ENHANCING DIABETIC RETINOPATHY DETECTION USING CNN ENSEMBLES AND GRAD-CAM ON RETINAL FUNDUS IMAGES

Abstract

Diabetic Retinopathy (DR) is a progressive eye disease and a leading cause of blindness among individuals with diabetes. Early detection is critical for preventing irreversible vision loss. In this paper, we propose an automated DR detection system using deep learning and ensemble methods to improve classification accuracy across five DR severity levels. The system was trained on the APTOS 2019 Blindness Detection dataset comprising 3,662 retinal fundus images. Three convolutional neural network (CNN) models-ResNet18, ResNet50, and EfficientNetB3-were implemented individually and then combined using ensemble techniques: majority voting, weighted voting, and stacked ensemble with a random forest meta-classifier. Image preprocessing techniques such as LAB color conversion, CLAHE, denoising, and data augmentation were used to enhance diagnostic features. The ensemble models significantly outperformed the individual CNNs, with the stacked ensemble achieving the best results: 85.27% accuracy, 0.933 ROC AUC, and 0.7352 PR AUC. The system's interpretability was improved using Grad-CAM, providing visual heatmaps of model decision regions. These results demonstrate that ensemble learning, coupled with interpretable AI, offers a robust and clinically relevant approach to DR detection.