A systematic review of advanced techniques for automated segmentation of the human upper airway in volumetric imaging

Abstract

Background: The human upper airway, comprising the nasal cavity, pharynx, and larynx, plays a vital role in respiratory and diagnostic processes. Volumetric imaging techniques, such as Computed Tomography (CT), Magnetic Resonance Imaging (MRI), and cone-beam computed tomography (CBCT) of the upper airway pose challenges in manual segmentation due to the complexity and variability of anatomical structures. Automated segmentation techniques, particularly those leveraging deep learning, have emerged as promising solutions to enhance accuracy, reproducibility, and clinical utility. Objective: This systematic review evaluates recent advancements in automated segmentation techniques for the human upper airway in volumetric imaging and assesses their clinical relevance. Methods: A comprehensive search of PubMed, Scopus, IEEE Xplore, Web of Science, and Embase was conducted for studies published between 2019 and 2024. Eligibility criteria included studies focusing on upper airway segmentation using imaging modalities like CT, MRI, and CBCT with measurable performance metrics. Performance, dataset size, computational efficiency, and clinical applicability were systematically analyzed. Results: Ten studies met the inclusion criteria. Advanced deep learning techniques, such as Mask R-CNN and DeepLabV3+, demonstrated superior performance (DSC: 0.95-0.96) compared to traditional methods like region-growing and thresholding (DSC: 0.84-0.85). Hybrid models, which integrate traditional algorithms with machine learning, showed improvements in segmentation accuracy and computational efficiency. Despite these advancements, challenges remain, including high computational demands, reliance on large annotated datasets, and limited integration into clinical workflows. Conclusions: Deep learning models excel in segmenting the upper airway's complex anatomy but require optimization to overcome computational and data-related challenges for clinical adoption. Future directions include leveraging multimodal imaging, improving computational efficiency, and developing interpretable models to enhance segmentation accuracy and clinical usability.