Fusion-Attention Diagnosis Network (FADNet): An end-to-end framework for optic disc segmentation and ocular disease classification

Abstract

Hundreds of millions of people suffer from blindness and severe vision impairment due to pathologic myopia and other ocular illnesses, posing substantial worldwide public health issues. Accurate diagnosis and timely treatment of these conditions heavily rely on the precise segmentation of key anatomical structures in fundus images, such as the optic disc, which is essential for identifying disease types for timely and effective clinical interventions. Although medical image analysis has made significant progress, existing methods often address segmentation and classification as separate tasks, resulting in limited performance and poor clinical applicability. In this work, we present an innovative end-to-end framework named Fusion-Attention Diagnosis Network (FADNet), which unifies ocular disease classification and optic disc segmentation tasks. The core innovation of FADNet lies in the Dynamic Weighted Feature Fusion strategy, which seamlessly integrates the segmentation mask into the original fundus image using a context-aware weighting mechanism. This approach amplifies the contribution of pathological regions, enhancing feature relevance for subsequent classification. The framework first employs an Attention U-Net to achieve accurate optic disc segmentation, followed by a ResNet-based classification network to diagnose ocular diseases from the fused image. Experiments on the iChallenge-PM and Retina datasets indicate that FADNet attains state-of-the-art performance, achieving accuracies of 97.1% in binary classification and 90.4% in multi-class classification, surpassing current methodologies. FADNet outperforms previous methods by its joint optimization strategy, which improves the synergy between segmentation and classification tasks, resulting in notable improvements in diagnostic accuracy and robustness. FADNet showcases its effectiveness and adaptability across multiple datasets, offering a comprehensive and practical solution for the automated diagnosis of ocular diseases, with significant potential for future deployment in clinical settings.