Accuracy and reliability of 3D cephalometric landmark detection with deep learning.

Background: Three-dimensional (3D) landmark detection is essential for assessing craniofacial growth and planning surgeries, such as orthodontic, orthognathic, traumatic, and plastic procedures. This study aimed to develop an automatic 3D landmarking model for oral and maxillofacial regions and to validate its accuracy, robustness and generalizability in both spiral computed tomography (SCT, 41 landmarks) and cone-beam computed tomography (CBCT, 14 landmarks) scans.

Methods: The model was implemented using an optimized lightweight 3D U-Net network architecture. Its accuracy, robustness and generalizability were thoroughly evaluated and validated through a multicenter retrospective diagnostic study. The model was trained and tested on a data set of 480 SCT and 240 CBCT cases. An additional inference on a different data set of 320 SCT and 150 CBCT cases was performed. Mean radial error (MRE) and success detection rate within 2-, 3-, and 4-mm error thresholds were measured as the primary evaluation metrics. Error analyses for landmark detection along each coordinate axis were performed. Consistency tests among observers were conducted.

Results: The average MRE for both SCT and CBCT was consistently below 1.3 mm and, notably, below 1.4 mm in complex conditions, such as malocclusion, missing dental landmarks, and the presence of metal artifacts. No significant differences in MRE and SDR at 2-4 mm were observed between external and internal SCT and CBCT sets. SCT bone landmarks were more precise than dental ones, with no difference between bone/soft tissue and dental/soft tissue. CBCT dental landmarks exhibited greater precision compared to bone landmarks. A detailed error analysis across the coordinate axes showed that the coronal axis had the highest error rates. The implementation of this model significantly improved the landmarking proficiency of senior and junior specialists by 15.9% and 28.9%, respectively, while also achieving a 6-9.5-fold acceleration in GUI interaction time.

Conclusions: This study shows that the AI-driven model delivers high-precision 3D localization of oral and maxillofacial landmarks, even in complex scenarios. The model demonstrates potential as a promising computer-aided tool to assist specialists in conducting accurate and efficient localization analyses; however, its robustness and generalizability require prospective clinical validation to ensure utility across varied experience levels.