Intelligent Construction Of 3D Maxillofacial Symmetry Reference Plane

With the widespread clinical application of digital dentistry, the automated construction of three-dimensional (3D) maxillofacial features has become a growing trend. This study aims to develop a deep learning model, the Maxillofacial Planar Reflective Symmetry Network (MPRS-Net), to achieve the automated construction of the 3D maxillofacial symmetry reference plane. A total of 400 clinical patients without significant maxillofacial deformities were included. Experts manually annotated the 3D maxillofacial symmetry reference planes which served as ground truth for supervised learning in the MPRS-Net model. PointNet was used to extract global features from these unordered point clouds. Additionally, an attention mechanism was introduced, to capture mathematical features highly correlated with the symmetry reference plane. Finally, the model outputted symmetry plane parameters to construct the 3D maxillofacial symmetry reference plane. To evaluate the model's performance, MPRS-Net was tested on an internal dataset of 40 cases, calculating the angle between the predicted and ground-truth symmetry planes. Furthermore, an external dataset of 40 cases from three independent institutions was used to evaluate the generalization and robustness. For the internal test set, the mean squared error (MSE) of the symmetry plane prediction was 0.22. Across both internal and external test sets, the average angle error was 0.65°± 0.38°. This study proposes a novel deep learning algorithm for the intelligent construction of 3D maxillofacial symmetry reference planes. The proposed method significantly enhances diagnostic efficiency and accuracy, reduces reliance on expert annotations, and improves overall clinical workflow.