Development of a Deep Learning Model to Automatically Identify Palatal Landmarks on Three‐Dimensional Maxillary Dental Casts

Objectives: To develop a deep learning model to automatically identify palatal landmarks on three-dimensional (3D) digital maxillary dental casts, and to evaluate model performance. Materials and Methods: Twelve palatal landmarks were manually annotated on each 3D digital maxillary dental cast from 377 individuals in the permanent dentition stage. Manually annotated landmarks were used as ground truth to develop and to evaluate a deep learning model for automatic landmark detection. A two-stage PointNet++ architecture was employed, where coarse landmark predictions were first generated, followed by localized refinement for improved precision. The model accuracy was evaluated by measuring the linear discrepancy between the final predicted and the ground-truth landmark positions. Results: A PointNet++-based hierarchical deep learning model, designed to extract both local and global features from point clouds, was developed. The model demonstrated a mean landmark detection error of 0.55 mm (SD ± 0.49) between predicted and ground-truth positions across 12 landmarks. The model also exhibited high predictive performance, correctly predicting 90% of landmarks within 1 mm and 98% within 2 mm of the ground truth. Conclusions: A deep learning model was developed for automated identification of 12 palatal landmarks on 3D maxillary dental casts, which demonstrated high performance. Our model will enable more efficient morphological assessment of the palate by substantially reducing the time for manual annotation in clinical and research settings.