A Monte Carlo study of deconvolution algorithms for partial volume correction in quantitative PET

In this study, we evaluated several deconvolution methods for partial volume (PV) correction within dynamic positron emission tomography (PET) brain imaging and compared their performance with a PV correction method based on structural imaging. The motivation for this study stemmed from the errors in structural imaging based PV correction that are caused by magnetic resonance (MR) image segmentation and MR-PET registration inaccuracies. The studied deconvolution methods included variants of the iterative Richardson-Lucy deconvolution, variants of the reblurred Van Cittert deconvolution and the linear Wiener deconvolution. Our material consisted of a database of 16 Monte Carlo simulated dynamic ¹¹C-Raclopride images with the same underlying physiology but differing underlying anatomy. We compared the binding potential (BP) values in putamen and caudate resulting from differing PV correction methods to the values computed based on the ground truth time activity curves (TACs). In addition, root mean square errors between TACs extracted from deconvolved images and the ground truth TACs were computed. The iterative deconvolution approaches featured better performance than the linear one. As expected, MR based PV correction under ideal conditions (perfect MR-PET registration and MR image segmentation) yielded more accurate quantification than the deconvolution based methods. However, the iterative deconvolution methods clearly improved the quantitative accuracy of computed physiological parameters (BP) as compared to the case of no PV correction. As variants of the reblurred Van Cittert deconvolution resulted in a lower anatomy-induced variance to the BP values, we consider them to be more interesting than Richardson-Lucy type deconvolution methods.