Abstract / Summary
Abstract Background Airway mucus plugging is a clinically consequential manifestation of chronic obstructive pulmonary disease (COPD), associated with elevated mortality and exacerbation risk, yet visual CT assessment remains subjective and labor-intensive. Existing automated methods rely on supervised classifiers trained on manually labelled exemplars, limiting scalability. Methods A self-supervised anomaly detection framework requiring no plug-level annotations was developed. A Lean Joint Embedding Predictive Architecture (LeJEPA) was pre-trained on 521 unlabeled chest CT scans to establish a manifold of normal airway topology, and Sparse-inferred Gaussian Regularization (SIGReg) compelled normal features to conform to a standard isotropic distribution, enabling pathological deviations to be identified as statistical outliers. The framework was applied to 8,521 inspiratory CT scans from COPDGene Phase 1. Results The framework yielded a continuous mucus plug burden score (range 0–18) that was inversely correlated with post-bronchodilator FEV₁ % predicted (Spearman ρ = −0.18, P < 0.001 ), positively correlated with parametric response mapping air trapping (ρ = 0.21, P < 0.001 ), and independently associated with increased all-cause mortality (adjusted hazard ratio [aHR] 1.27; 95% confidence interval [CI] 1.12–1.44; P < 0.001 ), respiratory mortality (subdistribution hazard ratio [sdHR] 1.31; P < 0.001 ), and exacerbation frequency (adjusted rate ratio [aRR] 1.31; P < 0.001 ) following adjustment for demographics, smoking, lung function, and CT-derived measures. Subgroup analysis revealed a more pronounced mortality association among participants with FEV₁ < 50% predicted (P-interaction = 0.03). Conclusions By obviating the requirement for plug-level annotation, this approach constitutes a scalable and reproducible alternative to both visual and supervised automated methods, supporting its utility for large-scale COPD phenotyping and risk stratification.