Longitudinal amyloid and tau PET imaging in Alzheimer's disease: A systematic review of methodologies and factors affecting quantification

Deposition of amyloid and tau pathology can be quantified in vivo using positron emission tomography (PET). Accurate longitudinal measurements of accumulation from these images are critical for characterizing the start and spread of the disease. However, these measurements are challenging; precision and accuracy can be affected substantially by various sources of errors and variability. This review, supported by a systematic search of the literature, summarizes the current design and methodologies of longitudinal PET studies. Intrinsic, biological causes of variability of the Alzheimer's disease (AD) protein load over time are then detailed. Technical factors contributing to longitudinal PET measurement uncertainty are highlighted, followed by suggestions for mitigating these factors, including possible techniques that leverage shared information between serial scans. Controlling for intrinsic variability and reducing measurement uncertainty in longitudinal PET pipelines will provide more accurate and precise markers of disease evolution, improve clinical trial design, and aid therapy response monitoring.


INTRODUCTION
2][3] In this preclinical stage, the primary pathological hallmarks are the accumulation of amyloid beta (Aβ) and the aggregation of neurofibrillary tau tangles, both of which can be imaged in vivo using positron emission tomography (PET).
Although cerebrospinal fluid (CSF), and, more recently, plasma assays, can provide global estimates of amyloid and tau loads, molecular imaging with PET provides spatial information about the location of those protein deposits within the brain and how they change over time. 4,5inical interpretation of PET images based on visual reads are simple to implement and effective in identifying marked elevations in uptake, but a further opportunity of PET comes from quantification of the ADprotein loads and their variation over time, particularly in the earliest phases of accumulation that are challenging for visual reads. 6,7e deposition of Aβ and tau pathologies are believed to ultimately result in neuronal death, which then leads to cognitive and functional decline.Whereas the ordering of these pathologies has largely been inferred from cross-sectional studies and can predict disease progression trends at the group level, longitudinal observations can provide more precise estimates based on individual trajectories and help determine causality.In addition, they could help uncover the substantial heterogeneity of the disease that might be overlooked when analyzing group estimates. 8Therefore, longitudinal PET studies are valuable for improving our understanding of the start and subsequent spread of the disease and are crucial in clinical trials to assess therapeutic efficacy.In particular, the change in Aβ or tau over time has been used as secondary or exploratory endpoints in clinical trials including the ones investigating the anti-Aβ monoclonal antibodies aducanumab, 9 donanemab, 10 lecanemab, 11 and gantenerumab. 12In addition, by determining baseline characteristics that identify individuals who are most likely to accumulate amyloid or tau, longitudinal studies can help inform enrichment strategies for trials, where improved statistical power for detecting evidence of disease modification will come from larger reductions with less variability.
To optimize the value of quantitative PET, biological and technical factors that influence the image acquisition, image reconstruction, and data analysis should be accounted for.The lack of homogeneity in processing pipelines and inherent limitations of acquisition parameters introduce uncertainties in measuring Aβ and tau protein abnormalities with sufficient precision. 13This is well illustrated by the fact that the annual accumulation rates of Aβ and tau are on the same order of magnitude as the test-retest variability of their respective tracers.
Key factors that impact longitudinal amyloid-PET measurement uncertainty were first identified and characterized by Schmidt et al. 13 This current work aims to update the framework first proposed there.
Since its publication, the number of research studies involving longitudinal amyloid and tau PET has increased substantially, allowing us to evaluate and update its recommendations.First, a systematic review of the literature to identify current strategies for measuring amyloid and tau PET accumulation over time is presented.Both amyloid and tau PET were included, as they provide complementary information

RESEARCH IN CONTEXT
1. Systematic review: The authors reviewed the literature using traditional sources and meeting abstracts and presentations.The review focuses on dementia-related research studies involving longitudinal positron emission tomography (PET), and recent developments in scan acquisition, data processing, and analysis.
2. Interpretation: Our findings provide an overview of the current methodologies used in longitudinal PET studies.
Sources of variability affecting the measure of change in amyloid and tau deposition are described, including how to address them, and which methodological features should be reported.

