Diabetic Retinopathy Severity Level Classification Based on Fundus Image Using Convolutional Neural Network (CNN)

Abstract

Diabetic retinopathy is an eye disease and is a complication of diabetes mellitus. The longer a person suffers from diabetes mellitus, the more likely they are to experience diabetic retinopathy. Diabetic retinopathy is divided into two types, namely Non-Proliferative Diabetic Retinopathy (NPDR) with 4 phases (normal, mild, moderate and severe) and Pre-proliferative Diabetic Retinopathy (PDR). To classify the severity of this disease requires an expert doctor and takes a long time. This study applies the Convolutional Neural Network (CNN) method to fundus image input to classify the severity of diabetic retinopathy, namely mild, moderate, severe, or regular. The fundus image dataset for training and testing was taken from the APTOS 2019 dataset. The pre-processing stage of the fundus image includes: resizing, Contrast Limited Adaptive Histogram Equalization (CLAHE), and gaussian filtering. After that, classification is carried out using the CNN Model, consisting of a convolution layer, a pooling layer, a dropout layer, and a fully connected layer. The results of the CNN model implementation show a classification accuracy of 75% in the training process and 73% in the model validation process. Meanwhile, in the confusion matrix testing process, the accuracy is 68%, the precision is 69%, and the recall is 68%.