Evaluation of partial volume correction and analysis of longitudinal [18F]GTP1 tau PET imaging in Alzheimer's disease using linear mixed-effects models

Abstract

Purpose: We evaluated the impact of partial volume correction (PVC) methods on the quantification of longitudinal [¹⁸F]GTP1 tau positron-emission tomography (PET) in Alzheimer's disease and the suitability of describing the tau pathology burden temporal trajectories using linear mixed-effects models (LMEM).

Methods: We applied van Cittert iterative deconvolution (VC), 2-compartment, and 3-compartment, and the geometric transfer matrix plus region-based voxelwise methods to data acquired in an Alzheimer's disease natural history study over 18 months at a single imaging site. We determined the optimal PVC method by comparing the standardized uptake value ratio change (%ΔSUVR) between diagnostic and tau burden-level groups and the longitudinal repeatability derived from the LMEM. The performance of LMEM analysis for calculating %ΔSUVR was evaluated in a natural history study and in a multisite clinical trial of semorinemab in prodromal to mild Alzheimer's disease by comparing results to traditional per-visit estimates.

Results: The VC, 2-compartment, and 3-compartment PVC methods had similar performance, whereas region-based voxelwise overcorrected regions with a higher tau burden. The lowest within-subject variability and acceptable group separation scores were observed without PVC. The LMEM-derived %ΔSUVR values were similar to the per-visit estimates with lower variability.

Conclusion: The results indicate that the tested PVC methods do not offer a clear advantage or improvement over non-PVC images for the quantification of longitudinal [¹⁸F]GTP1 PET data. LMEM offers a robust framework for the longitudinal tau PET quantification with low longitudinal test-retest variability.

Clinical trial registration: NCT02640092 and NCT03289143.

Keywords: Alzheimer's disease; linear mixed-effects models; longitudinal change; neuroimaging; tau PET.

Figure 1

Surface maps and violin plots of [¹⁸F]GTP1 SUVR displaying the distribution of tau burden by tau level at baseline in the Tauriel study in prodromal and mild AD participants: low tau (TMP SUVR ≤1.30), mid tau (1.30 < TMP SUVR ≤ 1.75) and high tau (TMP SUVR >1.75). The pattern of the uptake from low to high tau burden in the temporal meta-regions of interest is maintained across the cortex. AD, Alzheimer's disease; MT, mesial temporal regions; SUVR, standardized uptake value ratio; TMP, temporal meta-regions of interest.

Figure 2

Relationship between the measure of group separation (cohort × time interaction t-score) and the within-subject variability or residual error. Each point in the plot shows the t-score and residual error in a given region (symbols) for a given PVC method (or non-PVC). Filled symbols indicate significant differences between AD groups (diagnostic groups on top and tau-level groups in the bottom row) and cognitively unimpaired participants (t-scores >1.96, p < 0.05). CU, cognitively unimpaired; MT, mesial temporal regions; mTMP, PVC, partial volume correction; RBV, region-based voxelwise; TMP excluding MT; VC, van Cittert iterative deconvolution; WCG, whole cortical gray.

Figure 3

Average annualized change in SUVR and the corresponding 95% confidence intervals in the (A) NHS participants and (B) the Tauriel study estimated from the LMEM slope analysis and at the follow up visits (NHS: weeks 52 and 78; Tauriel: weeks 49 and 73) relative to baseline. Subjects are grouped by diagnostic cohort or tau level. Average and confidence intervals were estimated applying bootstrap. %ΔSUVR, standardized uptake value ratio change; CU, cognitively unimpaired; LMEM, linear mixed effects models; MT, mesial temporal regions; mTMP, NHS, natural history study; PVC, partial volume correction; RBV, region-based voxelwise; SUVR, standardized uptake value ratio; TMP excluding MT; VC, van Cittert iterative deconvolution; W, week; WCG, whole cortical gray.

Figure 4

Forest plots illustrate the Spearman correlations (±95% CI) between annualized SUVR change and baseline SUVR in the NHS (AD participants only; left) and the Tauriel study (right) across different regions of interest. In the NHS, the two LMEM slope analysis including tau level × time or diagnostic × time interactions are shown. The Tauriel LMEM included the Diagnostic × tau level × time interaction. AD, Alzheimer's disease; LMEM, linear mixed-effects models; MT, mesial temporal regions; mTMP, TMP excluding MT; NHS, natural history study; SUVR, standardized uptake value ratio; W, week; WCG, whole cortical gray.

Figure 5

Heat maps of the annualized %ΔSUVR estimated from the LMEM slope analysis in AD participants in the NHS and Tauriel study. %ΔSUVR, standardized uptake value ratio change; AD, Alzheimer's disease; LMEM, linear mixed-effects models; MT, mesial temporal regions; mTMP, TMP excluding MT; NHS, natural history study; WCG, whole cortical gray.