Future directions:
The manuscript proposes a framework to help increase the precision and accuracy of longitudinal studies, with experimental designs and PET-processing pipelines specific to those studies.Examples include: (1) further investigation of sex, race, and ethnicity as about the state and spread of AD and can present unique sources of variability.We then review and update the main sources of intrinsic, biological, and longitudinal PET measurement uncertainty, emphasizing the latest techniques in areas where progress has been made.So far, most longitudinal studies rely on processing and analyzing each timepoint independently, which in turn generates independent errors.
As the literature on longitudinal measures of brain atrophy using volumetric MRI has shown, [14][15][16][17] strategies that combine shared information through a joint analysis framework of all scans acquired on the same subject could reduce within-subject longitudinal measurement uncertainty.

MATERIAL & METHODS
Following preferred reporting items for systematic reviews and metaanalyses (PRISMA) guidelines, 18  The PubMed search identified 965 candidate publications, which then went through a two-step review (Figure 1).The first step involved screening the abstracts using the following inclusion criteria: (1) consisted of original work published in a peer-reviewed journal; (2)   reported a longitudinal PET study, that is, including at least one baseline and one follow-up PET scan acquired on the same individual; and (3) used an amyloid tracer ( 11 C-Pittsburg compound B ( 11 C-PiB), 18 Fflutemetamol, 18 F-florbetapir, 18 F-florbetaben, 18 F-NAV4694), a tau tracer ( 18 F-flortaucipir, 18 F-MK6240, 18 F-RO948, 18 F-PI2620, 18  in which the methodology used to obtain the outcome measure was not adequately described and (2) multiple papers that used identical methodology on the same cohort to answer different questions.In the latter case, this collection of papers was treated as a single entry for the purpose of the review.
After this two-step process, 80 publications remained.In line with PRISMA guidelines, an independent researcher assessed the final selection to ensure that all included studies met the inclusion and exclusion criteria.No discrepancies were noted.For each article included in the review, we pooled data and reported several features, as described in Figure 1.The search for those features was performed screening the methodology section of each article.Supplementary material and previously published worked were also searched, when referred to in the methodology.Finally, the quality of the studies was assessed based on several factors: the research question should be clearly stated, the population sufficiently defined, the methodology adequately described and research outcomes possibly generalizable.
Attention was also paid to potential selection, attrition, and reporting biases.A table listing all included publications and their key characteristics can be found in Supplementary Material.
This review is registered on PROSPERO (ID: CRD42021254695).

