Semi supervised generative adversarial network for automated glaucoma diagnosis with stacked Discriminator models

Abstract

Glaucoma newline optic disc newlineOptic cup newlineGenerative adversarial network newline fundus image newline newlineSemi supervised generative adversarial network for automated glaucoma diagnosis with stacked newlineDiscriminator models newline newlinep.101-107 xv -108p. newline newlineGlaucoma diagnosis is a highly demanding task and fundus newlineimaging is widely accepted as the most effective method of detecting early newlinesigns of glaucoma. The purpose of this study is to broadly examine the fundus newlineimage for glaucoma diagnosis using different Convolution Neural Network newline(CNN) architectures. A semi-supervised Generative Adversarial Network newline(GAN) is designed and also six DL networks from three different newlinearchitectures; Visual Geometry Group (VGG), Residual Network (ResNet) newlineand Inception architectures are employed for fundus image classification. newlineThe discriminator model in the conventional GAN classifies the newlinegenerated samples from the GAN generator model into real or fake. After newlinetraining, the discriminator model is discarded as the GAN model is only newlineinterested to generate examples. However, the discriminator model can be newlineused to develop a pattern recognition model. In this study, the discriminator newlinemodel of conventional GAN is extended into semi-supervised learning for newlinefundus image classification. At first, the fundus images are pre-processed by a newlinesegmentation approach to get the exact region of Interest (ROI). Based on the newlinebrightest region around the Green channel, an initial ROI is extracted and newlinethen spatially weighted fuzzy c-mean clustering approach is employed to get newlinethe exact OD region.The supervised GAN model with k classes is designed at first and newlinethen the semi-supervised GAN model is designed by taking the output prior to newlinethe activation function of the supervised model. CNNs such as VGG16, newlineVGG19, ResNet18, ResNet50, GoogleNet and InceptionV3, pooling layers newlineare employed to down-sample the image whereas in GAN down-sampling is newlineachieved by strided convolutions. Instead of using Rectified Linear activation newlineUnit (ReLU) in CNNs, GAN uses a variation of ReLU called LeakyReLU newlinethat allows a