摘要：

目的： 开发基于深度学习（deep learning，DL）的三维点云重建网络，利用上颌同侧侧切牙、尖牙、双尖牙形态实现上颌中切牙的个性化三维重建，以期为前牙美学修复提供参考。方法： 收集192例口腔扫描模型，利用Exocad软件分割11~14牙冠，经MATLAB软件标准化坐标系后，随机选取182例作为训练集、10例作为测试集，训练形态估计与位姿估计的双网络，以重建上颌中切牙三维点云。将输出结果网格化，与原始牙冠拟合对齐，评价重建效果。构建11牙冠切端1/3缺损模型，对比镜像法、本算法及标准数据库法的修复效果。结果： 测试集重建结果显示：倒角距离（chamfer distance，CD）为0.405，地球移动距离（earth mover's distance，EMD）为0.152，均方根误差（root mean square error，RMSE）为（0.128±0.030） mm。上颌中切牙缺损修复中，本研究算法[RMSE为（0.128±0.030） mm]与镜像法[RMSE为（0.130±0.021） mm]修复效果相当，但2种方法均显著优于数据库法[RMSE为（0.233±0.038） mm，P<0.001]。结论： 本研究提出的三维点云重建网络可基于邻牙形态高精度重建上颌中切牙，为个性化前牙修复提供可靠技术路径。

关键词: 深度学习, 牙冠设计, 上颌中切牙

Abstract:

Objective: To develop a deep learning (DL)-based three-dimensional (3D) point cloud reconstruction network that utilizes the morphological features of the ipsilateral maxillary lateral incisor, canine, and premolars to predict and reconstruct the anatomical morphology of the maxillary central incisor, providing insights for personalized anterior tooth aesthetic restoration. Methods: A total of 192 intraoral scan models were collected. Exocad software was used to segment crowns of teeth #11-14. After standardizing coordinate systems in MATLAB, 182 cases were randomly selected as the training set and 10 were the test set. A dual-network architecture (morphology and pose estimation) was trained to reconstruct maxillary central incisor point clouds. Reconstructed outputs were meshed, aligned with original crowns for evaluation, and applied to incisal third defect models of tooth #11 to compare restoration outcomes among the proposed algorithm, mirroring technique, and standard database method. Results: The test set achieved a chamfer distance (CD) of 0.405, earth mover's distance (EMD) of 0.152, and a root mean square error (RMSE) of (0.128±0.030) mm. For incisal defect restoration, the proposed algorithm [RMSE：(0.128±0.030) mm] demonstrated comparable accuracy to mirroring technique [RMSE: (0.130±0.021) mm], but significantly outperformed the database method [RMSE: (0.233±0.038) mm, P<0.001]. Conclusion: The proposed 3D point cloud reconstruction network enables high-precision maxillary central incisor restoration based on adjacent tooth morphology, offering a reliable technical solution for personalized anterior dental rehabilitation.

Key words: deep learning, dental crown design, maxillary central incisor

中图分类号: 

• R782