Deconvolution-based partial volume correction in Raclopride-PET and Monte Carlo comparison to MR-based method.

In this work, we evaluated three iterative deconvolution algorithms and compared their performance to partial volume (PV) correction based on structural imaging in brain positron emission tomography (PET) using a database of Monte Carlo-simulated images. We limited our interest to quantitative radioligand PET imaging, particularly to (11)C-Raclopride and striatal imaging. The studied deconvolution methods included Richardson-Lucy, reblurred Van Cittert, and reblurred Van Cittert with the total variation regularization. We studied the bias and variance of the regional estimates of binding potential (BP) values and the accuracy of regional TACs as a function of the applied image processing. The resolution/noise tradeoff in parametric BP images was addressed as well. The regional BP values and TACs obtained by deconvolution were almost as accurate than those by structural imaging-based PV correction (GTM method) when the ideal volumes of interests (VOIs) were used to extract TACs from the images. For deconvolution methods, the ideal VOIs were slightly eroded from the exact anatomical VOI to limit the bias due to tissue fraction effect which is not corrected for by deconvolution-based methods. For the GTM method, the ideal VOIs were the exact anatomical VOIs. The BP values and TACs by deconvolution were less affected by segmentation and registration errors than those with the GTM-based PV correction. The BP estimates and TACs with deconvolution-based PV correction were more accurate than BPs and TACs derived without PV correction. The parametric images obtained by the deconvolution-based PV correction showed considerably improved resolution with only slightly increased noise level compared to the case with no PV correction. The reblurred Van Cittert method was the best of the studied deconvolution methods. We conclude that the deconvolution is an interesting alternative to structural imaging-based PV correction as it leads to quantification results of similar accuracy, and it is less prone to registration and segmentation errors than structural imaging-based PV correction. Moreover, PV-corrected parametric images can be readily computed based on deconvolved dynamic images.