LAGNet: Label Attention Graph Networks for Ocular Disease Classification Using Fundus Images

Huanwen Liang, Leya Li, Yu Lu, Qianying Liu, Huilin Ge, Xianghua Fu

Abstract: Diagnosis of ocular diseases from fundus images presents a formidable challenge for clinicians due to the complexity of diseases, co-occurrence of multiple diseases, and the need for high-level expertise. In this paper, we introduce a novel framework, the Label Attention Graph Network (LAGNet), specifically designed to address these challenges. It leverages the Convolution Vision Transformer (CvT) as an image feature extractor and incorporates the Attention Graph Neural Network (AGNN) to establish label attention, enabling the model to learn intricate internal correlations among ocular disease labels. Extensive experiments were conducted using the ODIR-5K dataset, with LAGNet demonstrating superior performance. Specifically, it achieved an Overall Precision (OP) of 54.42%, an Overall Recall (OR) of 55.61%, and an Overall F1-Score (OF1) of 55.07%, representing an improvement of approximately 6.7% in OP, 12.2% in OR, and 9.6% in OF1 compared to SOTA. This research not only contributes to the advancement of multi-label medical image classification but also holds substantial promise for practical clinical applications, potentially leading to improved patient care and outcomes. Our code and models are available at: https://github.com/lvyupku/LAGNet.