human reference. Thus, GANs are a very promising approach artificial intelligence field. In this work, GANs with optimized hyperparameters selection are proposed to generate skin lesion medical images that are indistinguishable from the real images and analyze the quality of generated images using image quality metrics including Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Fréchet Inception Distance (FID). Additionally, the impact of hyperparameters including learning rate, batch size, latent space size, and number of epochs is analyzed against GANs performance. The experimental results indicate that the learning rate is the most effective hyperparameter for GAN stability and performance. Moreover, latent dimension, batch size, and number of epochs have a relatively small impact when paired with an appropriate learning rate thus good learning rate provides flexibility in optimizing these hyperparameters.