Self-Attention Guided Two-Stream Graph Convolutional Network for Improved Multilabel Retinal Disease Diagnosis

Abstract

Retinal disorders are the leading cause of partial and permanent vision loss across the globe, making early diagnosis and timely intervention paramount for effective treatment planning. Recent deep learning methods for retinal disease diagnosis from fundus images acquired using optical imaging sensors primarily focus on single-disease detection, despite the fact that a fundus image may contain multiple diseases. However, the multilabel retinal disease classification task remains challenging due to high interclass similarity, large variations in lesion shape and size, variations in sensor-acquired image quality, and class imbalance. While limited attempts have been made to this end, they remain limited in effectively capturing the complex relationships among diseases. In addition, they fall short of extracting salient and richer visual features and their relationships, resulting in suboptimal performance. To address these issues, we propose a self-attention guided two-stream graph convolutional network (SAT-GCNet), a unified framework that seamlessly models visual features and disease relationships via graph-based learning for effective multilabel retinal disease diagnosis. Specifically, in the first stream, we employ multiple CNN backbones to capture diverse visual patterns from fundus images and subsequently introduce a self-attention guided adaptive GCN (SAA-GCN) block to model feature correlations. While in the second stream, domain-specific label embeddings are generated and used in the SAA-GCN block to effectively capture disease correlations and incorporate rich semantic information, thereby guiding the model for better prediction. Extensive experimental evaluations on two benchmark datasets, OIA-ODIR, and MuReD, demonstrate that SAT-GCNet achieves superior performance compared to current state-of-the-art approaches, while also providing enhanced reliability.