Automatic multi-anatomical skull structure segmentation of cone-beam computed tomography scans using 3D UNETR

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Anatomy

Background:

• Accurate medical and dental image segmentation is crucial for clinical decision support systems.
• Current manual segmentation is time-consuming, averaging 7 hours per case.
• Automated segmentation streamlines diagnosis, therapy planning, and follow-up.

Purpose of the Study:

• To develop a novel, automated tool for precise full-face segmentation.
• To significantly reduce the time required for medical image segmentation.
• To integrate state-of-the-art UNETR models within the MONAI framework.

Main Methods:

• Utilized the UNETR architecture from the MONAI framework.
• Trained and validated models on 618 de-identified Cone-Beam Computed Tomography (CBCT) head scans.
• Employed a 5-fold cross-validation strategy for robust evaluation.

Main Results:

• Achieved highly accurate and robust full-face segmentation.
• Demonstrated a significant time reduction to approximately 5 minutes per case.
• Obtained a high Dice score of 0.962±0.02, indicating excellent performance.

Conclusions:

• The proposed tool offers a substantial improvement in efficiency and accuracy for medical image segmentation.
• The UNETR-based approach demonstrates generalizability across diverse CBCT data.
• The open-source availability of the code facilitates clinical adoption and further research.