Comparison of atlas-based bone segmentation methods in whole-body PET/MRI

The aim of this work is to compare bone extraction accuracy of commonly used atlas-based segmentation algorithms for generation of pseudo-CT images in whole-body PET/MRI. Eight different bone segmentation methods from whole-body MR images were implemented. The study comprised 23 patients who underwent sequential PET/CT and PET/MRI scans, thus enabling to produce pairs of aligned 3D Dixon MR and CT images. The voxel-wise weighting method, which employs local normalized cross-correlation similarity measure to give appropriate weights locally, outperformed other pseudo-CT generation approaches, yielding a Dice similarity index of 0.77±0.04 (P<;0.05) compared to 0.60±0.02 (P<;0.05) achieved by bone segmentation using simple averaging of all CT atlases. This approach is promising for MRI-guided attenuation correction in whole-body PET/MRI.