CG-TRAN: A novel multi-label retinal disease classification model with partially known pathologies

Abstract

Early detection of retinal diseases is vital to preventing partial or permanent blindness. However, the diagnostic process is often impeded by the complexity of interrelated lesions and the challenge of incomplete or missing pathology labels, which require specialized expertise in ophthalmic diagnosis. To address these limitations, we propose CG-Tran, a novel multi-label classification model that leverages partially known pathology information to diagnose retinal diseases. This approach integrates a pathology graph neural network with graph-based feature extraction to handle partially known pathologies, enabling more accurate multi-label classification of retinal diseases. To model the intricate interrelationships among ocular diseases, CG-Tran employs BERT-GNN to learn label interactions and construct a comprehensive fundus pathology graph. Additionally, an enhanced attention mechanism incorporates known pathology label features, bridging the gap between incomplete pathology information and fundus image data. These innovations collectively empower the model to overcome the challenges of missing or incomplete pathology labels. The model’s performance is rigorously evaluated on the Multilabel Retinal Disease (MuReD) dataset. Results demonstrate that CG-Tran significantly improves diagnostic accuracy, especially as more pathology labels become available. Under conditions with 0% and 75% partially known labels, CG-Tran achieves mean average precision (mAP) scores of 69.9% and 72.1%, respectively—outperforming the baseline model by 1.0% and 1.9%. This innovative architecture excels in multi-label classification tasks, particularly in recognizing and distinguishing complex and interrelated retinal lesions with partially known pathology. It offers a promising solution for early detection and accurate diagnosis of retinal diseases, addressing critical limitations in existing diagnostic methods.