RetinoFusionNet: A Scalable and Interpretable Vision Transformer Framework for Diabetic Retinopathy Detection

Diabetic Retinopathy (DR) is a leading cause of preventable blindness, highlighting the need for automated screening systems that combine accuracy, efficiency, and interpretability. The present study introduces RetinoFusionNet, a prototype-guided Vision Transformer (ViT) that unifies multi-resolution patch embedding, cross-scale attention, and class-specific prototype reasoning to capture both localized lesions and broader retinal structures. By segmenting fundus images into varied patch sizes, the model effectively extracts fine and global features, while cross-scale attention establishes dependencies across distant abnormalities. Prototype-based learning provides interpretable visual anchors that align predictions with clinically recognized disease patterns, enhancing trust in automated decisions. Comprehensive evaluation on EyePACS, APTOS 2019, and Messidor-2 datasets demonstrates state-of-the-art accuracy with only a 4.1–4.5% cross-dataset drop, outperforming ViT and ProtoPNet, which show a decline of 8.3–12.1%. RetinoFusionNet also achieves a per-image inference time of 78 ms, reduces memory usage by 42% compared to standard ViTs, and operates at just 14.6 GFLOPs, confirming its robustness and deployment feasibility. By combining precision, computational efficiency, and transparency, RetinoFusionNet is established as a practical and scalable solution for large-scale DR screening, particularly in resource-limited clinical settings.