Characteristics of longitudinal PET studies identified
Some important trends can be observed from the papers included in the review (Figure 2).First, there is a growing use of longitudinal, quantitative amyloid and tau PET, as evidenced by the rapidly increasing number of articles published since the initial framework laid out by Schmidt et al. in 2015. 13Most longitudinal studies included in the current review used amyloid tracers (80%).This trend has started to shift with the increased availability of tau tracers allowing more studies to acquire enough longitudinal data to report findings, with tau studies representing 50% of the publications in the past 2 years.Second, more than half of the analyses have been performed on a few cohorts: the Alzheimer's Disease Neuroimaging Initiative (ADNI) and the Australian Imaging Biomarkers and Lifestyle Study of Ageing (AIBL), two valuable resources available to the global research community, as well as the Mayo Alzheimer's Disease Research Center (ADRC) and the Mayo Clinic Study of Aging (MCSA).As a result, the most common tracers used are the ones that have been selected for these studies: 11 C-PiB and 18 F-florbetapir. 18F-Flortaucipir, the only regulatory approved tau tracer thus far, was used in all tau studies except one using a second-generation tracer ( 18 F-MK6240).Third, because these larger cohorts acquire longitudinal PET data on multiple cameras, processing pipelines need to account for any differences across scanners.
Baseline and follow-up scans tend to be acquired on the same scanners; however, the proportion of subjects that change center or scanner in-between timepoints is rarely stated.Finally, the slow progressive build-up of amyloid plaques and tau tangles in AD led to longitudinal studies acquiring data over several years.In the selected articles, most individuals underwent PET scans around 2 years apart; however, 24% of studies reported a time interval between acquisitions of <2 years.
As can be expected, most studies acquired PET scans at two timepoints (64%), although a non-negligible proportion of studies acquired more (21% studies with three timepoints, 15% with four or more timepoints).
Our literature search highlighted that many previous studies were not consistently reporting key details, which would allow the reader to properly analyze and interpret the results.Recently published PET reporting guidelines require the specification of image resolution, reconstruction, post-reconstruction filtering, attenuation correction methods, and motion correction methodology. 19The significance and impact of those parameters on longitudinal studies and the potential importance of reporting specific technical features will be discussed further herein.In particular, reconstruction parameters were rarely reported, but could be inferred from the scanner.In the case of large cohorts such as ADNI, ADRC, or AIBL, those features can also be deduced from the main methodology papers for those studies.
However, only subsets of those data sets are often used, leading to difficulties in identifying the actual technical parameters relevant to one particular study.
One of the areas where technical parameters were often not reported was image pre-processing.Motion correction was mentioned in only ≈25% of studies, of which the main technique used was frameto-frame realignment.Furthermore, whether partial volume correction (PVC) was applied was reported in half of the studies; of those, 90% reported the specific approach implemented.The three main PVC techniques implemented were geometric transfer matrix (GTM) 20 in 48% of studies, two-compartment PVC 21 in 21% of studies, and three-compartment PVC 22 in 14% of the papers.
For studies focusing on global cortical or regional AD-protein burden, 98% of articles provided enough information about the relevant methodology for generating these regions of interest.In terms of reference regions used for standard uptake value ratio (SUVr) computations, 11 C-PiB studies favored the use of the cerebellar gray matter, and 18 Fflortaucipir more specifically the inferior cerebellar gray matter.Nearly ≈70% of studies used the whole cerebellum or the cerebellar gray matter while several publications also compared multiple reference regions.The overwhelming majority of studies relied on magnetic resonance imaging (MRI) to extract anatomic information.In that case, most commonly and when specified, the PET image is co-registered with the MRI (51%), and the analysis is then done in native PET space, native MR space, or aligned to Montreal Neurological Institute (MNI)152 atlas (35%) to ensure a consistent presentation of anatomy in a stereotactic coordinate system.Finally, most of the reviewed articles used SUVr from static acquisitions as the main outcome measure.Although full dynamic imaging provides many benefits, there are practical limitations, even more so for longitudinal studies.This may be why some studies with dynamic acquisitions available chose to only use SUVr from the late pseudoequilibrium frames.The Centiloid scale has started to emerge as a useful metric to provide some level of harmonization across cohorts, being used in five articles in the past 2 years-including two clinical trials. 10,23,24 summary, the review highlights that current studies primarily use publicly available data sets to assess change in amyloid over 2 years with static acquisitions.Image reconstruction and processing vary F I G U R E 2 Overview of the literature published in peer-reviewed journals relating the use of longitudinal amyloid or tau positron emission tomography (PET) studies.Percentages represent the number of studies that report using a specific feature (radiotracer, reference region, and so on).Because some studies used multiple tracers or compared different reference regions, these percentages may not add up to 100%.ADNI: Alzheimer's Disease Neuroimaging Initiative; ADRC: Alzheimer's Disease Research Center; DIAN: Dominantly Inherited Alzheimer Network; MCSA: Mayo Clinic Study of Aging; SUVr: standard uptake value ratio; SPM: statistical parametric mapping; AAL: automated anatomic labeling; FMM: 18 F-flutemetamol; FBB: 18 F-florbetaben; FBP: 18 F-florbetapir; PiB: 11 C-PiB; FTP: 18 F-flortaucipir.across studies; however, longitudinal data analyses most commonly rely on the annualized difference in SUVr, with the SUVr being computed independently at each timepoint.Although some elements of the analysis are commonly reported, several key features are omitted, making complete evaluation of the techniques difficult.

