Deep Learning-Powered Framework for Enhanced Diabetic Retinopathy Detection and Classification

Abstract

Abstract: Diabetic Retinopathy (DR) is a complication of Diabetes that affects the retina and, under the proper diagnosis, can lead to blindness. Diagnosis and accurate detection are crucial, especially for timely control and management. Deep Learning has re-emerged as a strong candidate for DR detection, and CNNs have provided a broad promise of potential automated DR detection systems. This work proposes a novel strategy to improve the current methods of detecting DR using advanced deep learning methods. The proposed system combines a strong CNN framework with transfer learning that enables the use of pre-trained models together with rigorous image preprocessing for variability in retinal image databases. The proposed methodology makes the detection efficient and accurate enough to identify DR in its preliminary stages. This work explores a new and efficient deep-learning method for better diagnosis and classification of DR using an improved CNN model architecture. For the proposed model, the following performances were obtained: accuracy = 97.10%, precision = 97.00%, and sensitivity = 96.50% when evaluated on a dataset of 3,000 retinal fundus images outperforming VGG16 and ResNet50 by almost 3%, 5%, and 2.5% respectively. Significant contributions include transfer learning, hybrid feature extraction, and preprocessing, which help overcome dataset variance and computation time issues. In this way, this solution is helpful for the diagnostic process, as it helps to reduce vision loss and becomes useful for healthcare professionals as they get the opportunity to change the state for the better in time. This research explains how deep Learning can improve important issues in the healthcare sector and how it can be used in medical diagnosis.