A Multi-Scale Attention-Driven Deep Learning Model for Accurate and Interpretable Diabetic Retinopathy Severity Classification

Diabetic Retinopathy (DR), a complication of prolonged diabetes, can result in progressive and irreversible vision impairment if not detected promptly. Manual assessment of retinal fundus images remains a standard diagnostic practice but is labor-intensive and subject to variability among clinicians. Recent advances in deep learning (DL) have enabled automated DR detection, yet many existing models struggle with distinguishing subtle severity levels and maintaining performance across diverse datasets. To address these limitations, this research introduces MSCAN-DR, a novel convolutional model that integrates multi-scale feature extraction with an attention-guided refinement mechanism. The architecture employs parallel convolution blocks of varying kernel sizes to extract diverse spatial features, followed by feature alignment and attention modules to emphasize lesionrelevant regions. Evaluation demonstrates that the model's high effectiveness, achieving 98.63% classification accuracy, with 95% sensitivity and 100% specificity on the test dataset and 99% classification accuracy, with 95% sensitivity and 99.5% specificity on the training dataset. Enhanced preprocessing and lightweight design ensure both robustness and efficiency. MSCAN-DR offers a reliable and computationally efficient framework for automated DR grading, supporting timely diagnosis in clinical and screening environments.