4
Measuring change in amyloid and tau deposition

Methodologies for computing rates of change in AD-protein deposition
Various methodologies are in use to measure the change in AD-protein deposition.The approach that requires the least amount of processing involves reconstructing and processing each scan independently.The annualized rate of change (ARC) between baseline and follow-up scans is the difference in protein load, most often in terms of SUVr, divided by the time interval between acquisitions.This rate of change can also be expressed as a percentage of change relative to the baseline value.
They are also flexible, allowing variations in the number of scans and time intervals between scans.From our literature screening, we estimate that around 17% of studies used mixed-effects models; however, no clear consensus was found among the studies as to whether common covariates, such as baseline age, sex, apolipoprotein E (APOE) ε4 carriership, and years of education, were essential to include in models in order to sufficiently explain variability in either the uptake values at baseline or rates of change over time.As a result, depending on the nature of the sample, different models should be tested in order to determine whether these covariates explained baseline values as well as their interaction between these covariates and time.
Alternative approaches to characterize changes in PET images have been proposed that incorporate features outside of the uptake relative to a pre-specified reference region.Shokouhi and colleagues calculated Pearson correlations between regional uptake from baseline and follow-up 18 F-fluorodeoxyglucose (FDG) data. 38This approach may be sensitive to changes over time, but it will not reflect whether changes represent hyper-or hypometabolism, or increased disagreement between scans.Another approach that combines information from two images into one metric is to derive a "slope" image created using a difference map. 39Rather than focusing on magnitude of intensities within a voxel or region, the overlap index (OI) identifies the within-subject change of spatial extent of voxels with elevated tau PET signal.This approach appears to be particularly sensitive to early tau build-up. 40Finally, one approach did not look at longitudinal changes between the two timepoints themselves, but rather in the change of a neuroimaging index based on differences in probability distribution between each image and a template of AD individuals as calculated by the Wasserstein distance. 41The template used in the longitudinal studies was based on a leave-one-out study of 30 AD subjects from ADNI, so how well it captures the heterogeneity and subtypes within AD is not clear.Except for the OI, the ability of these methods to detect small changes compared to more standard approaches remains unexplored.
Regardless of the method used to compute the rate of change, numerous biological and technical factors influence the precision of that measure.As a result, the annual accumulation rates of Aβ and tau are on the same order of magnitude as the test-retest tracer variability.[50] Variability was found to be noticeably higher for AD subjects compared to healthy controls. 44,46,49,503][54] The metric used to assess the rate of change should, therefore, always be reported with an estimation of its variability.

Variation in longitudinal PET measurements
Variability in longitudinal PET measurements can be decomposed into three elements: (1) the true variation in protein accumulation, (2) measurement uncertainty, and (3) biologic variability.
Following van Belle's statistical guidelines, 55 we refer to uncertainty as measurement uncertainty, that is, the degree of precision with which the protein load is measured, while the term variability refers to inherent, biological heterogeneity in the longitudinal PET data.This variability cannot be reduced but can be partially accounted for in the design of PET studies (e.g., standardization strategies, subject selection) and in some cases statistical modelling.On the other hand, uncertainty can be reduced by increasing the precision of PET measurements both cross-sectionally and longitudinally.These elements will be discussed in the following sections.Intrinsic variability in protein load over time

