Artificial intelligence applications in CBCT-based assessment of craniofacial airway volume and shape in sleep-disordered breathing: a systematic review

Background Sleep-disordered breathing (SDB), including obstructive sleep apnea (OSA), is strongly influenced by craniofacial airway morphology. Cone-beam computed tomography (CBCT) enables three-dimensional airway assessment; however, manual segmentation is time-consuming and subject to variability. Artificial intelligence (AI) has emerged as a potential tool to automate airway analysis. This systematic review evaluates the technical performance and clinical applicability of AI in CBCT-based craniofacial airway assessment for SDB. Methods This review was conducted in accordance with PRISMA 2020 guidelines and prospectively registered in PROSPERO (CRD 420251065028). A comprehensive search of PubMed, Ovid, Scopus, and Google Scholar was performed for studies published from 2015 onward. Eligibility criteria included diagnostic or validation studies assessing AI-based CBCT airway segmentation, landmark detection, or OSA prediction. Dual independent screening was conducted, and risk of bias was assessed using QUADAS-2. Results Fourteen studies met the inclusion criteria. Most implemented deep learning architectures, particularly convolutional neural networks-based models such as U-Net and SpatialConfiguration-Net. Reported Dice Similarity Coefficients ranged from 0.90 to 0.97, and intraclass correlation coefficients for volumetric measurements typically exceeded 0.90, indicating strong agreement with manual segmentation. AI systems substantially reduced annotation time compared with manual methods. However, heterogeneity was considerable across AI architectures, anatomical targets, datasets, and outcome metrics. External validation was limited, and only one study evaluated OSA prediction against polysomnography. Conclusions AI demonstrates high technical accuracy and efficiency in CBCT-based airway segmentation and morphometric analysis. Nevertheless, current evidence largely reflects internal technical validation rather than confirmed clinical utility. Additional prospective, multicenter studies incorporating clinical outcome validation are required before routine clinical implementation can be fully supported.