Curriculum-guided adaptive representation decoupling for multimodal semi-supervised chest disease classification

Chest diseases remain a major source of morbidity and mortality, while automated chest X-ray diagnosis is constrained by annotation scarcity and heterogeneous clinical information. Frontal and lateral radiographs provide complementary views within one X-ray modality, whereas pre-diagnostic clinical context constitutes a heterogeneous textual modality. We propose Curriculum-guided Adaptive Representation Decoupling (CARD) for multimodal semi-supervised chest disease classification. Label-aware Dynamic Curriculum Learning combines prediction persistence with dynamic confidence thresholds to partition unlabeled sample-label pairs into easy, stable, and hard groups. Persistence-guided Adaptive Feature Decoupling transfers these reliability cues to cross-view representation learning. It balances view-consistent alignment against preservation of projection-specific radiographic information. The learned multi-view radiographic representation is then fused with clinical-text features for multimodal prediction. Under the matched radiographic protocol, CARD achieved a Macro-AUROC of 75.9%, exceeding STiL by 1.1 percentage points. Adding matched pre-diagnostic clinical text increased Macro-AUROC to 77.0%. These findings support reliability-aware curriculum learning, adaptive cross-view representation decoupling, and clinically aligned multimodal fusion.