Related Experiment Video
Updated: Sep 6, 2025
09:10
Digital Hybrid Model Preparation for Virtual Planning of Reconstructive Dentoalveolar Surgical Procedures
Published on: August 5, 2021
1.9K
Segmentation of Dental Restorations on Panoramic Radiographs Using Deep Learning
Csaba Rohrer1, Joachim Krois1,2, Jay Patel3
1Oral Diagnostics, Digital Health and Health Services Research, Charité-Universitätsmedizin Berlin, 10117 Berlin, Germany.
Diagnostics (Basel, Switzerland)
|June 24, 2022
Summary
Training Convolutional Neural Networks (CNNs) on smaller image tiles significantly improved dental restoration segmentation accuracy on panoramic radiographs. This tiling approach enhances model performance and speeds up convergence for better dental image analysis.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Dental Radiology
Background:
- Convolutional Neural Networks (CNNs), specifically U-Net, are standard for medical image segmentation.
- Segmenting dental restorations in panoramic radiographs presents challenges due to variations in shape, size, and type.
- Existing methods may struggle with class imbalance, favoring frequent and large features.
Purpose of the Study:
- To evaluate the effectiveness of training U-Net models on tiled panoramic radiographs for dental restoration segmentation.
- To compare the performance of models trained on full images versus various numbers of image tiles.
- To assess the impact of tiling on model convergence speed and accuracy, particularly for less frequent features like root canal fillings.
Main Methods:
- Pixelwise annotation of 1781 panoramic radiographs for dental restorations (fillings, crowns, root canal fillings) by dental experts.
- Training U-Net models using different numbers of equally spaced rectangular image crops (tiles) from panoramic radiographs.
- Employing five-times-repeated three-fold cross-validation for robust model evaluation.
- Comparing performance metrics (F1-score) between full image training and tiled image training.
Main Results:
- Training with an increased number of tiles demonstrably improved segmentation performance and accelerated model convergence.
- The F1-score increased from 0.7 for the full image to 0.83, 0.92, and 0.95 with 6, 10, and 20 tiles, respectively.
- A significant performance boost (+294%) was observed for segmenting root canal fillings, addressing class imbalance issues.
Conclusions:
- Training Convolutional Neural Networks on tiled panoramic radiographs enhances pixelwise classification performance for dental restorations.
- The tiling strategy effectively overcomes segmentation bias towards more frequent and extended classes, improving overall accuracy.
- This approach accelerates model convergence and improves the segmentation of challenging, less frequent dental features.
