Quantitative CT Analysis in Chronic Hypersensitivity Pneumonitis: A Convolutional Neural Network Approach

Chronic hypersensitivity pneumonitis (cHP) is a heterogeneous condition, where both small airway involvement and fibrosis may simultaneously occur. Computer-aided analysis of CT lung imaging is increasingly used to improve tissue characterization in interstitial lung diseases (ILD), quantifying disease extension, and progression. We aimed to quantify via a convolutional neural network (CNN) method the extent of different pathological classes in cHP, and to determine their correlation to pulmonary function tests (PFTs) and mosaic attenuation pattern.The extension of six textural features, including consolidation (C), ground glass opacity (GGO), fibrosis (F), low attenuation areas (LAA), reticulation (R) and healthy regions (H), was quantified in 27 cHP patients (age: 56 ± 11.5 years, forced vital capacity [FVC]% = 57 ± 17) acquired at full-inspiration via HRCT. Each class extent was correlated to PFTs and to mosaic attenuation pattern.H showed a positive correlation with FVC%, FEV1% (forced expiratory volume), total lung capacity%, and diffusion of carbon monoxide (DLCO)% (r = 0.74, r = 0.78, r = 0.73, and r = 0.60, respectively, p0.001). GGO, R and C negatively correlated with FVC% and FEV1% with the highest correlations found for R (r = -0.44, and r = -0.46 respectively, p0.05); F negatively correlated with DLCO% (r = -0.42, p0.05). Patients with mosaic attenuation pattern had significantly more H (p = 0.04) and lower R (p = 0.02) and C (p = 0.0009) areas, and more preserved lung function indices (higher FVC%; p = 0.04 and DLCO%; p = 0.05), but did not show more air trapping in lung function tests.CNN quantification of pathological tissue extent in cHP improves its characterization and shows correlation with PFTs. LAA can be overestimated by visual, qualitative CT assessment and mosaic attenuation pattern areas in cHP represents patchy ILD rather than small-airways disease.