Fundus Image Generation using EyeGAN

Abstract

Deep learning models are widely used in various computer vision fields ranging from classification, segmentation to identification, but these models suffer from the problem of overfitting. Diversifying and balancing the datasets is a solution to the primary problem. Generative Adversarial Networks (GANs) are unsupervised learning image generators which do not require any additional information. GANs generate realistic images and preserve the minute details from the original data. In this paper, a GAN model is proposed for fundus image generation to overcome the problem of labelled data insufficiency faced by researchers in detection and classification of various fundus diseases. The proposed model enriches and balances the studied datasets for improving the eye disease detection systems. EyeGAN is a nine-layered structure based on conditional GAN which generates unbiased, good quality, credible images and outperforms the existing GAN models by achieving the least Fréchet Inception Distance of 226.3. The public fundus datasets MESSIDOR I and MESSIDOR II are expanded by 1600 and 808 synthetic images respectively.