AttResAMD: An Attention-Driven Deep Learning Framework for Expert-Level Automated Classification of Age-Related Macular Degeneration from Fundus Photography.

Age-related macular degeneration (AMD) represents a leading cause of vision impairment worldwide, and its early detection is critical for initiating timely interventions and optimizing disease management. In this study, we propose an attention-driven deep learning framework, AttResAMD, for expert-level automated detection and classification of AMD from fundus images, by integrating channel-wise and spatial-wise attention mechanisms. The AttResAMD model demonstrated robust performance in a three-class classification task (normal, dry AMD, and wet AMD), achieving micro-average areas under the receiver operating characteristic curve (AUC) of 0.971 and 0.962 in the internal and external testing cohorts, respectively. In an independent comparative evaluation cohort, the diagnostic accuracy of five experienced ophthalmologists ranged from 82.1% to 89.7% for AMD detection, and from 65.0% to 82.5% for subclassifying dry versus wet AMD. In contrast, AttResAMD achieved superior performance with an accuracy of 97.4% for AMD detection and 82.5% for subtype differentiation. In summary, AttResAMD shows strong potential as a non-invasive, cost-effective, and scalable tool for automated AMD screening and classification, supporting clinical decision-making in ophthalmic care.