C[Formula: see text]Net: A co-occurrence and consistency-aware framework for structured multi-label fundus diagnosis.

Early concurrent screening of multiple retinal pathologies is vital for preventing vision loss but poses substantial challenges for deep learning models. These include complex disease co-occurrences, severe class imbalance (long-tailed distributions), and the imperative for diagnostic logic (e.g., avoiding conflicting “Normal” and “Disease” predictions. This paper presents C[Formula: see text]Net, a unified Co-Occurrence and Consistency-Aware Network designed to address these coupled challenges simultaneously. Specifically, C[Formula: see text]Net integrates three complementary components: (1) a Co-Occurrence-Aware Supervised Contrastive Loss (CoaSupCon) that explicitly encodes positive label dependencies to capture disease correlations; (2) an Auto-Adaptive Multi-Label Focal Loss (Auto-MLFocal) that automatically learns per-class balancing and focusing parameters to mitigate long-tail imbalance; and (3) a Diagnostic Consistency Mechanism (DCM) that enforces negative dependencies and clinical logic via conflict suppression and default-to-other constraints. By unifying positive and negative dependency modeling, C[Formula: see text]Net promotes robust and clinically reliable predictions. Extensive evaluations on the ODIR-5K and MuReD benchmarks under a rigorous 5-fold cross-validation protocol demonstrate that C[Formula: see text]Net achieves consistently better performance than representative state-of-the-art methods. On ODIR-5K, it achieves a macro-average F1 (AF1) of 72.1% and a macro-average Precision (mAP) of 74.4%. On MuReD, it reaches an AF1 of 77.1% and a mAP of 80.5%. Results confirm that C[Formula: see text]Net not only improves diagnostic accuracy but also improves logical reliability, offering a unified framework for multi-label medical imaging.