CNN Transfer Learning for Two Stage Classification of Diabetic Retinopathy using Fundus Images

Abstract

Diabetic retinopathy is a common retina disease caused by diabetes, which is increasing rapidly worldwide along with diabetic Mellitus. It is very difficult to diagnose in the beginning as it is asymptomatic, which leads to blindness. Presently, there are about 422 million diabetes patients, and it is projected that the number will be 552 million in 2030, as per a report from the World Health Organization. With the advancement in Artificial Intelligence, Deep Learning, and Computer Vision fields, many CNN-powered models have been developed for the detection and classification of DR using color fundus images. Early diagnosis of DR will increase recovery and decrease the possibility of vision loss threats. In this study, we aim to classify Diabetic Retinopathy in two stages, diseased versus healthy images. Three different pre-trained CNN models, namely, VGG16, InceptionV3, and MobileNet were deployed through transfer learning. Messidor and Messidor-2, two publicly available datasets are used in the training and testing of these CNN models. In terms of classification accuracy, the highest result of 84% was obtained using InceptionV3, whereas MobileNet and VGG16 shows 83% and 78% accuracy, respectively. The Highest 86 % precision for the healthy class and 88% sensitivity for the diseased class is shown by VGG16.