Tri-attention-augmented residual convolutional neural network with explainable AI for automated early diabetic retinopathy screening in resource-constrained environments

Diabetic retinopathy (DR) is a leading cause of vision loss among patients with diabetes, affecting the retina and optic disc. Early detection is challenging due to the slow progression and minimal symptoms in the initial stages. This study aims to develop a fully automated deep learning (DL) based system for early identification and screening of DR using an enhanced tri-attention module (TAM) augmented deep convolutional neural network (DCNN) explainable AI (XAI). A DCNN model was designed and trained to classify DR into four categories: healthy eye (L-0), mild DR (L-1), moderate DR (L-2), and severe DR (L-3). To enhance feature learning, TAM–squeeze-and-excitation (SE), channel attention (CA), and spatial attention (SA)–were incorporated into the architecture. Furthermore, XAI techniques, including Gradient-weighted Class Activation Mapping (Grad-CAM), were applied to generate heatmaps that highlight clinically relevant retinal regions, giving qualitative visual interpretation. The proposed attention-augmented DCNN achieved an accuracy of 99.4% on external testing and 97.01% on internal testing, surpassing the baseline CNN and previous state-of-the-art (SOTA) approaches. The integration of attention modules improved precision, recall, and AUC across all DR stages. Grad-CAM Visually highlights regions that appear consistent with lesion-rich retinal areas. These comprehensive experimental findings on publicly available APTOS datasets demonstrate that the proposed Tri-Attention (TA)-induced XAI model improves DR classification performance under the evaluated dataset conditions in comparison with existing SOTA models. Consequently, the proposed framework demonstrates significant potential for integration into ophthalmic image analysis workflows. Although the current study is limited to retinal fundus images, future adaptation to OCT-assisted diagnostic systems may be explored through appropriate preprocessing and image representation strategies.