TY - JOUR T1 - Objective evaluation of voxelwise partial volume correction methods in brain PET imaging JF - Journal of Nuclear Medicine JO - J Nucl Med SP - 72 LP - 72 VL - 61 IS - supplement 1 AU - Murat Bilgel AU - Yansong Zhu AU - Olivier Rousset AU - Dean Wong AU - Susan Resnick AU - Arman Rahmim Y1 - 2020/05/01 UR - http://jnm.snmjournals.org/content/61/supplement_1/72.abstract N2 - 72Introduction: Partial volume effects impair the accurate quantification of PET scans, and are important consideration in voxelwise statistical analyses. A number of methods have been proposed for voxelwise partial volume correction (PVC); however, comparison of PVC methods in real data sets is difficult due to the lack of objective criteria. Objectives: develop objective criteria to assess the performance of voxelwise PVC methods and apply these criteria to compare three PVC methods using 15O-H2O brain PET scans. Methods: We used cross-sectional 15O-H2O brain PET data acquired on a GE Advance scanner for 175 participants from the Baltimore Longitudinal Study of Aging to assess the performance of the following voxelwise PVC methods: reblurred Van Cittert (VC) [1], Region-Based Voxelwise (RBV) [2], and Parallel Level Set (PLS) with a non-smooth optimization technique (split Bregman) [3]. Images were reconstructed as a single time frame using filtered backprojection, yielding ~7.5 mm FWHM resolution at the center of the field of view [4]. For each participant, we coregistered their inhomogeneity-corrected and skull-stripped structural MRI with their PET. Anatomical regions were defined on the MRI using a multi-atlas labeling approach [5] and mapped onto the PET using the coregistration result. We used the cerebellar gray matter as the reference region to compute standardized uptake value ratios (SUVR). To aid in the evaluation of the performance of voxelwise PVC methods, we calculated the error in SUVR estimates. Assuming that voxelwise intensities are independent, and that within a given region, they are identically distributed as a normal with mean μ and variance σ2 , a first-order Taylor series expansion of the definition of SUVR yields Var(SUVR)≍(μT/μR)2[σT2/(n μT2) + σR2 / (m μR2)], where n and m are the number of voxels within the target (T) and reference (R) regions, respectively. A lower variance is desirable, since it indicates a smaller error in the SUVR estimate. To assess the performance of voxelwise PVC methods, we used the following criteria: the lowest uncertainty in regional SUVR estimates, given by Var(SUVR), and the greatest negative age association at baseline (after adjusting for sex), given by the number of voxels with statistically significant differences within the cortical gray matter. Results: PLS yielded the lowest Var(SUVR), resulting in a lower uncertainty in regional SUVR compared to not performing PVC, while not affecting the overall SUVR levels (Figure 1). In the voxelwise analysis, PLS yielded the greatest number of statistically significant voxels where older individuals had lower SUVR (Figure 2). Conclusion: PLS can improve statistical analysis by reducing the uncertainty around regional SUVR estimates, and may enable the discovery of effects that would not be considered statistically significant without PVC or other PVC methods. Acknowledgments: This research was supported in part by the Intramural Research Program of the National Institute on Aging, National Institutes of Health, and R21 Grant No. AG056142. Figure 1. Swarm and box plots of (A) the expected value of SUVR and (B) square root of the SUVR variance for the right posterior cingulate gyrus. Results were similar in other cortical gray matter regions. None=No PVC, VC=reblurred Van Cittert, RBV=Region-Based Voxelwise, PLS=Parallel Level Set. Figure 2. Normalized histograms of two-sided t-values for the age term in the linear regression model with voxelwise SUVR as outcome (adjusted for sex). Each row corresponds to a different PVC method. Red vertical dashed lines indicate the t-values that correspond to a multiple comparison-corrected (via permutation tests) p<0.05 threshold. Number of statistically significant cortical gray matter voxels with a negative and positive association are indicated to the left and right of the dashed lines, respectively. Sagittal slices illustrate the extent of statistically significant clusters on the left medial cortical surface. ER -