Segmentation-Based Attention Mask for Enhancing Fundus Image Diagnosing

Abstract

Abstract. Approach to improve the classification of ocular diseases from fundus images by leveraging semantic segmentation as an attention mechanism. Key anatomical structures – optic disc, optic cup, and retinal vessels – are segmented using deep learning models and combined into weighted attention masks. These masks guide a classifier based on EfficientNetB6 to focus on clinically relevant regions, resulting in significant improvements in diagnostic accuracy. The method enhances detection sensitivity for subtle disease features and increases model interpretability.