Automated Dental Landmarks Recognition in Special Models via Neural Network

Background: Accurate dental landmark recognition is essential in orthodontics and prosthodontics applications but can be challenged by malocclusion and tooth wear. This study expends automated landmark recognition to include such complex cases, aiming to improve anatomical analysis and establish a treatment evaluation framework for Chinese patients. Methods: The study incorporated a dataset of 919 pairs of pretreatment intraoral scan (IOS) models. Initially, a subset of 100 IOS models was segmented and labeled using MeshLab software to generate 3-D dental grid data. Subsequently, 500 IOS models were meticulously labeled by three researchers to train and develop a neural network termed multiscale automated landmark recognition. This network reformulates the task of dental landmark recognition as a geodesic distance field problem on the tooth surface, effectively capturing both local and global geometric features essential for precise detection. The remaining 319 IOS models were then subjected to both manual labeling and automated recognition using the trained network, and the consistency between the two methods was evaluated. Finally, arch width measurements were compared between the two recognition approaches. Results: The neural network shows high concordance with manual labeling, with a mean Euclidean distance discrepancy of 0.375 mm, and no significant difference in arch width measurements between the two methods (P > 0.05). Conclusion: Based on the 0.500 mm clinically acceptable difference standard defined by the ABO objective grading system, the neural network developed in this study demonstrates superior performance in the automated recognition of IOS models exhibiting malocclusion and tooth wear.