Genetic and demographic factors
Age, the female sex, other genetic factors (presence of the APOE ε4 allele, mutations in the genes encoding the presenilin 1 and 2 and amyloid precursor proteins) are established risk factors for AD.Reviews that summarise how these factors can account for inter-individual heterogeneity have been published. 13,56These risk factors accumulate over decades to convey different levels of risk.In a longitudinal study, the patient serves as their own control and quite often these confounds have far more impact explaining variability at baseline than on measured rates of change in protein deposition over a typical 2-year interval.
Although APOE is the largest genetic risk factor for sporadic AD, it is important to note that the level of risk conveyed by carrying an ε4 allele has been shown to vary by race and ethnicity.ε4 carriership increases risk of developing AD for White individuals but not significantly for African Americans and Hispanics. 57,58Several studies have assessed the impact of APOE ε4 carriership on the rate of AD-protein accumulation, with differing results.Some studies found that fibrillar amyloid accumulates at greater rates in APOE ε4 carriers than in noncarriers 34 ; some indicate that this difference occurs only during early stages of the disease 32,59 ; and others reported that carriership is not related to the rate of change but rather the age at onset. 31,60y elements to consider are that APOE ε4 carriers are more likely to accumulate Aβ pathology versus non-carriers, and importantly tend to develop the disease earlier, 34,61 which could explain higher rates of Aβ accumulation for APOE ε4 carriers in study groups with lower amyloid burden. 31,59A gene dose-dependent effect has also been suggested. 62reover, an APOE effect was also found for tau accumulation, with accumulation possibly being more prominent in ε4 carriers. 63Finally, the ε2 allele might have a mild protective role by slowing down amyloid accumulation. 27,32e female sex is another well-recognized risk factor, 64 although the influence of sex on longitudinal PET biomarkers has been relatively little studied. 65In contrast to amyloid, emerging studies focusing on tau suggest diverging courses of protein deposition over time.For instance, Smith et al. found a higher rate of tau accumulation in females compared to males. 66For cognitively normal individuals with high amyloid load, it has also been shown that females exhibited higher tau load in the entorhinal cortex compared to men. 65Further studies will be needed to confirm those observations and better understand the factors driving those differences.
Race and ethnicity are being investigated as additional causes of biological variability.In a single-site cohort from Washington University in St. Louis, there was no evidence of a cross-sectional difference in neocortical SUVr (as measured by 11 C-PiB or 18 F-florbetapir) between African American and non-Hispanic White subjects. 679][70] Results on the effects of race and ethnicity on tau PET uptake are also mixed.An additional analysis of the cohort in Morris et al. 67 showed no evidence of a difference in flortaucipir PET uptake based on race. 71However, Black/African Americans had elevated 18 F-flortaucipir PET binding in the choroid plexus (with a possible spill-in effect onto the hippocampus), 72 bilateral occipital, temporal, and superior frontal areas 73 compared to White individuals.This could be linked to non-specific binding of 18 F-flortaucipir to melatonin in the meninges, which is found in higher quantities in darker skin types. 74Further studies should be performed because differences in assessed tau load due to non-specific binding or other physiological issues could impact the definition of positivity thresholds.Although non-Hispanic White individuals are over-represented in longitudinal AD studies, including clinical trials, 75 investigating whether there is an impact of race and ethnicity on the rate of protein accumulation will improve our understanding of differing risk factors for AD and more comprehensively characterize the underlying AD pathology. 76erall, although these demographic and genetic factors increase the risk of developing AD, it is not clear whether this risk is conferred through increased rates of accumulation in AD pathology or if similar levels of pathology result in earlier onset.

Disease subtypes and staging models
[79] Identifying subtypes can help uncover group patterns that better characterize the heterogeneity of the disease.Multiple classification schemes have been developed to reflect AD variants based on heterogeneous clinical presentations and imaging patterns.For instance, quantification on a regional level and beyond dichotomization of amyloid and tau status to positive and negative could help predict interindividual as well as group trajectories.[86] These various patterns of protein deposition are not fully understood and could be linked to specific genetic susceptibility. 86

