Classifying chronic obstructive pulmonary disease using computed tomography imaging and 2D and 3D convolutional neural networks

Convolutional Neural Network (CNN)-based models using Computed Tomography (CT) images classify Chronic Obstructive Pulmonary Disease (COPD) patients with high accuracy, but studies have used various different input images and it is unclear what input images are optimum, particularly in a milder COPD cohort. We propose a novel approach using 2D airway-optimized topological multi-planar reformat (airway-optimized tMPR) images as well as novel 3D fusion methods and compared the performance of these models with various established 2D/3D CNN-based methods in a population-based mild COPD cohort. Participants from the CanCOLD study were evaluated. We implemented several 2D/3D models adapted from the literature. Existing CNN-based models were trained using 2D collages of axial/coronal/sagittal slices, and colored and binary airway images. 3D models consisting of 15 axial inspiratory/expiratory slices were selected, and input and output combination methods were investigated. For the proposed models, 2D airway-optimized tMPR images were constructed using cut-surface renderings to convey shape and interior/contextual information. 3D output fusion of axial/coronal/sagittal images, as well as output fusion of the axial and 3D airway tree, were also investigated. Finally, the output fusion of 2D airway-optimized tMPR methods and 3D lungs combined method was investigated. 742 participants were used for training/validation and 309 for testing. The 2D and 3D methods adapted from the literature had accuracy ranging from 61%-72% in the mild COPD cohort. The 2D airway-optimized tMPR model achieved 73% accuracy. The proposed 3D model of combining axial/coronal/sagittal images had an accuracy of 75%. The proposed model output combining 2D colored airways and inspiratory combined 3D images, and the 3D collage of axial/coronal/sagittal images, resulted in 74% and 73% accuracy, respectively. However, the output fusion of the airway-optimized tMPR and 3D lung model of combining axial/coronal/sagittal images reached the highest accuracy of 78%. While the CNN model with 2D airway/lung-optimized images had improved performance with reduced computational resources as compared to the 3D models proposed, as well as the other published CNN-based models, the combination of this 2D method with the 3D CNN model of combining axial/coronal/sagittal images achieved the highest performance in this mild cohort.