Subtle MR Guidance for Partial Volume Correction of PET Images: A Comparison of Techniques

Quantitative accuracy of PET is influenced by the partial volume effect. Different MR-based partial volume correction (PVC) methods have been proposed to tackle this issue. Early methods like geometric-transfer matrix (GTM) require segmentation of MR images into different regions and also make the strong assumption that activity distribution in each region is constant. Though they have shown good performance with perfect MR guidance, imperfect MR information may degrade their performance. Our hypothesis is methods that make subtle usage of MRI images can prove more advantageous and depict greater feasibility and strength for translation to routine usage. In this work, we compared different MR-guided post-reconstruction PVC methods with perfect and imperfect MR guidance due to errors in registration and segmentation. We considered PVC methods that perform subtle MR guidance, including our two recently proposed methods: anatomical-guided non-local mean (NLMA) and split Bregman based parallel level set (SBPLS), which solves PLS with non-smooth optimization technique and shows advantages compared to ordinary smooth optimization framework, as well as symmetric (sBowsher) and asymmetric Bowsher (aBowsher), and PLS solved with smooth optimization technique. Region-based voxel-wise (RBV) method, a GTM-based method, was also taken for comparison. Reblurred Van Citter (VC) deconvolution method was chosen as a reference method that does not use MR information at all. These methods were evaluated with realistic simulation experiments based on the BrainWeb phantom. PET-MR registration mismatches were additionally modeled. Imperfect segmentation was obtained by segmenting MR images with SPM12. Bias and coefficient of variability (COV) were computed for quantitative comparison. Results indicated PVC methods with subtle use of MR guidance to show stronger resilience to imperfect MR information compared to methods with stronger assumptions invoked in MR-guided PVC.