Deep Learning-Based Segmentation of Airway Morphology from Endobronchial Optical Coherence Tomography

Abstract

Background: Manual measurement of endobronchial optical coherence tomography (EB-OCT) images means a heavy workload in the clinical practice, which can also introduce bias if the subjective opinions of doctors are involved. Objective: We aim to develop a convolutional neural network (CNN)-based EB-OCT image analysis algorithm to automatically identify and measure EB-OCT parameters of airway morphology. Methods: The ResUNet, MultiResUNet, and Siamese network were used for analyzing airway inner area (Ai), airway wall area (Aw), airway wall area percentage (Aw%), and airway bifurcate segmentation obtained from EB-OCT imaging, respectively. The accuracy of the automatic segmentations was verified by comparing with manual measurements. Results: Thirty-three patients who were diagnosed with asthma (n = 13), chronic obstructive pulmonary disease (COPD, n = 13), and normal airway (n = 7) were enrolled. EB-OCT was performed in RB9 segment (lateral basal segment of the right lower lobe), and a total of 17,820 OCT images were collected for CNN training, validation, and testing. After training, the Ai, Aw, and airway bifurcate were readily identified in both normal airway and airways of asthma and COPD. The ResUNet and the MultiResUNet resulted in a mean dice similarity coefficient of 0.97 and 0.95 for Ai and Aw segmentation. The accuracy Siamese network in identifying airway bifurcate was 96.6%. Bland-Altman analysis indicated there was a negligible bias between manual and CNN measurements for Ai (bias = −0.02 to 0.01, 95% CI = −0.12 to 0.14) and Aw% (bias = −0.06 to 0.12, 95% CI = −1.98 to 2.14). Conclusion: EB-OCT imaging in conjunction with ResUNet, MultiResUNet, and Siamese network could automatically measure normal and diseased airway structure with an accurate performance.