A Gaussian mixture model for combining single threshold and adaptive threshold segmentation of bone microstructure.

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Quantitative Bone Analysis

Background:

• High-resolution CT (HR-pQCT, PCCT) segmentation is crucial for assessing bone microstructure, quality, and fracture risk.
• Traditional global thresholding methods struggle with image noise and inhomogeneities, leading to poor microstructure representation.
• Adaptive thresholding preserves fine features but lacks quantitative bone mineral density (BMD) information.

Purpose of the Study:

• To introduce a novel Gaussian mixture model (GMM) based method for bone microstructure segmentation.
• To compare the accuracy of the GMM method against global and adaptive thresholding techniques.
• To evaluate the GMM method's ability to retain quantitative information while improving segmentation accuracy.

Main Methods:

• Developed a GMM technique to model the normalized image histogram as a bi-modal Gaussian distribution.
• Defined an analytical threshold and an intensity range for uncertain tissue classes for adaptive segmentation.
• Validated the GMM method using simulated images and applied it to HR-pQCT and PCCT wrist scans, comparing against micro-CT reference segmentations.

Main Results:

• The GMM method accurately determined optimal thresholds in simulated images when components were distinct.
• For HR-pQCT, GMM segmentation accuracy (83.0±2.3%) was comparable to full adaptive segmentation (86.2±3.7%).
• For PCCT, GMM achieved higher accuracy (77.9±2.3%) than adaptive segmentation (76.4±2.4%).

Conclusions:

• The proposed GMM method offers accurate bone microstructure segmentation comparable to conventional techniques.
• GMM provides a tunable approach, allowing adjustments for higher sensitivity or specificity in segmentation.
• This method enhances the quantitative assessment of bone microstructure from CT imaging data.