Tooth3dNet: A preliminary exploration for automatic 3D morphology design of dental crowns with a deep generative network.

OBJECTIVES: Artificial intelligence is opening new avenues for digital dental restoration. To advance the automatic design of dental crowns, this study proposed a novel deep generative network to enable automatic generation of 3D crown shapes for various types of missing teeth-from incisors to molars-based solely on the surrounding oral environment. METHODS: The proposed network, named Tooth3dNet, adopted a Transformer-based encoder-decoder architecture, featuring specially designed optimized query generation and a critical incremental point cloud reconstruction module. It was trained directly on a self-created large-scale 3D intraoral scan point cloud dataset. ANOVA and paired t-test statistical analysis were implemented to compare the performance of the proposed network and other state-of-the-art methods. RESULTS: Multi-metric evaluations showed that our network achieved a mean reconstruction error of 0.200 mm and a Hausdorff distance of 0.683 mm, outperformed current state-of-the-art point cloud generation networks significantly (P < 0.05). The generated point clouds produced reconstructed surfaces that retained the anatomical features of the teeth. CONCLUSIONS: Our method demonstrates good performance in generating primary 3D morphology of dental crowns. CLINICAL SIGNIFICANCE: Although the present method is limited to occlusal surface generation and some areas of the proximal surfaces, it introduces a novel approach for crown design and lays the groundwork for future development of fully automated, full-crown morphology generation.