Benchmarking Deep Learning Models for Tooth Structure Segmentation.

Area of Science:

• Artificial Intelligence in Dentistry
• Deep Learning for Medical Imaging
• Radiographic Analysis

Background:

• Deep learning (DL) model selection for dental applications is often unsystematic.
• Comprehensive benchmarking of DL architectures in dentistry is lacking.
• Tooth structure segmentation on dental radiographs is a critical task.

Purpose of the Study:

• To systematically benchmark various deep learning architectures for tooth structure segmentation on dental bitewing radiographs.
• To evaluate the impact of different network architectures, encoders, and initialization strategies on segmentation performance.

Main Methods:

• Developed 72 distinct DL models by combining 6 network architectures with 12 encoders (ResNet, VGG, DenseNet families).
• Applied 3 initialization strategies (ImageNet, CheXpert, random) to 216 trained models.
• Utilized a dataset of 1,625 annotated radiographs, 5-fold cross-validation, and F1-score for performance quantification.

Main Results:

• Initialization with ImageNet or CheXpert weights significantly outperformed random initialization (P < 0.05).
• Deeper, more complex models did not consistently outperform simpler alternatives.
• VGG-based models demonstrated robustness, while ResNet-based models achieved peak performance.

Conclusions:

• Pre-trained weights are recommended for training DL models in dental radiographic analysis.
• Less complex architectures can be viable alternatives when computational resources are limited.
• Models optimized for non-dental tasks may not generalize effectively to dental-specific challenges.