ARNet-DR: Attention-Resnet Model for Diabetic Retinopathy Classification

Abstract

Diabetic retinopathy is a leading cause of blindness, it's largely preventable with early detection. Catching it in time is critical to stopping permanent vision loss. Right now, diagnosing it depends on specialists manually examining retinal images. This approach isn't just time-consuming; it's also highly subjective, leaving room for human error and inconsistency. To address these challenges, we propose ARNet-DR, a ResNet50-based architecture enhanced with a kernel attention mechanism for robust <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$D R$</tex> classification. The attention modules enable effective modeling of contextual dependencies in fundus images, prioritizing subtle yet discriminative features relevant to <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$D R$</tex> staging. ARNet-DR was assessed on the class-imbalanced APTOS 2019 Blindness Detection dataset, where it attained a training accuracy of 98.24 %, a test accuracy of 94.5%, and an F1-score of 0.99. These outcomes highlight the model's robustness in addressing class imbalance and image diversity, surpassing existing baseline approaches. Owing to its high precision and strong generalization performance, ARNetDR emerges as a viable decision-support aid for ophthalmologists and a significant advancement toward scalable, AI-powered diabetic retinopathy screening.