CT vessel features improves COPD classification

Makimoto,K.,Genkin,D.,Hogg,J.,Bourbeau,J.,Tan,W.,Kirby,M.

DOI: https://doi.org/10.1183/13993003.congress-2024.pa5241

IF: 24.3

2024-11-01

European Respiratory Journal

Abstract:Background: Computed tomography (CT) features extracted from the lung and airways can classify COPD with machine learning (ML) models (Puchakayala Radiol 2023). However, pulmonary vasculature changes occur in early COPD (Santos Eur Respir J 2002), and it is unknown if CT pulmonary vessel features can improve COPD classification. Aim: To compare ML models for classifying COPD using combinations of CT features extracted from the lung, airways, and vessels. We hypothesize that the inclusion of CT vessel features to lung and/or airway features will improve ML performance for COPD classification compared to each feature set alone. Methods: Participants that were never-/ever-smokers from the population-based CanCOLD study underwent spirometry and CT imaging. 107 CT PyRadiomics features were extracted from the lung, airways, vessels: 18 first-order, 14 shape, and 75 texture features. ML models including 6 demographics (age, sex, race, BMI, smoking-status, pack-years), low-attenuation-areas-below -950HU, and CT PyRadiomics were evaluated for classifying COPD (FEV1/FVC<0.70). The area under the receiver operating characteristic curve (AUC) and DeLong's test were used to compare models. Results: 1302 participants were evaluated (n=669 no COPD; n=363 mild COPD; n=270 moderate/severe COPD). The model with CT lung parenchyma+airway+vessel features (AUC=0.87; p<0.05) obtained a higher performance than models with lung (AUC=0.78), airway (AUC=0.84), or vessel features alone (AUC=0.81). The lung+vessel (AUC=0.83) and the airway+vessel model (AUC=0.85) obtained higher performance than the lung only model (AUC=0.78; p<.05). Conclusion: Inclusion of CT features extracted from the pulmonary vessels improves classification of COPD in a mild COPD cohort.

respiratory system