Blood flow
[91] Longitudinal studies might be particularly impacted by the presence of CBF changes over time, 91 which may occur with increasing age, disease progression, and therapeutic interventions. 92A simulation study by Ottoy et al. suggested that reduced CBF is expected in subjects with mild cognitive impairment (MCI) and AD patients, in particular in the posterior cingulate, which agrees with previous results by Cselényi et al. 93,94 These simulations were based on fully dynamic 18 F-florbetapir data and showed that changes in relative CBF can produce apparent changes in SUVr in patients with AD, increasing with the disease severity.This was also observed on 18 F-florbetaben data. 89ood-flow alterations can be accounted for through kinetic modelling of dynamic acquisitions to derive non-displaceable binding potential (BP ND ) or distribution volume ratio (DVR = BP ND +1).
Heeman et al. compared methodologies to obtain DVR from 18 Fflutemetamol and 18 F-florbetaben PET data and found that reference Logan quantification was more robust to changes in blood flow than the simplified reference tissue model (SRTM). 95These findings are in line with similar 11 C-PiB studies. 96Furthermore, the impact of blood flow on SUVr measurements has been shown to be more dependent on the reference region than the target, as CBF changes in the white matter (WM) will have more effect than when the cerebellum is used as the reference. 89,93nally, Cselényi et al. extracted a blood-flow component in cortical SUVr and found significant negative relationship between R 1 representing the ratio of tracer delivery and SUVr (≈0.04 unit increase in SUVr for every 0.1 unit decrease in R 1 at a constant DVR). 94The R 1 has demonstrated low test-retest variability over time (≈1.70% for a global composite region), which might make it particularly suitable as a surrogate measure of CBF in longitudinal studies. 91,97Overall, SUVr from 18 F-flutemetamol and 18 F-florbetaben data tend to overestimate protein binding via DVR, which should more closely reflect ground truth, with this effect being proportional to the Aβ load. 98This overestimation was also observed in a 18 F-THK-5351 tau PET study. 99ngitudinal studies with individuals that are cognitively impaired or have higher amyloid burden could be particularly affected by this bias, leading to an under-or overestimation of their rate of amyloid accumulation.Despite the challenges of implementing longitudinal, dynamic PET studies, 13% of articles in the reviewed literature used dynamic acquisitions.

6
Reducing PET measurement uncertainty

Subject motion
Head motion can degrade the image quality and subsequently the accuracy of protein quantification.For instance, simulating realistic motion, Chen et al. found an ≈4% relative change between corrected and noncorrected FDG data. 100Regional and voxel-wise analyses, particularly relevant to study tau spreading over time, can be impacted severely by motion-related errors as regions that are smaller, more lateralized with a high degree of curvature (such as the lateral orbital frontal cortex) are more likely to be misquantified. 100tion can occur between the attenuation correction scan and the PET acquisition leading to emission-attenuation mismatch between timeframes and within (averaged) PET timeframes, as detailed previously by Schmidt et al. 13 To help address the first two, in the case of a PET-CT (computerized tomography) acquisition, the amyloid PET Quantitative Imaging Biomarkers Alliance (QIBA) profile recommends co-registering the PET frames to the CT image prior to attenuation and scatter corrections, and if not feasible, to remove frames or whole scan when motion exceeds 4 mm or 4 degrees. 101Motion correction is still an active area of research, with recent developments following progresses in optical tracking and neural nets. 102,103Within-frame motion remains the most challenging to detect and, therefore, correct; however, novel techniques relying on multiple reconstructions from very short PET frames have been developed that can address this issue. 104e full range of motion-correction techniques is beyond the scope of this review, but are summarized elsewhere. 105udies focusing on when to perform motion correction could be valuable, as mis-registration and interpolation (particularly when motion is minimal), can also introduce inaccuracies in the final PET measure. 106,107In addition, serial PET measurements will likely be highly affected by motion, as differing types and magnitudes of motion across timepoints could lead to an under-or overestimation in rates of change computations, even more so in regional analyses.Although motion correction has a significant impact on the accuracy of the final metric, most often the use of such techniques is not reported (in 25% of articles screened), contrary to the latest PET-reporting guidelines. 108

