Fundus images classification for Diabetic Retinopathy using Deep Learning

Abstract

Diabetes is a worldwide chronic disease, which can even affect life and has several complications. Diabetic Retinopathy is the most serious complication of diabetes. Early detection still has a chance of cure, but there are many cases of serious blindness. Today's machine learning and deep learning are significant technology, where perform excellently in many classification fields. In this paper, we modify the architecture of the VGG-16 and ResNet-50 models to classify the severity grading of Diabetic Retinopathy with the dropout concept. In addition, contrast-limited adaptive histogram equalization (CLAHE) is used in data pre-processing to improve the quality of the fundus image of diabetic retinopathy, and data expansion is used to solve the problem of data imbalance and improve training over-fitting. After the pre-processing of the fundus image and the models are modified with dropout, the confusion matrix is used to evaluate the model. The classification accuracy of the two models is 94.03% and 97.21%. The average sensitivity is over 70%, and the specificity is over 90%.