Enhancing fundus image classification with semantic segmentation-based attention mask

Abstract

This research proposes a method for classifying ocular diseases in fundus images using semantic segmentation as an attention mechanism. Unlike conventional approaches that rely on the entire retinal image, the proposed framework emphasizes anatomically relevant regions extracted via segmentation of the optic disc, optic cup, and retinal vessels. These segmentation masks are integrated into the classification pipeline to enhance feature learning. A EfficientNetB6-based classifier is utilized to evaluate the impact of this strategy. Experimental results demonstrate improvements in classification performance across multiple evaluation metrics.