Duration of acquisition
The duration of the acquisition can influence the accuracy of the PET measure.A continuous, dynamic acquisition from tracer injection to pseudo-equilibrium allows quantification of tracer uptake with standard kinetic-modelling techniques.As mentioned previously, this approach is more accurate than SUVr, since kinetic modelling can account for the effects of blood flow.On the other hand, SUVr derived from static acquisition reduces cost and burden to the patient due to the shorter acquisition time required.Motion is also likely to be reduced during these shorter acquisition periods.
Given that the kinetics of some tracers require nearly 2 hours from injection to reach pseudo-equilibrium, dual time-window protocols have recently been proposed to allow kinetic modelling while reducing the total acquisition time.A patient is first scanned during tracer injection and wash-in, and then a second time when the tracer has likely reached pseudo-equilibrium.Kinetic modelling is then performed by combining the two sections of the time-activity curve using interpolation to fill the region of the curve in between the two acquisitions.An optimized dual-time window for amyloid tracers 18 F-florbetaben and 18 F-flutemetamol, which was designed to minimize the bias and the number of outliers, was 0-30 min post injection (p.i.) and 90-110 min p.i. 109 Bullich et al. reported similar timings (0-30 and 120-140 min p.i.).
Dual time window acquisitions have also been investigated for tau tracers.Protocols similar to the ones described previously herein have also been implemented for 18 F-flortaucipir (0-30 min p.i. and 80-100 min), 110 with one caveat for 18 F-flortaucipir being that many studies suggest that the tracer uptake does not reach a plateau but continually increases after the time of injection until the end of acquisition. 111,112A dual time-window strategy was also achievable for 18 F-MK-6240, with the acquisition phases being 0-30 min p.i. and 90-120 min. 113Kolinger et al. noted that the dual-phase protocol might be less robust to blood-flow variations than a full dynamic acquisition.In order to limit inter-and intra-individual variability arising from variations in time windows, a time correction was implemented by Pontecorvo et al. 53 For each voxel, a linear regression is fitted through all Long axial field of view PET scanners Current PET scanners have an axial field of view (FoV) of ≈20 cm, which can cause lower parts of the brain to be cropped or to be noisy due to their proximity to the edge of the FoV.This is especially important as the cerebellum is very often used as a reference region, which can lead to an overestimation of the SUVr, as noted by Schmidt et al. 13 In longitudinal studies, differences in the positioning of the subject within the FoV at different timepoints can bias the estimate of change in protein load.
This can be addressed by acquiring data on long axial FoV scanners, 114 such as the two first total body PET-CT scanners, the EXPLORER (United Imaging Healthcare Co, Shanghai, China) 115 and the Biograph Vision Quadra (Siemens Healthineers, Knoxville, TN, USA), 116 with axial FoV of 194 and 106 cm, respectively.[118] Although more sensitive PET scanners could be particularly advantageous for longitudinal studies, they might not be well suited for claustrophobic patients, and their high cost might be a major limitation to their widespread use, especially for AD-related brain imaging.

