A latent variable deep generative model for 3D anterior tooth shape.

Area of Science:

• Computer-aided design and manufacturing
• Biomedical engineering
• 3D modeling and reconstruction

Background:

• Digital dentistry relies on accurate 3D tooth shape representation.
• Generating realistic and diverse 3D tooth shapes is crucial for restorative and cosmetic procedures.
• Existing methods may lack integration with current digital design workflows.

Purpose of the Study:

• Introduce PointFlow, a 3D generative technology for creating 3D tooth shapes.
• Integrate PointFlow with conventional digital design workflows.
• Evaluate the clinical applicability of PointFlow for tooth reconstruction.

Main Methods:

• Trained a deep generative model (PointFlow) on 1337 3D scans of natural anterior teeth.
• Encoded 3D tooth geometries into compact latent codes representing morphological features.
• Generated new tooth shapes by sampling from the latent space and evaluated using 3D shape metrics.
• Assessed clinical applicability through reconstruction of 60 artificially damaged teeth.

Main Results:

• PointFlow effectively captured the diversity of anterior tooth shapes.
• Generated tooth shapes demonstrated superior performance on generative metrics compared to the reference dataset.
• The model successfully reconstructed missing regions in damaged teeth with an average Chamfer Distance of 0.2738 ± 0.095 mm.

Conclusions:

• Deep generative models can accurately learn tooth characteristics.
• PointFlow shows potential for generating high-quality 3D tooth shapes.
• The technology is applicable for further clinical use in dentistry.