Chronic Obstructive Pulmonary Disease: Thoracic CT Texture Analysis and Machine Learning to Predict Pulmonary Ventilation

Background Fixed airflow limitation and ventilation heterogeneity are common in chronic obstructive pulmonary disease (COPD). Conventional noncontrast CT provides airway and parenchymal measurements but cannot be used to directly determine lung function. Purpose To develop, train, and test a CT texture analysis and machine-learning algorithm to predict lung ventilation heterogeneity in participants with COPD. Materials and Methods In this prospective study ( : NCT02723474; conducted from January 2010 to February 2017), participants were randomized to optimization ( = 1), training ( = 67), and testing ( = 27) data sets. Hyperpolarized (HP) helium 3 ( He) MRI ventilation maps were co-registered with thoracic CT to provide ground truth labels, and 87 quantitative imaging features were extracted and normalized to lung averages to generate 174 features. The volume-of-interest dimension and the training data sampling method were optimized to maximize the area under the receiver operating characteristic curve (AUC). Forward feature selection was performed to reduce the number of features; logistic regression, linear support vector machine, and quadratic support vector machine classifiers were trained through fivefold cross validation. The highest-performing classification model was applied to the test data set. Pearson coefficients were used to determine the relationships between the model, MRI, and pulmonary function measurements. Results The quadratic support vector machine performed best in training and was applied to the test data set. Model-predicted ventilation maps had an accuracy of 88% (95% confidence interval [CI]: 88%, 88%) and an AUC of 0.82 (95% CI: 0.82, 0.83) when the HP He MRI ventilation maps were used as the reference standard. Model-predicted ventilation defect percentage (VDP) was correlated with VDP at HP He MRI ( = 0.90, < .001). Both model-predicted and HP He MRI VDP were correlated with forced expiratory volume in 1 second (FEV ) (model: = -0.65, < .001; MRI: = -0.70, < .001), ratio of FEV to forced vital capacity (model: = -0.73, < .001; MRI: = -0.75, < .001), diffusing capacity (model: = -0.69, < .001; MRI: = -0.65, < .001), and quality-of-life score (model: = 0.59, = .001; MRI: = 0.65, < .001). Conclusion Model-predicted ventilation maps generated by using CT textures and machine learning were correlated with MRI ventilation maps ( = 0.90, < .001). © RSNA, 2019 See also the editorial by Fain in this issue.