MRI based attenuation correction for PET/MRI via MRF segmentation and sparse regression estimated CT

Abstract

MR-based attenuation correction (AC) is a prerequisite to fully harnessing the power of the recently introduced hybrid PET/MRI scanner. Assigning attenuation coefficients based upon MR anatomical images alone remains challenging. In this study, we sought to develop a novel approach based upon hidden Markov random field segmentation (hMRFS) and sparse regression (SR) to estimate CT from T1w images for AC in PET reconstruction in the head. The performance of the proposed method was evaluated using patient-specific PET simulation. We compared the mean absolute (MARE) and full width tenth maximum (FWTM) of relative errors of the reconstructed PET images using attenuation maps from the proposed (μprop), averaged atlas (μatlas) and CT segmentation methods (a.k.a. silver standard) and found that our proposed approach produced significantly lower MARE and FWTM in the errors of the reconstructed PET images. Thus, even with T1w contrast alone, we are able to achieve the accuracy on a par with the previous reports using multispectral MRI data.