Detection of Retinal Disorders in Optical Coherence Tomography using Deep Learning

Abstract

We address a paramount problem domain in the analysis of Optical Coherence Tomography (OCT) images: the detection of retinal disorder in retinal OCT images. The objective is to identify the existence of normal macula and each of the three types of disorders, specifically the Choroidal Neovascularization (CNV), the Diabetic Macular Edema (DME) and the Drusen, in the OCT images. Here, we formulate a detection system based on a Deep-Learning methodology for the screening of patients with frequent blinding retinal diseases, which are treatable if detected in time. For achieving the goal, we used Densely Connected Convolution Neural Network (DenseNet). Various experiments were conducted to determine a justifiable DenseNet network with optimal parameters. We achieved an accuracy of 98.00% on the training set, 97.20% on the validation set and 97.65% on the test set, with a precision of 0.9640, recall/sensitivity of 0.9557 and specificity of 0.9915. Also, the Area under curve (AUC) of the model evaluates to an average of 0.9775 based on One-vs-All approach.