Artificial intelligence-based upper airway segmentation for evaluating volume changes following genioplasty in patients with obstructive sleep apnea

Background : Obstructive sleep apnea (OSA) is strongly linked to upper airway anatomical compromise, with mandibular retrognathia being a risk factor. Genioplasty is commonly performed for both aesthetic and functional enhancement. Convolutional neural networks (CNNs) enable reliable segmentation of cone-beam computed tomography (CBCT) images. This study aimed to evaluate AI-based upper airway segmentation from CBCT in OSA patients who underwent genioplasty. Methods : A total of 170 CBCT images were utilized, divided into a training/validation set (n=110) and a test set (n=60). The test set consisted of 30 matched preoperative(T0) and postoperative(T1) image pairs from OSA patients with microgenia who underwent advancement sliding genioplasty. A SegResNet CNN model was employed for fully AI-based segmentation of subregional upper airway volumes, with performance assessed via dice similarity coefficient (DSC), volume similarity (VS), and 95 percentile Hausdorff Distance (95% HD). Correlations between clinical indicators, volume changes, and model metrics were analyzed. Results : The model exhibited a mean DSC value of 0.900-0.907, a mean VS value of 0.949-0.950 and a mean 95%HD of 1.485-1.588. Postoperatively, both subregions showed significant volume increases (velopharynx: 8888.19 ± 3106.34 vs. 10615.96 ± 3501.67; oropharynx: 6330.92 ±3218.49 vs. 7905.11 ± 4413.17, p<0.05), and oropharyngeal expansion weakly correlated with chin advancement magnitude. Conclusions : The present SegResNet-based model achieved fast and accurate upper airway segmentation from pre- and postoperative CBCT scans of OSA patients underwent genioplasty, establishing a basis for developing efficient analytical models to predict surgical outcomes for OSA patients. Clinical trial number : not applicable.