[Deep learning algorithms for intelligent construction of a three-dimensional maxillofacial symmetry reference plane]

Area of Science:

• Computer Vision
• Medical Imaging
• Biomedical Engineering

Context:

• Accurate registration of 3D maxillofacial point cloud data is crucial for digital dental applications.
• Existing methods may lack efficiency or precision in establishing maxillofacial symmetry.
• Deep learning offers potential for automated and intelligent registration solutions.

Purpose:

• To develop an original-mirror alignment deep learning algorithm (MDGR-Net) for intelligent registration of 3D maxillofacial point cloud data.
• To utilize a dynamic graph-based registration network for calculating rotation and translation matrices.
• To establish a novel methodology for constructing 3D maxillofacial symmetry reference planes.

Summary:

• The maxillofacial dynamic graph registration network (MDGR-Net) was developed using 2,000 3D maxillofacial datasets.
• The algorithm constructs feature vectors, establishes correspondences, and calculates transformation matrices for point cloud registration.
• Principal Component Analysis (PCA) was applied to derive the symmetry reference plane, with evaluation using R-squared and angle error metrics.

Impact:

• MDGR-Net achieved high accuracy (R²=0.91 for rotation, R²=0.98 for translation) in registration.
• The method rapidly constructs 3D maxillofacial symmetry reference planes (3 seconds per case).
• This approach enhances diagnostic and therapeutic efficiency in clinical dental applications, particularly for specific malocclusion types.