ACCURACY AND GENERALIZABILITY OF AN OPEN-SOURCE DEEP LEARNING MODEL FOR FACIAL BONE SEGMENTATION ON CT AND CBCT SCANS: AN EX VIVO STUDY.
Authors
Affiliations (3)
Affiliations (3)
- Department of Orthodontics and Dentofacial Orthopedics, School of Dental Medicine, University of Bern, CH-3010 Bern, Switzerland. Electronic address: [email protected].
- Department of Orthodontics and Dentofacial Orthopedics, School of Dental Medicine, University of Bern, CH-3010 Bern, Switzerland; Jeddah Second Health Cluster, 23816 Jeddah, Saudi Arabia.
- Université Paris Cité, Université Sorbonne Paris Nord, Inserm UMR 1333 Oral Heath, F-92120 Montrouge, France; Service Médecine Bucco-Dentaire, AP-HP, Hôpital Pitié-Salpêtrière, F-75013 Paris, France. Electronic address: [email protected].
Abstract
To evaluate the accuracy and generalizability of DentalSegmentator, an open-source deep learning tool, for automated reconstruction of facial skeletal surface models from computed tomography (CT) scans acquired under different ex vivo imaging conditions. Ten human dry skulls were scanned using one CT scanner and three cone beam CT (CBCT) protocols, including an ultra-low-dose protocol, on two CBCT devices. High-accuracy reference surface models were acquired using an optical scanner. CT and CBCT scans were automatically segmented using the open-source DentalSegmentator model and reconstructed as 3D surface models. Three facial regions (forehead, zygomatic process, and maxillary process) were defined for quantitative assessment. Accuracy was measured as the mean absolute distance (MAD) and the standard deviation of absolute distances (SDAD) between segmented and reference models after best-fit superimposition. Repeated segmentations were identical, confirming perfect reproducibility. Across all acquisition settings and regions, DentalSegmentator produced highly accurate skeletal surface models, with an overall median MAD of 0.088 mm (IQR 0.073) and SDAD of 0.061 mm (IQR 0.028). Small but significant differences were detected between imaging systems (MAD: p < 0.001; SDAD: p = 0.003), with CT scans showing slightly lower trueness than CBCT images. DentalSegmentator produced accurate facial skeletal surface models across diverse CT and CBCT settings, demonstrating excellent generalizability, including under low-radiation conditions. Minor differences in trueness between imaging systems were small and unlikely to affect clinical or research applications. Deep learning provides a robust foundation for automated 3D craniofacial surface extraction, supporting broader adoption of AI-driven workflows in clinical and research settings.