Comparison of state-of-the-art atlas-based bone segmentation approaches from brain MR images for MR-only radiation planning and PET/MR attenuation correction

Magnetic Resonance (MR) imaging has emerged as a valuable tool in radiation treatment (RT) planning as well as Positron Emission Tomography (PET) imaging owing to its superior soft-tissue contrast. Due to the fact that there is no direct transformation from voxel intensity in MR images into electron density, itchr('39')s crucial to generate a pseudo-CT (Computed Tomography) image from MRI for the task of MR-guided attenuation correction and RT planning. This study set out to investigate the performance of two state-of-the-art atlas-based pseudo-CT generation approaches from MR brain images.

Bone segmentation was performed on 43 brain CT and MR pairs using atlas-based local weighting (Atlas-LW) and atlas registration combined with pattern recognition (AT-PR) techniques. The accuracy of bone extraction performed by these two approaches was investigated for the entire bony structures as well as cortical bone using standard segmentation metrics such as Dice similarity (DSC). Moreover, the accuracy of the CT value estimated from MR images was evaluated using mean absolute error (MAE), and root mean square error (RMSE) metrics.

Overall, Atlas-LW technique exhibited higher segmentation accuracy resulting in DSC=0.79±0.61 and 0.84±0.03, while AT-PR method led to DSC=0.72±0.8 and 0.77±0.05 for cortical and total bone, respectively. Moreover, Atlas-LW approach estimated CT values for the total bone tissue with MAE=17.6±138.0 HU and RMSE=466.2±75.0 HU compared to MAE=-10.9±147.0 HU and RMSE=522.5±89.7 HU obtained from AT-PR approach.

The Atlas-LW exhibited superior the performance in this study which demonstrates its potential to be employed in PET/MR attenuation correction (AC) and MR-only RT planning