Deep CNN for COPD identification by Multi-View snapshot integration of 3D airway tree and lung field

Chronic obstructive pulmonary disease (COPD) is a complex and irreversible respiratory disease with potential morphological abnormalities of the airway and lung fields. To date, whether and how these abnormalities can be used to identify COPD is unknown. This study developed a deep convolutional neural network (CNN) integrating the airway tree and lung field morphologies to identify COPD. We represent 3D airway and lung fields through multi-view 2D snapshots and their integration via deep CNN, to estimate the possibility of COPD. We constructed two datasets named Dataset 1 including 380 participants (190 COPD and 190 healthy controls) for training and validation and Dataset 2 including 201 participants (101 COPD and 100 healthy controls) for testing. First, the 3D airway tree and lung field are automatically extracted from computed tomography (CT) images, and 2D snapshots in nine views are captured. Second, the proposed ResNet-26 is trained with each view of snapshots as input. Finally, majority voting of nine models is performed to identify COPD. The accuracy (ACC) of the single-view ResNet-26 model (ventral, dorsal, and isometric view of airway; front, rear, left, right, top, and bottom view of lung field) is 0.900, 0.873, 0.889, 0.868, 0.824, 0.876, 0.861, 0.839, and 0.884, respectively. For the multi-view ResNet-26 model of airway tree and lung field, the ACC is 0.913 and 0.895, respectively. For the model integrating all nine views, the ACC eventually reaches as high as 0.947. The deep CNN model identifies COPD through integrating morphology of the airway tree and lung field extracted from CT images. A different view of 2D snapshots represents various characteristics of the 3D airway tree and lung field. The integration of multiple views can improve the performance of COPD prediction. The CNN model provides a potential method of identifying COPD via CT scans.