Retinal Fundus Images Classification to Diagnose the Severity of Diabetic Retinopathy using CNN

Abstract

Diabetic Retinopathy is a disease obstructing retinal blood vessels that become the factor of blindness for Diabetes Mellitus patients. If this disease is late to be treated, the patients can have blindness. The examination of this disease is still performed manually by ophthalmologists. As this may need a relatively long time, a system is needed to classify the severity levels of Diabetic Retinopathy. The dataset used in this study was retinal fundus images with the severity levels of Diabetic Retinopathy consisting of 5 classes: NO DR, Mild, Moderate, Severe, and Proliferative DR. In this study, the diagnosis of diabetic retinopathy used Convolutional Neural Network(CNN) with the following architectures: Alexnet, DenseNet121, InceptionV3, Resnet50, VGG16 and Xception to classify the severity levels of Diabetic Retinopathy. This study aimed to compare the accuracy level of 6 architectures in classifying the severity levels of Diabetic Retinopathy. Two trial scenarios were used, namely, first, using a CNN without additional optimization. Second, using a CNN with Stochastic gradient descent, ADAM, and RMSprop optimization. Evaluation results show that model VGG16 with the ADAM optimization produced an accuracy value of 77%. The accuracy value is the highest among the other models on the dataset trialed.