Fully Convolutional Mandible Segmentation on a valid Ground- Truth Dataset.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Anatomy

Background:

• Computed tomography (CT) imaging of the mandible often contains artifacts (e.g., dental restorations, trauma) complicating manual segmentation.
• Manual segmentation for ground truth generation leads to significant inter-physician variability and uncertainty, impacting algorithm evaluation.
• Excluding challenging slices is not ideal for treatment outcomes, necessitating robust segmentation methods.

Purpose of the Study:

• To develop an automatic segmentation method for the human mandible using deep learning on CT images.
• To establish a reliable and objective ground truth dataset for evaluating mandible segmentation algorithms.
• To address the limitations of manual segmentation in the presence of artifacts and anatomical variations.

Main Methods:

• Trained deep learning networks for automatic segmentation of the mandible in CT images.
• Defined strict inclusion/exclusion criteria for dataset creation, focusing on complete physiological mandibles without teeth.
• Clinical experts manually segmented all cases twice to ensure ground truth reliability and inter-observer agreement.

Main Results:

• The study presents a novel dataset and ground truth for mandible segmentation, addressing challenges of artifacts and bone discontinuities.
• Manual segmentation by experts showed high qualitative and quantitative agreement, validating the ground truth.
• The developed deep learning approach facilitates accurate automatic segmentation of the mandible.

Conclusions:

• This work provides a novel, rigorously validated dataset and ground truth for mandible segmentation.
• The findings support the use of deep learning for reliable automatic mandible segmentation in clinical CT images.
• This approach offers a solution to the uncertainties associated with manual segmentation in complex cases.