Comparison of atlas-based techniques for whole-body bone segmentation

Area of Science:

• Medical Imaging
• Image Segmentation
• Radiology

Background:

• Accurate bone segmentation from MRI is essential for PET/MRI attenuation correction.
• Existing atlas-based methods vary in effectiveness for whole-body bone extraction.

Purpose of the Study:

• To evaluate and identify the most accurate atlas-based segmentation method for whole-body bone extraction from MRI.
• To inform the development of improved MRI-guided PET attenuation maps.

Main Methods:

• Compared various atlas-based segmentation techniques including intensity averaging, majority voting, global/local weighted atlas fusion, shape-based averaging, and Hofmann's method.
• Utilized normalized mutual information (NMI), normalized cross-correlation (NCC), and mean square distance (MSD) as similarity measures.
• Evaluated segmentation accuracy on 23 whole-body CT/MR image pairs using Dice similarity coefficient (DSC) and relative volume difference (RVD).

Main Results:

• Local weighted atlas fusion using MSD achieved the highest accuracy (DSC = 0.81 ± 0.03).
• Global weighted methods showed moderate improvement (DSC = 0.65 ± 0.05) over non-weighted intensity averaging (DSC = 0.60 ± 0.02).
• Shape-based averaging and Hofmann's methods did not improve accuracy compared to intensity averaging.

Conclusions:

• Voxel-wise local weighted atlas fusion demonstrates superior performance for whole-body bone segmentation from MRI.
• Image denoising, normalization, and parameter optimization are critical for accurate bone extraction.
• This approach holds promise for enhancing MRI-guided PET attenuation map generation.