Deep learning for automatic mandible segmentation on dental panoramic x-ray images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oral and Maxillofacial Radiology

Background:

• Mandible bone structure analysis is crucial for understanding systemic diseases like osteoporosis.
• Manual mandible segmentation in dental panoramic X-rays (PAN) is laborious, prone to variability, and challenging due to image quality issues.
• Accurate automatic mandible segmentation (AMS) is essential for large-scale oral health research.

Purpose of the Study:

• To develop precise and robust deep learning algorithms for automatic mandible segmentation (AMS) on PAN images.
• To evaluate the performance of U-Net and HRNet architectures, with and without data augmentation, for AMS.
• To enhance segmentation accuracy through morphological refinement and an ensemble approach.

Main Methods:

• Training four deep learning models (U-Net, HRNet) on two datasets (in-house: 393 pairs, third-party: 116 pairs).
• Implementing data augmentation and a morphological refinement routine to improve segmentation.
• Developing an ensemble model combining the four best-performing segmentation models.

Main Results:

• The ensemble model with morphological refinement achieved superior performance, reaching 98.27% accuracy, 97.60% DICE, and 97.18% IoU on the test set.
• All trained models demonstrated high performance, exceeding 95% across all metrics.
• The study achieved the highest performance compared to previous AMS research on PAN images.

Conclusions:

• Deep learning-based automatic mandible segmentation (AMS) provides a precise, robust, and efficient alternative to manual segmentation.
• The proposed ensemble model significantly advances AMS for PAN images, supporting research on systemic diseases.
• The robust results, validated on a diverse dataset, confirm the model's generalizability across various patient demographics and oral conditions.