This master’s thesis from the University of Ulsan presents a deep learning model for automatic segmentation and labeling of 18 segmental bronchi on chest CT in patients with pulmonary disease. To address the time-consuming nature and inter-observer variability of manual labeling, a nnU-Net V2-based automatic segmentation model was developed, analyzing the impact of combining binary masks and distance maps as input features. Multi-institutional CT data from 370 patients with pulmonary diseases including asthma and COPD were used, with ground-truth labels established through cross-validation by two expert radiologists. Among five 3D nnU-Net model configurations, Model 5—integrating original CT, binary masks, and distance maps—achieved the best performance with internal validation DSC of 0.781±0.071 and external validation DSC of 0.693±0.083, effectively reducing false positive and wrong location errors. AVIEW COPD (Coreline Soft, Seoul, Korea) was utilized for CT image processing. This study contributes to improving the clinical applicability of automatic segmental bronchi segmentation.
목록으로
