Evaluation of different partial volume correction methods for longitudinal tau PET imaging: Neuroimaging / Optimal neuroimaging measures for tracking disease progression

Background Previous studies have evaluated whole‐brain and region‐specific accumulation rates of tau using longitudinal 18 F‐Flortaucipir ( 18 F‐FTP) PET. 18 F‐FTP is well known for presenting significant off‐target binding in regions such as the striatum, the white matter and the choroid plexus. However, the use of partial volume correction methods that do not model this off‐target binding is still common in the community, which might hamper optimal longitudinal 18 F‐FTP quantification. The aim of the present study is to study the performance of different PVC methods for longitudinal 18 F‐FTP PET studies. Method Serial 18 F‐FTP PET and T1 images from 101 cognitively unimpaired (CU) and 77 cognitively impaired (CI) subjects were obtained from the Alzheimer’s Disease Neuroimaging Initiative (ADNI). T1 images were segmented using the tools provided by the Computational Anatomical Toolbox (CAT12). Tissue probability maps were binarized to generate GM, WM and CSF masks. The GM was multiplied by the Hammersmith atlas to generate a GM patient‐specific atlas. The CSF was segmented in three different levels based on the maximum uptake in the PET image, so that the choroid plexus was segmented in high/low uptake regions. A traditional region deconvolution PVC, the reblurred Van‐Cittert (RVC) and 2 voxel‐based PVCs, the Region‐based voxel‐wise correction (RBV) and the iterative Yang (iY) were studied (Figure 1). Region‐specific tau accumulation rates were calculated for a medial temporal lobe region of interest, as well as for Braak areas generated by region aggregation. Longitudinal rates of change were estimated for amyloid‐β‐positive CU and CI subjects using linear mixed models and compared with a reference control group of cognitively normal, amyloid‐β‐negative older adults under age 70. Result Voxel‐based PVCs systematically improved the rate of change across all the studied regions in Aβ‐positive CU and Aβ‐positive CI, compared to deconvolution PVC and no PVC. Among the tested voxel‐based PVCs, the RBV was slightly superior to the IY method. Deconvolution‐based PVC provided similar change estimates as not applying PVC (Figure 2) Conclusion Voxel‐based PVCs that model off‐target binding maximize the annualized rate of change of 18 F‐FTP uptake. Our findings suggest that simple deconvolution techniques should be substituted by Voxel‐based PVCs for longitudinal 18 F‐FTP imaging.