Stand‐alone graph‐based lung airway labeling with quantitative validations

BACKGROUND: Previous studies have shown that protecting airways connected to high-functioning lung parenchyma reduces the risk of post-treatment radiation-induced lung injury. To account for airway sparing in radiotherapy planning, a critical requirement is to characterize the geometry and radiation response of individual airway segments. PURPOSE: To report on automated individual airway labeling framework with branch dimensions extracted from CT-derived virtual bronchoscopy images. METHODS: We developed an automated graph-based topological labeling algorithm with bifurcation detections and context-based denoising. In-house modules for estimating airway diameters and tube length were also developed. Both labeling and meta-information estimation were validated against a commercial virtual bronchoscopy software in data from 20 lung cancer patients. Labeling metrics were segmentation surface scores: i.e., average symmetric surface distance (ASSD), normalized surface distance (NSD), and 95th percentile surface Hausdorff (HD95surf). Meta-information metrics were absolute errors of branch major diameter, minor diameter, and tube length. Mean absolute error (MAE) and root mean square error (RMSE) were calculated to assess meta-information deviations in both validations. RESULTS: Mean and standard deviations were 0.56 ± 0.21 mm for ASSD, 87.66 ± 3.20% for NSD, and 2.51 ± 0.68 mm for HD95surf. Deviations of meta-information metrics in labeling validation were MAE/RMSE: 0.33 ± 0.09/0.67 ± 0.16 mm for major diameter, 0.24 ± 0.07/0.53 ± 0.15 mm for minor diameter, and 3.27 ± 0.82/5.53 ± 1.53 mm for tube length. Deviations of meta-information metrics in meta-information validation were MAE/RMSE: 1.81 ± 0.30/2.31 ± 0.41 mm for major diameter, 1.25 ± 0.27/1.14 ± 0.26 mm for minor diameter, and 2.20 ± 0.32/3.24 ± 0.49 mm for tube length. CONCLUSION: The proposed open-source framework provides stand-alone, fully-automated airway labeling with very good agreement with a commercial virtual bronchoscopy system. Our method shows strong potential for use in functionally-guided lung sparing in radiation treatment plan optimization.