目的: 建立一种可实现三维颌面点云数据智能配准的本体-镜像关联深度学习算法，基于颌面动态图结构的配准网络(maxillofacial dynamic graph registration network，MDGR-Net)模型，实现三维颌面对称参考平面的自动化构建，以期为口腔临床数字化设计与分析提供参考。方法: 收集2018年10月至2022年10月就诊于北京大学口腔医院无显著颌面畸形临床患者400例，通过数据增强的方式获得2 000例三维颌面数据用于MDGR-Net算法训练与测试，其中训练集1 600例、验证集200例、内部测试集200例，MDGR-Net模型包含构造本体与镜像点云(X和Y)中关键点的特征向量，基于特征向量获取点云X和Y中关键点的对应关系，以及通过奇异值分解(singular value decomposition，SVD)计算旋转和平移矩阵R，t。基于MDGR-Net模型实现本体点云与镜像点云的智能配准，获得本体-镜像联合点云，并采用主成分分析(principal component analysis，PCA)算法获得MDGR-Net关联法对称参考平面。基于决定系数(coefficient of determination，R²)指标对内部测试集平移及旋转矩阵进行模型评价，并对200例内部测试集与40例外部测试集临床数据，基于MDGR-Net关联法与“真值”迭代最近点(iterative closest point，ICP)关联法构建的三维颌面对称参考平面进行角度误差评价。结果: 基于200例内部测试集三维颌面数据测试MDGR-Net旋转矩阵R²为0.91，平移矩阵R²为0.98。在内部与外部测试集上，角度误差平均值分别为0.84°±0.55°、0.58°±0.43°，临床构建40例三维颌面对称参考平面仅需3 s，在正畸骨性Ⅲ类、高角、安氏Ⅲ类错牙合畸形受试者表现最佳。结论: 基于点云智能配准的MDGR-Net关联法为口腔临床三维颌面对称参考平面构建提供了新的解决方案，可显著提升诊疗效率和效果，降低专家依赖性。

Objective: To develop an original-mirror alignment associated deep learning algorithm for intelligent registration of three-dimensional maxillofacial point cloud data, by utilizing a dynamic graph-based registration network model (maxillofacial dynamic graph registration network, MDGR-Net), and to provide a valuable reference for digital design and analysis in clinical dental applications. Methods: Four hundred clinical patients without significant deformities were recruited from Peking University School of Stomatology from October 2018 to October 2022. Through data augmentation, a total of 2 000 three-dimensional maxillofacial datasets were generated for training and testing the MDGR-Net algorithm. These were divided into a training set (1 400 cases), a validation set (200 cases), and an internal test set (200 cases). The MDGR-Net model constructed feature vectors for key points in both original and mirror point clouds (X, Y), established correspondences between key points in the X and Y point clouds based on these feature vectors, and calculated rotation and translation matrices using singular value decomposition (SVD). Utilizing the MDGR-Net model, intelligent registration of the original and mirror point clouds were achieved, resulting in a combined point cloud. The principal component analysis (PCA) algorithm was applied to this combined point cloud to obtain the symmetry reference plane associated with the MDGR-Net methodology. Model evaluation for the translation and rotation matrices on the test set was performed using the coefficient of determination (R²). Angle error evaluations for the three-dimensional maxillofacial symmetry reference planes were constructed using the MDGR-Net-associated method and the "ground truth" iterative closest point (ICP)-associated method were conducted on 200 cases in the internal test set and 40 cases in an external test set. Results: Based on testing with the three-dimensional maxillofacial data from the 200-case internal test set, the MDGR-Net model achieved an R² value of 0.91 for the rotation matrix and 0.98 for the translation matrix. The average angle error on the internal and external test sets were 0.84°±0.55° and 0.58°±0.43°, respectively. The construction of the three-dimensional maxillofacial symmetry reference plane for 40 clinical cases took only 3 seconds, with the model performing optimally in the patients with skeletal Class Ⅲ malocclusion, high angle cases, and Angle Class Ⅲ orthodontic patients. Conclusion: This study proposed the MDGR-Net association method based on intelligent point cloud registration as a novel solution for constructing three-dimensional maxillofacial symmetry reference planes in clinical dental applications, which can significantly enhance diagnostic and therapeutic efficiency and outcomes, while reduce expert dependence.