Classifying chronic obstructive pulmonary disease status using computed tomography imaging and convolutional neural networks: comparison of model input image types and training data severity.

Purpose: Convolutional neural network (CNN)-based models using computed tomography images can classify chronic obstructive pulmonary disease (COPD) with high performance, but various input image types have been investigated, and it is unclear what image types are optimal. We propose a 2D airway-optimized topological multiplanar reformat (tMPR) input image and compare its performance with established 2D/3D input image types for COPD classification. As a secondary aim, we examined the impact of training on a dataset with predominantly mild COPD cases and testing on a more severe dataset to assess whether it improves generalizability. Approach: CanCOLD study participants were used for training/internal testing; SPIROMICS participants were used for external testing. Several 2D/3D input image types were adapted from the literature. In the proposed models, 2D airway-optimized tMPR images (to convey shape and interior/contextual information) and 3D output fusion of axial/sagittal/coronal images were investigated. The area-under-the-receiver-operator-curve (AUC) was used to evaluate model performance and Brier scores were used to evaluate model calibration. To further examine how training dataset severity impacts generalization, we compared model performance when trained on the milder CanCOLD dataset versus the more severe SPIROMICS dataset, and vice versa. Results: A total of Conclusions: The CNN-based model with the combined 2D tMPR images and 3D lung view as input image types had the highest performance for COPD classification, highlighting the importance of airway information and that the fusion of different types of information as input image can improve CNN-based model performance. In addition, models trained on CanCOLD demonstrated strong generalization to the more severe SPIROMICS cohort, whereas training on SPIROMICS resulted in lower performance when tested on CanCOLD. These findings suggest that training on milder COPD cases may improve classification performance across the disease spectrum.