An Interpretable Hybrid Deep Learning Framework for Accurate Diabetic Retinopathy Detection Using Grad-CAM

Diabetic Retinopathy (DR) is an eye condition and a significant cause of sight loss, especially in regions where skilled ophthalmologists are scarce. Fundus photography requires the clinician's time and expertise for diagnosis and, hence, is not appropriate for high-volume screening. To overcome them, the following solution is proposed: DRNet is a hybrid deep learning framework that combines an attention-enhanced convolutional neural network (CNN) with Grad-CAM to provide gradient-based visual explanations. DRNet is a multi-branch network, with one branch tasked with capturing global retinal structures and the other with lesion appearance, e.g., microaneurysms and hard exudates. A learnable attention-weight block amalgamates these, producing a diagnostic accuracy of 96.96% when tested on standard DR datasets. DRNet can improve the sensitivity of early-stage disease detection by focusing on clinically meaningful regions of the fundus image. Grad-CAM generates a heat-map of explanations, meeting expert standards and enhancing clinical trust. Comparison results indicate that DRNet outperformed state-of-the-art models, especially in early-stage classification, across accuracy, transparency, and clinical usability.