ABSTRACT
Diabetic retinopathy is a disease caused in human beings as a result of diabetic complications. The light-sensitive tissues present in human eye are responsible for vision. Diabetic retinopathy in its severe form can damage the blood vessels of these tissues, causing permanent loss of vision to a patient. A considerable number of people across the globe are susceptible to this problem. The worst part of the problem is that the patients may remain asymptomatic at the initial stages of the disease. However, detection of the disease at an early stage can save a patient from losing his/her vision. One of the techniques of detecting the stage of this disease is through the study of retinal images. This chapter presents a comprehensive comparative analysis of different neural network models, viz. VGG16, InceptionV3, Resnet50, and MobileNetV; in detecting diabetic retinopathy by classifying the severity of the disease on a scale of 0−4. Retinal images with multiple properties like varying contrast, intensity, brightness, etc., were used to train the models 330and predict the stage of disease the patient is in at present. The results of the analysis reveal that the model MobileNetV2 records maximum training and testing accuracy amongst the rest. Moreover, accuracy graph, loss graph, confusion matrix, and classification report for every model have also been presented, for increasing number of epochs. It can be seen that the overall performance of all the models improves with an increase in the number of epochs.
