OBJECTIVES: The purpose of this study was to propose an automatic landmark identification method using curvature to improve the reproducibility of landmark identification and compare its performance with that of a previously established method. METHODS: A total of 30 patients with facial deformities associated with mandibular prognathism were included. Computed tomography (CT) images were utilized to construct three-dimensional (3D) surface models, followed by an analysis of their surface curvature distribution. A statistical shape model (SSM) was created as a deformable mean model to identify the six landmarks. These landmarks were automatically identified in each patient model by registering the SSM in the individual patient models. Two registration methods were employed: the proposed curvature-based and previously established methods. Both methods involved rigid and non-rigid registration processes
however, the proposed method included additional curvature-based registration using a curvature-driven, non-rigid Iterative Closest Point (ICP) algorithm. The Euclidean distances between the manually and automatically identified landmarks were measured and compared between the two methods. RESULTS: The Euclidean distance was significantly lower in the gonion and right coronoid process when the proposed method was used compared to the previous method. No significant differences were observed in the condylion or left coronoid process. CONCLUSIONS: These findings suggest that the curvature-based registration successfully automates landmark identification on 3D mandibular images, providing higher accuracy in convex regions and improved reproducibility in landmark identification.