Conditional Patch-based Generative Adversarial Network for Retinal Vessel Segmentation

Abstract

In various medical applications, retinal vessel segmentation is a key step to diagnose diseases in fundoscopic images. This research propose a deep learning base patch-based Generative Adversarial Network (GAN) to map complex boundaries of retinal vessels to a probability masks. The proposed conditional patch GAN which utilizes the generator and patch discriminator conditioned on the sampled ground truth distribution to segment out the thick and thin retinal vessels and robust to the inadequate brightness and contrast regions and distinguish the vessels and background. Moreover, the generator network is combined with a patch based discriminator where conditional distribution is imposed on both the generator and discriminator to learn the features from conditional data. Finally, model generates a binary map of vessel segmentation and achieves comparable accuracy on benchmark datasets.