3D Tooth Pose Estimation with Neighbor-Guided Neural Alignment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for tooth pose estimation in virtual 3D representations are labor-intensive, time-consuming, and prone to errors, especially when dealing with heavily ground teeth, and existing algorithms are inefficient and take too long to provide accurate results.
Innovation Solution
A computer-implemented method using a trained neural network to estimate tooth poses by segmenting a virtual 3D representation, determining initial poses based on geometric parameters of neighboring teeth, and normalizing the data for faster and more accurate tooth pose estimation, utilizing networks like PointNet for 3D point clouds.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual annotation of coordinate system axes is used for each tooth representation, then tooth pose accuracy is improved, but the process becomes labor- and time-intensive
Solution Approach 1:
The patent replaces manual mechanical annotation processes with an automated neural network system that processes 3D tooth models and predicts tooth poses algorithmically, eliminating the need for manual coordinate system annotation while maintaining accuracy
Solution Approach 2:
The system enables self-service by allowing the tooth pose estimation to be performed automatically through the neural network without requiring manual intervention, making the process autonomous and efficient
2Extent of automation
If Principal Component_analysis (PCA) algorithm is used for tooth pose estimation, then automated pose determination is achieved, but computational time increases to multiple seconds
Solution Approach 1:
The patent changes the algorithmic parameters by switching from PCA to a trained neural network model that has learned optimal pose estimation parameters during training, achieving faster computation while maintaining or improving accuracy
Solution Approach 2:
The neural network is pre-trained on a large dataset of tooth models with known poses, performing preliminary learning offline so that during actual use, pose estimation can be done rapidly without extensive real-time computation
3Extent of automation
If PCA algorithm is used for tooth pose estimation, then automated processing is achieved, but accuracy decreases for heavily ground teeth
Solution Approach 1:
The patent changes the estimation parameters by using a neural network that learns robust features from training data including heavily ground teeth, enabling accurate pose estimation for challenging cases where PCA fails
Solution Approach 2:
The neural network learns from copied examples in the training dataset, creating a model that has seen various tooth morphologies including ground teeth, allowing it to generalize accurately to new cases without being limited by algorithmic assumptions
Data Source
Figure 1
Figure 2
Figure 3A~3B
AI summary
According to an embodiment a computer-implemented method for tooth pose estimation is disclosed. The method comprises obtaining a virtual 3D representation (100) representing a patient's dentition, segmenting the virtual 3D representation (100) to obtain at least a first segmented tooth representation (101) and a second segmented tooth representation (105), wherein the at least first and second segmented tooth representation (101, 105) represent neighboring teeth in the patient's dentition. Furthermore, the method comprises determining an initial tooth pose (302) for the first segmented tooth representation (101) using a geometric parameter of the first segmented tooth representation (101) and a geometric parameter of the second segmented tooth representation (105). A normalized tooth representation (301) of the first segmented tooth representation (101) is obtained by transforming the first segmented tooth representation (101) using the initial tooth pose (302) for the first segmented tooth representation (101). The normalized tooth representation (301) of the first segmented tooth representation (101) is then input into a trained neural network (400). An output from the trained neural network (400) is produced, comprising a correct tooth pose for the first segmented tooth representation (101).