10 - Grading of diabetic retinopathy using deep learning techniques

Abstract

Diabetic retinopathy (DR) is the most common prevalent cause of avoidable vision impairment worldwide, primarily affecting the age group of people between 20 and 60. Given that diabetes is the leading cause of blindness and that one in every three diabetic patients has some degree of DR, there is an urgent need to raise diabetics’ awareness of this serious health risk. The importance of DR screening programs, as well as the difficulty of obtaining reliable early DR diagnosis at a reasonable cost, necessitates the creation of a computer-aided diagnosis tool. In the proposed system, the input image is taken from the Indian Diabetic Retinopathy Image Dataset and given to the ALEXNET, VGG-16, RESNET50, INCEPTION V3, GOOGLENET convolutional neural networks, and the various performance like sensitivity, specificity, accuracy, positive predictive value, negative predictive value, receiver operating characteristic curve and area under curve is compared with two- and multiclass grading of DR. The two-class VGG-16 gives an accuracy of 93% and five-class GOOGLENET gives 87.4%, higher compared to other proposed pretrained networks.