Single Inspiratory Chest CT-based Generative Deep Learning Models to Evaluate Functional Small Airway Disease

This study developed and validated Deep Learning (DL) models to generate Parametric Response Mapping (PRM) and predict functional Small Airway Disease (fSAD) using only single inspiratory chest CT. Two methods, P-PRM and G-PRM, were proposed. G-PRM utilized biphase deformable registration and a generative model to synthesize expiratory CT, followed by PRM generation. PRM reference standards were produced using the aview software, which registered inspiratory–expiratory CT scans. Using data from 308 patients, 5-fold cross-validation was conducted, followed by evaluations on internal and external test sets. G-PRM outperformed P-PRM in fSAD detection, achieving higher sensitivity (86.3% vs 38.9%) and AUC (0.86 vs 0.70). The model demonstrated strong reproducibility and generalizability across various conditions and CT equipment, with particularly high diagnostic performance in the PRISm group. This approach offers potential for early diagnosis using PRM while reducing radiation exposure.