Gated Feature Fusion Networks for Diabetic Retinopathy Grading

Diabetic retinopathy (DR) is a leading cause of vision loss worldwide, underscoring the critical need for effective DR screening and intervention. Screening via direct ophthalmoscopy is notoriously known for being extremely cumbersome and subjective. This has led to the design of automated systems aimed at fundus image analysis and interpretation. In this work, we propose a novel approach for hierarchical feature fusion, which merges multi-level outputs from ResNet50 and GoogLeNet employing a gated classifier. Intricate lesion details as well as broader retinal features are aimed to be captured by this dual-stream architecture, and the shortcomings of single-model approaches are overcome. Extensive experiments on benchmark DR datasets have validated the method’s effectiveness, with state-of-the-art accuracies of 86.1% on APTOS2019, 88.0% on EyePACS, 84.2% on DDR, 79% on IDRiD, and 81% on Messidor. The strength and improved generalization of the proposed method are thereby confirmed, and its capability to enhance automated DR screening and support clinical decision-making is demonstrated.