DM-CFO: A Diffusion Model for Compositional 3D Tooth Generation With Collision-Free Optimization

Area of Science:

• * Dental Digitalization
• * Computer Graphics
• * Artificial Intelligence

Background:

• * Current 3D tooth generation methods struggle with optimizing the layout and shape of missing teeth.
• * 3D Gaussian-based generation often overlooks collision detection, leading to intersecting objects.
• * Explicit geometric information is frequently absent in existing 3D generation techniques.

Purpose of the Study:

• * To develop an advanced approach for compositional 3D tooth generation that optimizes both layout and shape.
• * To introduce a method that effectively prevents collisions between generated 3D tooth models.
• * To enhance the realism and multiview consistency of automatically designed 3D dental models.

Main Methods:

• * Proposed DM-CFO approach combining graph generation via diffusion models and 3D Gaussian-based collision detection.
• * Progressive restoration of missing tooth layouts during denoising, guided by text and graph constraints.
• * Alternating updates of Gaussian parameters and jaw structure using score distillation sampling (SDS).
• * Introduced a regularization term to penalize intersections based on 3D Gaussian distances.

Main Results:

• * DM-CFO demonstrated significant improvements in multiview consistency and realism on three tooth-design datasets.
• * The method successfully addressed challenges in compositional 3D tooth generation.
• * Collision conflicts were effectively penalized and reduced in the generated 3D models.

Conclusions:

• * DM-CFO offers a robust solution for compositional 3D tooth generation in dental digitization.
• * The approach enhances the quality and accuracy of automatically designed 3D dental models.
• * This method represents a significant advancement in creating realistic and collision-free 3D tooth models.