Related Experiment Video
Updated: Aug 20, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.9K
Improving performance of deep learning models using 3.5D U-Net via majority voting for tooth segmentation on cone
Kang Hsu1,2, Da-Yo Yuh1, Shao-Chieh Lin3,4
1Department of Periodontology, School of Dentistry, Tri-Service General Hospital, National Defense Medical Center, Taipei, Taiwan, ROC.
Scientific Reports
|November 17, 2022
Summary
Deep learning models for segmenting teeth in cone beam computed tomography (CBCT) show varied performance. A novel 3.5D U-Net, enhanced with majority voting and morphological operations, significantly improved segmentation accuracy and reliability.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Dental Radiology
Background:
- Deep learning enables automatic teeth segmentation in cone beam computed tomography (CBCT).
- Segmentation performance varies significantly based on different deep learning training strategies.
- Accurate teeth segmentation is crucial for various dental diagnostic and treatment planning procedures.
Purpose of the Study:
- To propose and evaluate a novel 3.5D U-Net architecture for improved teeth segmentation on CBCT images.
- To compare the performance of the proposed 3.5D U-Net against various other U-Net configurations and training strategies.
- To assess the impact of post-processing techniques like erosion and dilation (E&D) on segmentation accuracy.
Main Methods:
- Retrospective enrollment of 24 patients who underwent CBCT scans.
- Training and evaluation of multiple U-Net models (2D, 2.5D, 3D, and 3.5D variants) for teeth segmentation.
- Application of majority voting and mathematical morphology operations (erosion and dilation) for performance enhancement.
- Performance evaluation using Dice Similarity Coefficient (DSC), accuracy, sensitivity, specificity, PPV, and NPV with fourfold cross-validation.
Main Results:
- Significant variations in segmentation performance were observed across different U-Net training strategies (P < 0.05).
- The 3.5D U-Net variants, particularly 3.5Dv5 U-Net, demonstrated significantly higher DSC and PPV compared to initial U-Nets (P < 0.05).
- Erosion and dilation (E&D) significantly improved DSC, accuracy, specificity, and PPV (P < 0.005).
- The 3.5Dv5 U-Net achieved the highest DSC and accuracy among all tested models.
Conclusions:
- The segmentation performance of U-Net models for teeth on CBCT can be substantially improved through majority voting and E&D post-processing.
- The proposed 3.5Dv5 U-Net architecture offers superior performance for automatic teeth segmentation in CBCT imaging.
- This enhanced segmentation accuracy has significant implications for improving dental diagnostics and treatment planning.
