Enhancing Diabetic Retinopathy Detection: A Deep Learning Approach with Advanced Image Preprocessing

Abstract

Diabetic Retinopathy (DR) is one of the main causes of blindness globally, which calls for early intervention in order to avoid permanent vision loss. In this study, a new approach is proposed for DR severity level classification with automatic detection using deep learning and retinal fundus images. An image preprocessing pipeline is introduced to enhance the quality of the images and critical features of the retina. The dataset used for the study was the APTOS 2019 dataset which consists of 3662 high-resolution retinal images divided into 5 levels of severity. For optimal results, the model was prepped with grayscale, contrast adjustments, and contour filling. VGG16, InceptionV3, and MobileNetV2 were the three CNN architectures that were implemented and tested against each other using accuracy, precision, recall, F1 score, and ROC-AUC. Out of all participants, MobileNetV2 was able to exceed the most expectations achieving 99.35% accuracy. Unlike the traditional methods, he was able to capture essential features of the retina, decrease false identified positives, and improve the classification of the stages of DR. The results accentuate the promise of AI-powered diagnostic solutions for fully automated systems for DR screening and sets stage for research involving retinal data as alternative markers for potential cardiovascular disease risks.