Automated Early Detection of Diabetic Retinopathy Using Deep Learning

Overview

Diabetic retinopathy (DR) is a microvascular eye condition that can lead to vision loss and blindness in individuals with diabetes. DR has been on the rise over the past 30 years and is the leading cause of blindness in American adults. The number of patients with DR is estimated to increase to 191 million worldwide by 2030. Traditional diagnostic approaches by ophthalmologists are time-consuming, expensive, and often inaccessible in developing countries. This project proposes a multi-phase deep learning approach to automatically detect the severity of DR in retinal images. First, a preprocessing pipeline was built to fit retinal scans onto a circular crop and apply a Gaussian filter. A Region Proposal Network and a Mask R-CNN were trained using the DDR dataset to detect four key abnormalities that occur in retinas with DR (hemorrhages, soft exudates, hard exudates, microaneurysms). Using these models, segmentation maps were generated from the images in the APTOS-2019 retina dataset with opacities of the segments corresponding to the confidences of the predictions. These segmentation maps were overlaid onto the original retinal images, and the modified images were used to finetune an EfficientNet-B5-based CNN model to determine the severity of DR. The ensemble model pipeline achieved accuracies of 84.9% for the 5-class scale and over 90% for the 2-class and 3-class scales on the APTOS-2019 dataset, exceeding benchmark accuracies in previous works. The developed framework has the potential for a rapid, cost-effective, and accessible tool that can be readily used worldwide for the early detection of DR.

Competition history

• AJAS 2025 Category not listed

Related projects

Closest projects by meaning, across every fair and year in the corpus.

Browse more like this

• All Biomedical and Health Sciences projects
• AJAS Biomedical and Health Sciences projects
• Every AJAS 2025 project