Objectives: Cone-beam computed tomography (CBCT) is extensively used in dentistry and maxillofacial surgery for three-dimensional imaging with relatively low radiation doses. However, low soft tissue contrast and artifacts, particularly from fixation devices, make accurate skin segmentation difficult. Manual segmentation remains the gold standard but is time-consuming. Existing automated approaches often fail to differentiate between skin surfaces, fixation devices, and background noise. This study aimed to develop SkinNet-CBCT, a deep learning-based model using the nnU-Net framework, for fully automated and robust skin segmentation in CBCT scans. Methods: A total of 262 CBCT scans were compiled from diverse open-access datasets (CTooth, CTooth+, and ToothFairy), encompassing various imaging conditions and anatomical features. Ground truth labels were created through manual segmentation using 3D Slicer. Data were divided into training (80 cases), validation (20 cases), and testing (162 cases) sets. SkinNet-CBCT was trained and optimized using the nnU-Net architecture. Model performance was evaluated using the Dice Similarity Coefficient (DSC), 95th percentile Hausdorff Distance (HD95), and Average Surface Distance (ASD), with emphasis on accuracy and robustness in differentiating soft tissue from fixation devices. Results: SkinNet-CBCT achieved near-perfect segmentation performance, with a mean DSC of 0.999 ± 0.0007, HD95 of 0.166 ± 0.093 mm, and ASD of 0.034 ± 0.027 mm. The model effectively distinguished skin boundaries from fixation devices, minimizing false segmentation and maintaining high geometric consistency across datasets. Conclusions: The SkinNet-CBCT model enables highly accurate, automated skin segmentation in CBCT images while reliably differentiating fixation devices from soft tissue. This advancement substantially reduces manual annotation time, supports clinical workflow efficiency, and facilitates secure, anonymized CBCT data sharing for research and AI-driven diagnostics in digital dentistry.
Minh et al. (Sun,) studied this question.