[Intelligent generation of occlusal surface morphology of dental crown using an octree-based diffusion model]

Objective: To explore a three-dimensional point cloud completion method based on an octree diffusion model for intelligent generation of personalized dental crown occlusal morphology, and to evaluate the similarity and functional adaptability of the generated crown occlusal surfaces compared with natural teeth. Methods: A total of 780 intraoral scan datasets of unilateral maxillary and mandibular second premolars to second molars were selected from the DCPR-GAN database (https://github.com/Sukhum169/DCPR-GAN). The data were obtained from Peking University School and Hospital of Stomatology and Nanjing Stomatological Hospital, involving patients aged 35-50 years. Among them, 720 cases were used for model training and 60 for testing. An octree-structured diffusion model was constructed to achieve two stages of three-dimensional crown morphology generation: unconditional generation and mask-conditioned generation. In the unconditional generation stage, Earth mover's distance (EMD) was used as the metric for morphological discrepancy, and the occlusal contact areas of generated crowns (DM-U) were compared with those of natural teeth. Statistical analysis was performed using the Kruskal-Wallis test. In the conditional generation stage, clinical intraoral scan data were collected from 12 healthy volunteers at Peking University School and Hospital of Stomatology between September 2024 and February 2026, including six males and six females aged 22-24 years. Crowns were generated for the diffusion model generated crown with the mask condition group (DM group) and compared among four groups: natural teeth (NA group), technician-designed crowns (TE group), and automatically generated crowns (AT group). Three-dimensional coordinate deviations and euclidean distance errors of five anatomical landmarks were analyzed. Root mean square error (RMSE) of the three-dimensional surface was calculated using natural teeth as the gold standard. Statistical analysis was performed using the Friedman test. Results: In the unconditional generation evaluation, EMD results showed a coverage rate of 50.88%, a minimum matching distance (MMD) of 1.74, and a 1-nearest neighbor accuracy (1-NNA) of 0.504, indicating high similarity between generated crowns and natural teeth in feature space. Comparison of occlusal contact area showed no statistically significant difference between the DM and NA groups (χ²=0.75, P=0.387). In the conditional generation evaluation, the RMSE values for the TE, AT, and DM groups were 0.39 (0.11), 0.37 (0.13), and 0.31 (0.10) mm, respectively, with no significant overall difference among the three groups (χ²=3.50, P=0.174). Conclusions: The three-dimensional point cloud completion method based on an octree diffusion model can effectively learn the occlusal morphology of natural crowns and achieve personalized reconstruction of crown occlusal surfaces. The generated crown occlusal morphology demonstrated similarity to natural teeth and functional adaptability comparable to technician-designed methods.