Single Bone Modeler: deep learning bone segmentation for cone-beam CT

Area of Science:

• Orthopedic imaging and computational anatomy.
• Medical image analysis and deep learning applications.
• 3D modeling for surgical planning.

Background:

• Accurate bone segmentation is vital for orthopedic diagnosis and surgical planning.
• Traditional methods struggle with complex structures in cone-beam computed tomography (CBCT) scans.
• Existing techniques lack precision for detailed extremity bone analysis.

Purpose of the Study:

• To introduce a novel deep learning workflow, the Single Bone Modeler (SBM), for precise bone segmentation in CBCT scans.
• To develop and evaluate a U-Net architecture for enhanced bone segmentation accuracy.
• To compare different training strategies (axial vs. multi-planar) for CBCT data.

Main Methods:

• A U-Net architecture was developed for bone segmentation, compared against SegNet.
• Axial and multi-planar training strategies were evaluated for CBCT data.
• Bone separation utilized a watershed algorithm, followed by 3D modeling.

Main Results:

• The U-Net model trained with multi-planar strategy achieved superior segmentation performance.
• Achieved Jaccard Index (JI) of 0.941 ± 0.031 and Dice Coefficient (DC) of 0.970 ± 0.015.
• The SBM workflow demonstrated significant improvement over benchmark methods in isolating specific bones.

Conclusions:

• The Single Bone Modeler (SBM) workflow enhances precision in segmenting bones from high-resolution CBCT scans.
• The proposed deep learning approach offers reliable and efficient extremity bone segmentation.
• Results indicate potential for improved orthopedic applications through enhanced imaging analysis.