Automatic 3D Facial Symmetry Reference Plane Construction Based on Facial Planar Reflective Symmetry net

Objectives: This study aimed to develop a novel deep-learning model to construct an symmetry reference plane (SRP) intelligently, called the facial planar reflective symmetry net (FPRS-Net). We also aimed to establish a method for defining a three-dimensional (3D) point-cloud region of interest (ROI) and high-dimensional feature computations suitable for this network model.
Methods: Four models of the FPRS-Net were prepared, each using a supervised and unsupervised learning approach based on full facial and ROI data (FPRS-Net_S, FPRS-Net_SR, FPRS-Net_U, and FPRS-Net_UR). These models were trained on 160 3D facial datasets, validated in 20 cases, and tested on another 20 cases. Model predictions were evaluated using an additional 40 clinical 3D facial datasets by comparing the angle error between the predicted plane and the ground truth plane.
Results: The FPRS-Net_SR and FPRS-Net_U models achieved an average angle error of less than 1° in predicting the 3D facial SRP. The successful detection rates of the FPRS-Net_SR and FPRS-Net_U models exceeded 90%. The application of the four FPRS-Net models to create an SRP in 40 cases of 3D facial data took less than 4 s.
Conclusions: Our study demonstrates a new solution for the intelligent construction of oral clinical 3D facial SRPs, improving diagnostic and therapeutic efficiency and effectiveness.
Clinical significance: This study proposes an innovative deep learning algorithm (FPRS-Net) to construct an symmetry reference plane (SRP), which can reduce workload, shorten the time required for digital design, reduce dependence on expert experience, and improve therapeutic efficiency and effectiveness in dental clinics.