Lesion‐Aware Prototypes and Dynamic Comorbidity Graphs for Long‐Tailed Multilabel Retinal Disease Classification

ABSTRACT Multilabel retinal disease classification from fundus images remains challenging due to severe long‐tailed label distributions, heterogeneous lesion scales, and patient‐specific disease comorbidity. To address these issues, we propose LPC‐Net, a unified framework that improves the recognition of rare diseases while preserving interpretability. LPC‐Net introduces: (i) an anatomically guided lesion‐aware multi‐instance learning module that aggregates multiscale patches from global, macula‐centered, and optic‐disc‐centered regions without requiring lesion annotations; (ii) a prototype‐balanced margin regularization that enforces geometric separation for tail‐class prototypes to mitigate feature collapse under imbalance; and (iii) a sample‐specific comorbidity graph that blends dataset‐level co‐occurrence priors with instance‐level predictions, enabling a lightweight label GNN to refine logits via residual corrections. Experiments on ODIR‐5K and RFMiD demonstrate consistent gains over strong baselines. LPC‐Net achieves 0.901 mAP on ODIR‐5K and 0.768 mAP on RFMiD, with notable improvements on tail classes. Under prespecified six‐output mappings and without target‐domain fine‐tuning, cross‐dataset evaluation yielded 0.846/0.618/0.681 (Macro‐AUC/Macro‐F1/mAP) for ODIR‐5K‐to‐RFMiD and 0.836/0.598/0.659 for RFMiD‐to‐BRSET. These results are interpreted as transfer under harmonized label definitions rather than universal diagnostic equivalence or clinical generalizability.