Comparative Analysis of Deep Learning Architectures for Automatic Tooth Segmentation in Panoramic Dental Radiographs:...

Area of Science:

• Artificial Intelligence in Medical Imaging
• Deep Learning for Medical Image Analysis
• Dental Radiography and Diagnostics

Background:

• Automatic tooth segmentation in panoramic dental radiographs is crucial for diagnostic accuracy.
• Evaluating U-Net based deep learning models with various encoder backbones is essential for optimizing performance and efficiency.

Purpose of the Study:

• To systematically benchmark U-Net deep learning models for automatic tooth segmentation in panoramic dental radiographs.
• To analyze the trade-off between segmentation accuracy and computational cost across different encoder backbones (ResNet, EfficientNet, DenseNet, MobileNetV3-Small).

Main Methods:

• U-Net models with pre-trained ImageNet encoder families were evaluated on the Tufts Dental Database (1000 images).
• A five-fold cross-validation strategy was employed.
• Segmentation performance was measured using Dice coefficient and Intersection over Union (IoU); computational efficiency was assessed by parameter count and GFLOPs.

Main Results:

• Overall segmentation quality was high (Dice: 0.9168-0.9259), with diminishing returns for increased backbone complexity.
• EfficientNet-B0 achieved a near-optimal balance of high accuracy (Dice: 0.9244 ± 0.0011) and low computational cost (5.98 GFLOPs).
• While EfficientNet-B7 had the highest nominal accuracy, differences were not statistically significant compared to EfficientNet-B0 and B4.

Conclusions:

• Larger deep learning models do not consistently yield superior performance for tooth segmentation.
• EfficientNet-B0 is identified as the most practical model, offering near-saturated accuracy with significantly reduced model size and computational demands.
• The findings highlight the importance of selecting computationally efficient models without compromising diagnostic accuracy in dental imaging applications.