Diabetic Retinopathy Severity Categorization in Retinal Images Using Convolution Neural Network

Abstract

As one of the most common causes of blindness, Diabetic Retinopathy (DR) is a devastating disease of the retina. Doctors utilize optical coherence tomography (OCT) and fundus photography to examine the retina's thickness, grade, structure, and edoema, as well as to detect scarring. To categorize and stage the disease, deep learning algorithms are usually used to analyze OCT or fundus pictures and extract unique properties for each stage of DR. Until the disease has progressed significantly, the signs and symptoms are difficult to detect. Patients with a high propensity for DR vision loss can benefit from early detection and monitoring. However, because to the intricacy of the image captured by colour fundus photography, human detection and classification of Diabetic Retinopathy is a difficult and error-prone task. DR levels have previously been detected and classified using machine learning techniques armed with several feature extraction techniques. This research presents a Resemblance Pixel Vector Set with Convolution Neural Network (RPVS-CNN) model for accurate detection of severity detection of DR using MR images. Diabetic Retinopathy can be objectively diagnosed and graded using the suggested method, eliminating the necessity for a retina specialist and increasing the number of people who can receive retinal care. A network with CNN architecture and data augmentation is developed that can identify the intricate aspects of the classification task including micro-aneurysms, exudate and haemorrhages on the retina and subsequently deliver a diagnosis automatically and without input from the end-user. The proposed model is compared with the traditional models and the results illustrate that the proposed model performance is enhanced.