Deep learning-based computed tomography assessment for lung function prediction in chronic obstructive pulmonary disease

Abstract

Deep learning models based on medical imaging enable numerical functional predictions in combination with regression methods. In this study, we evaluate the prediction performance of a deep learning-based model for the raw value and percent predicted forced expiratory volume in one second (FEV1) in patients with chronic obstructive pulmonary disease (COPD). To this end, ResNet50-based regression prediction models were constructed for FEV1 and %FEV1 based on 200 CT scans. 10-fold cross-validation was performed to yield ten models in aggregate. The prediction model for %FEV1 was externally validated using 20 data points. Two hundred internal CT datasets were assessed using commercial software, producing a regression model predicting airway [%WA] and parenchymal indices [%LAV]. The average Root Mean Squared Error(RMSE) value of the 10 predictive models was 627.65 for FEV1 as per internal validation and 15.34 for %FEV1. The externally validated RMSE for %FEV1 was 11.52, whereas that for %FEV1 was 23.18. The predictive model for %FEV1 yielded significant positive correlations corresponding to both internal and external validation. The proposed models exhibited better prediction accuracy for %FEV1 than for FEV1. Further studies are required to improve the accuracy further and determine the validity of longitudinal applications.