Improved distinct bone segmentation in upper-body CT through multi-resolution networks.

Area of Science:

• Medical Imaging
• Computer Vision
• Radiology

Background:

• Automated bone segmentation from CT scans is crucial for surgical planning and navigation.
• U-Net variants excel at semantic segmentation but face challenges with high-resolution, large-field-of-view CT data.
• Existing methods struggle with detail loss or localization errors due to limited spatial context in 3D U-Net architectures.

Purpose of the Study:

• To develop an improved method for distinct bone segmentation from upper-body CT scans.
• To address limitations of single-resolution 3D U-Nets in handling large fields of view and high-resolution inputs.
• To enhance the accuracy and efficiency of automated bone segmentation for clinical applications.

Main Methods:

• Proposed an end-to-end trainable segmentation network combining multiple 3D U-Nets at different resolutions.
• Implemented a multi-resolution architecture that captures low-resolution spatial information and transfers it to high-resolution networks.
• Evaluated the architecture against single-resolution networks and conducted ablation studies on feature concatenation and context network numbers.

Main Results:

• The best-performing network achieved a median Dice Similarity Coefficient (DSC) of 0.86 across 125 bone classes.
• Significantly reduced confusion between similar-looking bones in different anatomical locations.
• Outperformed previous 3D U-Net baseline results and other published distinct bone segmentation methods.

Conclusions:

• Multi-resolution 3D U-Nets effectively address shortcomings in bone segmentation from upper-body CT scans.
• The approach captures a larger field of view while mitigating computational complexity associated with high-resolution 3D data.
• Improved accuracy and efficiency in distinct bone segmentation from CT scans are achieved.