Layered deep learning for automatic mandibular segmentation in cone-beam computed tomography.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Anatomy

Background:

• Semi-automatic segmentation of the mandible from cone-beam computed tomography (CBCT) is crucial for surgical planning but is time-consuming and prone to user error.
• Existing methods require significant manual input, impacting workflow efficiency and consistency in creating 3D surface models.

Purpose of the Study:

• To develop and validate a layered deep learning algorithm for automated 3D mandibular surface model generation from CBCT data.
• To assess the time-efficiency, consistency, and accuracy of the AI model compared to semi-automatic segmentation (SA) and user-refined AI segmentations (RAI).

Main Methods:

• A cloud-based AI model utilizing a 3D U-Net architecture was trained on 160 anonymized CBCT scans.
• The AI model's performance was evaluated on timing, intra- and inter-operator consistency (IoU, DSC, HD, etc.), and accuracy against SA as the ground truth.
• Comparisons were made between the AI model, RAI, and SA methods.

Main Results:

• The AI model achieved segmentation in 17 seconds, a 71.3-fold decrease compared to SA (1218.4s).
• RAI and the AI model demonstrated superior intra- and inter-operator consistency over SA, with AI being consistently accurate by default.
• Accuracy metrics (IoU, DSC, HD) for AI and RAI were comparable to SA, indicating high fidelity in 3D model generation.

Conclusions:

• The layered 3D U-Net deep learning algorithm significantly improves time-efficiency and reduces operator error in creating 3D mandibular models from CBCT.
• The AI approach offers excellent accuracy and consistency, outperforming current semi-automatic segmentation methods.
• This AI-driven method represents a more efficient and reliable tool for clinical applications in orthognathic surgery and other dental fields.