Clinically oriented automatic three-dimensional enamel segmentation via deep learning.

Area of Science:

• Biomedical Engineering
• Dental Imaging
• Artificial Intelligence

Background:

• Accurate enamel segmentation is vital for dental treatments and evolutionary studies.
• Current methods lack non-destructive, chair-side assessment capabilities for enamel integrity and thickness.
• Automated, efficient, and accurate enamel analysis is needed in clinical dentistry.

Purpose of the Study:

• To develop a deep learning model, 2.5D Attention U-Net, for automated enamel segmentation.
• To train the model on small datasets for efficient and accurate enamel segmentation across all teeth.
• To enable precise, non-destructive, chair-side assessment of enamel thickness and integrity.

Main Methods:

• A fully automated computer-aided enamel segmentation model using 2.5D Attention U-Net was developed.
• The model was trained on manually annotated enamel data after annotation and augmentation.
• Performance was evaluated using Dice similarity coefficient, generating 3D enamel models and calculating thickness via ray-tracing.

Main Results:

• The 2.5D Attention U-Net model achieved a 96.6% Dice score for enamel segmentation.
• Quantitative analysis revealed enamel thickness variations, being thickest at incisal edges/cusps and thinning towards roots.
• Specific thickness patterns were identified across different tooth types and jaws, with gradual increases from incisors to molars.

Conclusions:

• The 2.5D Attention U-Net system provides automatic, efficient, and accurate enamel segmentation and analysis.
• This advancement enhances precise chair-side diagnosis and treatment planning for enamel-related conditions.
• The system represents a significant step forward in automated diagnostics for dental enamel.