Development of Artificial Intelligence-Supported Automatic Three-Dimensional Surface Cephalometry.

Objective: Surface-based three-dimensional (3D) cephalometry provides detailed clinical information for the analysis of craniofacial structures. This study aimed to develop an automated 3D surface cephalometry system using mesh fitting based on landmarks identified by artificial intelligence (AI) and to evaluate its accuracy.

Methods: A total of 185 CBCT images from adult Japanese patients (system training, n = 152; evaluation, n = 33) were used in this study. Cranial and mandibular images were generated via surface rendering of CBCT images. An experienced orthodontist manually recognised 19 and 45 3D landmarks for the cranium and mandible, respectively, and used them as the gold standard after they were checked by another experienced orthodontist. An AI system developed using PointNet ++ was trained to output landmark coordinates based on surface data and normal vectors. Mesh fitting (homologous modelling) was then conducted using the AI-identified landmarks. The errors in mesh fitting were evaluated.

Results: The mean errors for wire mesh fittings with AI-identified landmarks for the maxilla and mandible were 0.80 ± 0.57 mm and 1.45 ± 0.34 mm, respectively.

Discussion: An AI-based landmark identification system and mesh fittings that demonstrate clinically acceptable accuracy were presented. This system can be applied in clinical settings to quantify and visualise craniofacial structures in three dimensions.

Conclusion: The automated 3D surface cephalometry system utilising mesh fitting based on AI-identified landmarks showed clinically acceptable accuracy. This allows orthodontists to compare a patient's craniofacial surface with normative data, without the need for manual landmark identification.