Synthesizing Retinal Images using End-To-End VAEs-GAN Pipeline-Based Sharpening and Varying Layer

Abstract

This study attempts to synthesize a realistic-looking fundus image from a morphologically changed vessel structure using the newly proposed sharpening and varying vessels technique (SVV). This technique sharpens the reconstructed vessels and introduces variation to their structure to generate multiple images from a single input mask. This helps to reduce the reliance on expensive and scarce annotated medical data. The study also aims to overcome the limitations of current methods, such as unrealistic optic disc boundaries, extreme vessel tortuosity, and missed optic discs. This is mainly due to the fact that existing models penalize their weights based on the difference between real and synthetic images using only a single mask. Therefore, their emphasis is on generating the input mask while disregarding other important fundoscopic features. Inspired by the recent progress in Generative Adversarial Nets (GANs) and Variational Autoencoder (VAE), the proposed approach was able to preserve the geometrical shape of critical fundus characteristics. Visual and quantitative results indicate that the produced images are considerably distinct from the ones used for training. However, they also exhibit anatomical coherence and a reasonable level of visual. The data utilized in this study and the programming code necessary to recreate the experiment can be accessed at https://github.com/AliSaeed86/SVV_GAN.