Automatic segmentation of craniomaxillofacial complex and mandible in dental CBCT images using a TransUNet-based deep learning model

Accurate segmentation of craniomaxillofacial structures from cone beam computed tomography (CBCT) is essential for digital dentistry, but conventional manual and semi-automated methods are time-consuming and operator-dependent. This study aimed to develop and validate a TransUNet-based deep learning model for automated, holistic segmentation of the craniomaxillofacial complex and the mandible, both inclusive of the associated dentition. A total of 211 CBCT scans were included. The internal cohort comprised a training set ( n = 81), a validation set ( n = 20), and an internal testing set ( n = 30), stratified by scanner model. An independent external testing set ( n = 80) was used to evaluate model performance outside the development setting. TransUNet was implemented as the primary segmentation model, while U-Net and SwinUNet were trained using the same development data and harmonized training protocol as comparator models. Model performance was assessed using Dice similarity coefficient (DSC), intersection over union (IoU), precision, recall; root mean square (RMS) was additionally evaluated internally. Time efficiency was recorded. Internal comparisons were performed using paired t tests, whereas external comparisons were conducted using Friedman tests followed by paired Wilcoxon signed-rank tests with Holm adjustment. The TransUNet model’s mean algorithmic inference time was 38.4 ± 3.5 s per scan. In the internal testing set, TransUNet achieved significantly closer agreement with the ground truth than semi-automated segmentation across all evaluation metrics for both anatomical structures ( p < 0.05), with mean DSCs of 0.949 ± 0.021 for the craniomaxillofacial complex and 0.970 ± 0.014 for the mandible. In the external testing set, TransUNet achieved significantly higher values than both U-Net and SwinUNet for all metrics in both structures (all Holm-adjusted p < 0.001). The TransUNet-based deep learning model offers an accurate, time-efficient solution for segmenting craniomaxillofacial complex and mandible, both inclusive of the associated dentition, in the testing cohorts.