Automated Computed Tomography Segmentation of the Pharyngeal Airway and Palate to Accelerate Tübingen Palatal Plate Fabrication in Pierre Robin Sequence

Abstract

Infants with Pierre Robin sequence frequently develop upper airway obstruction due to micrognathia and glossoptosis. The Tübingen palatal plate repositions the tongue base anteriorly to improve airway patency; however, conventional fabrication requires serial intraoral impressions and repeated nasoendoscopy, which prolongs airway compromise. Computed tomography (CT) enables single-session virtual spur design, yet manual pharyngeal airway and hard palate segmentation is labor-intensive, delaying treatment. The authors evaluated convolutional neural network–based automated segmentation to accelerate CT-guided Tübingen palatal plate fabrication using 74 low-dose head-and-neck CT scans (50 pre-contrast, 24 phonation) annotated retrospectively by 3 raters. Two-dimensional and 3-dimensional U-net models were trained with 5-fold cross-validation; ablation experiments compared cropping versus resizing; sagittal, coronal, versus axial planes; multiclass versus one-versus-rest strategies; and batch splitting. Primary outcome: dice similarity coefficient (DSC); secondary outcomes: inference time and contouring time saved. The 2-dimensional U-net achieved the best accuracy-efficiency balance, with mean DSC 0.8835 (palate 0.8741; airway 0.8928). Cropping improved sagittal DSC from 0.8584 to 0.8690. Multiclass and one-versus-rest DSC were comparable; semi-supervised pretraining conferred minimal benefit. Inference required <60 seconds on a single graphics processing unit, reducing manual contouring by approximately 25 minutes per patient and enabling same-day computer-aided design/computer-aided manufacturing printing. Automated CT segmentation eliminates a major clinical bottleneck, supporting faster, safer, and more personalized airway management for Pierre Robin sequence infants, and warrants prospective validation.