Differentiable Collision-Supervised Tooth Arrangement Network With a Decoupling Perspective.

Area of Science:

• Digital dentistry
• Orthodontics
• Computer-aided design

Background:

• Digital orthodontic planning relies on accurate tooth arrangement.
• Current methods couple pose perception and motion regression, leading to poor 3D transformation accuracy.
• Existing methods fail to prevent unacceptable teeth overlaps or gaps.

Purpose of the Study:

• To decouple tooth arrangement prediction tasks and feature modeling.
• To improve the accuracy and efficiency of digital tooth arrangement.
• To introduce a differentiable collision-supervised network for precise 3D tooth positioning.

Main Methods:

• Proposed DTAN (Differentiable Tooth Arrangement Network) to decouple prediction tasks.
• Decoupled hidden features into geometric and positional components with feature consistency constraints.
• Introduced a novel differentiable collision loss function for 3D point cloud data.
• Developed C-DTAN (Controllable DTAN) for arch-width guided tooth arrangement.

Main Results:

• DTAN successfully decouples prediction tasks, enhancing 3D transformation perception.
• The differentiable collision loss effectively constrains inter-tooth relationships, preventing overlaps and gaps.
• C-DTAN provides controllable tooth arrangement based on arch width.
• Achieved drastically improved performance in accuracy and speed compared to existing methods on three constructed datasets.

Conclusions:

• The proposed DTAN and C-DTAN offer a significant advancement in digital orthodontic tooth arrangement.
• Decoupling prediction tasks and incorporating differentiable collision supervision enhance accuracy and prevent undesirable teeth arrangements.
• The methods are efficient, accurate, and controllable, with potential applications in other 3D point cloud tasks.