Classification and grading of lesions in diabetic retinopathy using convolutional neural networks based on VGG 19 architecture

Abstract

Diabetic Retinopathy (DR) is a type of eye disease that occur during diabetic condition which can harm the retina resulting in blindness. If not treated properly at its early stages, it develops severity. Eventually, it can block the light passing via the optical nerves which thereby, damages the blood vessel in retina making the patient blind. Therefore, our research aimed to eradicate this problem by identifying the various types of lesions using an automated segmentation approach based on deep neural convolutional network (ConvNet). Also, there can occur morphological variations in retina leading to less blood flow across the retina, which can decline the pericytes cells too. As initially stage of diabetic retinopathy has no symptoms the patient is not aware at onset of disease, which creates risk. Hence, early detection and automated diagnosis has become necessary to avoid visual damage. newlineIn this research, retinal defects of DR such as exudates, haemorrhages, microneurysms are accurately identified using proposed segmentation methods from digital fundus images and also the grades of DR as mild, moderate, severe, No PDR, PDR were labeled precisely from the obtained fundus images. This was achieved using Deep Convolutional Neural Network (DCNN), trained using VGG-19. The classification of diabetic retinopathy (DR) using color fundus images needs proper feature extraction methods to classify and detect the existence and relevance of various subtle small features , as well as an efficient classification system, drives this as cumbersome and labor intensive. newline