Abstract
Glaucoma is a major eye disease, leading to vision loss in the absence ofproper medical treatment. Current diagnosis of glaucoma is performed byophthalmologists who are often analyzing several types of medical imagesgenerated by different types of medical equipment. Capturing and analyzingthese medical images is labor-intensive and expensive. In this paper, wepresent a novel computational approach towards glaucoma diagnosis andlocalization, only making use of eye fundus images that are analyzed bystate-of-the-art deep learning techniques. Specifically, our approach leveragesConvolutional Neural Networks (CNNs) and Gradient-weighted Class ActivationMapping (Grad-CAM) for glaucoma diagnosis and localization, respectively.Quantitative and qualitative results, as obtained for a small-sized datasetwith no segmentation ground truth, demonstrate that the proposed approach ispromising, for instance achieving an accuracy of 0.91$\pm0.02$ and an ROC-AUCscore of 0.94 for the diagnosis task. Furthermore, we present a publiclyavailable prototype web application that integrates our predictive model, withthe goal of making effective glaucoma diagnosis available to a wide audience.