Optic Disc Segmentation using Nonconvex Rank Approximation from Retinal Fundus Images

Abstract

An essential step in the diagnosis of glaucoma is the accurate detection of the optic disc (OD). The increasing demand requires an effective and noninvasive retinal imaging tools to manage the growing retinal abnormality. A promising strategy in this area is the use of portable fundus cameras and handheld mobile cameras that are connected to a smartphone. In contrast to retinal images taken using traditional equipment, the fundus camera and smartphone images are frequently of poor quality and difficult to segment the optic disc due to nonuniform illumination and a small curved surface of the retina. To alleviate the segmentation problem, this paper proposed a non-convex rank approximation technique for efficient segmentation of optic disc. Adaboost, KNN, Randomforest and SVM are the four machine learning classifiers used after OD segmentation in order to compare the results of the proposed method and for better accuracy. This method achieved an accuracy of 89.25% using SVM classifier for REFUGE dataset.