Advance Detection of Diabetic Retinopathy: Deep Learning Approach

Abstract

Diabetic retinopathy (DR) is an ophthalmological ailment wherein the diabetic individuals suffer from the formation of blockages, lesions, or hemorrhages predominantly in light-sensitive portion of said retina. Because of the increase in blood sugar, vascular blockage drives new vessel creation, giving rise to mesh-like patterns. As lack of timely treatment of DR results in vision loss, early diagnosis and professional assistance plays a crucial role. This can be achieved with a computer-aided diagnostic (CAD) system based on retinal fundus images. Various steps are involved in a CAD system, including as the detection, segmentation, and categorization of abnormalities in fundus images. This study is an effort to expedite the first screening of DR so as to meet the need of the increasing population of diabetic patients in the future. On publicly accessible datasets, we have trained and validated reliable classification algorithms enabling timely detection of DR. Convolutional neural networks (CNN)-based advanced deep learning models are used to fully use data-driven machine learning techniques for this purpose. We also defined the issue as the detection of DR of any grade (Grades 1–4) vs. No-DR in a binary classification (Grade 0). For training the models, we used 56,839 fundus pictures from the EyePACS dataset. On a test set from EyePACS, the models were put to the test (14,210 images). As compared to the established methods, experimental findings demonstrate superior outcomes through DenseNet with pre-trained weights. In the model’s evaluation on the EyePACS datasets, it achieved good results with an of 97.55% in binary and 78% in multiclass-classification.