Segmentation of the upper airway using deep learning - nnUNet

OBJECTIVE: The objective of this study is to assess the performance of a deep learning-based framework for automated segmentation of the entire upper airway using CBCT and CT datasets. METHODS: A dataset comprised of 220 multi-source 3D images, including CBCT and CT from institutions in Canada, Chile, and France, covering both adult and pediatric scans, as well as pre- and post-operative scans, was used. A "one-center-out" validation was performed using CBCT scans from an Institution in India. Ground truth was established through manual segmentation using 3D Slicer. Deep learning models were implemented, focusing on a multi-source training approach with nnUNet. RESULTS: The multi-source model achieved a high average Dice score of 0.962. Furthermore, the nnUNet-155 model demonstrated an absolute volume difference of 3.31% compared to manual segmentation, with a prediction time of only 5 minutes per volume. CONCLUSIONS: This robust, efficient, and generalizable deep learning model provides a valuable tool for clinicians and researchers, enabling precise and consistent 3D analysis of the upper airway to support clinical decision-making and research across a wide range of patient demographics and imaging modalities. CLINICAL SIGNIFICANCE: This study introduces a robust and efficient deep learning model for automated 3D upper airway segmentation. This tool is clinically relevant as it provides clinicians with a precise, consistent, and time-saving method for analyzing the upper airway. By overcoming the limitations of manual segmentation-namely, labor intensity and inter-observer variability-this model facilitates more reliable clinical decision-making and research across diverse patient populations and imaging modalities.