Dose-reduction strategies
Dose-reduction strategies could be relevant for longitudinal studies, and worth investigating if PET is to be used more frequently to monitor disease progression and effects of treatment.These strategies are promising and provide clear benefits to patients.Apart from a reduced risk for patients, it could also allow more detailed investigations of differences between tracers if the dose is sufficiently reduced to stay within radioactive substance regulations.Although dose-reduction strategies have been implemented for FDG, 119,120 fewer studies exist for amyloid and tau PET acquisitions.Nai et al.
showed that reconstructing images from low-count PET images using 18 F-THK5351 and 11 C-PiB data introduces varying degrees of bias depending on the processing pipeline.The DVR resulted in the best quantitative accuracy, with <7% bias across regions. 121An ultralow-dose strategy has been developed by Chen et al., which uses a deep learning approach to synthesize PET images from multiple MR sequences and a low-count PET image. 122Long axial FoV scanners discussed previously could lead to lower dose requirements, although the impact on quantification beyond visual read should be evaluated.In longitudinal settings, dose-reduction strategies may allow more frequent PET acquisitions; however, distinguishing noise from the actual variation in Aβ or tau with an additional measurement bias and a shorter time interval in-between scans might not be feasible at this time.In longitudinal studies, PET quantification can particularly be affected by PVE for subjects with brain atrophy. 123Compared to amyloid, tau studies are more likely to require regional protein estimates and study individuals with more advanced AD and atrophy.They are, therefore, especially prone to PVE quantification errors.The adoption of PVC in routine image processing has been limited so far (41% of our selected publications).Importantly, these methodologies require an accurate specification of the point spread function model and the precise coregistration of PET and MR, as small misalignments can cause PVC methods to add noise rather than reducing it. 106,1075][126][127] Increases in rates of change due to PVC may affect subjects with higher amyloid more, as suggested by Brendel et al. who found a decrease in percentage change over time in healthy controls and MCI subjects, versus an increase in AD subjects. 125Such differences were also observed by Rullman et al., who found a significant effect of PVC on Aβ-positive subjects, but none on Aβ-negative subjects, 126 and a significant positive effect of PVC in MCI and AD groups in another study. 127These results suggest that atrophy-related PVE occurs at a higher rate in MCI/AD subjects, could mask the true PET signal increase, and could, to some extent, be recovered by PVC.
On the other hand, some studies found no apparent effect of PVC on the rate of change of protein deposition for 18 F-flortaucipir scans 128,129 and 11 C-PiB. 2,130The difference in findings could be attributed to shorter time intervals in-between PET acquisitions, making it more difficult to detect PVC-related changes.
Finally, the choice of PVC implementation can have a substantial effect on variability.The most commonly used techniques in longitudinal PET studies include the two-compartment model or Meltzer method, 21 the three-compartment or Müller-Gärtner method, 22 and the geometric transfer matrix (GTM) by Rousset et al. 20 Some studies used the region-based voxel-wise (RBV) correction 123 and iterative Yang technique. 131Schwarz et al. evaluated the performance of the former three techniques longitudinally.Although GTM was the most used technique in our literature screening, it was found to have less relative precision assessed by the residuals from mixed-effects models compared to Meltzer and Müller-Gärtner style PVC (longitudinal variability of ≈4% to 8% compared to ≈2% for traditional methods). 132he GTM method seems to offer better group separation between impaired and unimpaired individuals, but it suffers from larger levels of uncertainty.
Although PVC was reported in only 40% of the articles in his review, the findings support a wider use of PVC for longitudinal studies, especially for individuals with brain atrophy-provided some quality control is in place to ensure an accurate PET/MR registration when MR is used.In reviewing the relevant literature, GTM was the most common PVC technique implemented in nearly half of the studies using PVC; however, traditional two-compartment and three-compartment PVC seemed to provide more stable estimates of amyloid accumulation and might, therefore, be preferable to GTM.Further studies comparing the precision, accuracy, and robustness of these PVC techniques in longitudinal settings are warranted.

Attenuation correction
The Erlangen, Germany) and the Signa PET/MRI (GE Healthcare, Waukesha WI, USA).However, several alternative methods have been proposed to improve the accuracy of attenuation correction for PET/MR systems. 134tenuation correction technique was reported in 39% of the reviewed studies: 31% based on CT and 60% on transmission scans.
No MR-based methodologies were mentioned specifically, although it is probable that the default attenuation correction from the scanner manufacturer was used in those cases.
Regardless of the methodology, the accuracy of the attenuation correction strongly depends on an accurate estimation of the attenuation-correction map and co-registration to the PET scan, with any misalignment likely due to subject motion. 135Chen et al. simulated misalignment, and estimated a bias in SUVr of 2.5%, 4%, and up to 10% in respectively low, moderate, and large motion models. 100y motion-induced cross-sectional uncertainty could have an even higher impact for longitudinal imaging, although to the best of our knowledge, no study on that topic has been conducted.An accurate registration of the μ-map to the PET scan or the implementation of a motion correction algorithm could be used to address this issue.

Reference region
Even in fully dynamic acquisitions, kinetic modelling using arterial blood sampling is rarely used, thus a reference region consisting of tissue with minimal uptake needs to be defined.The reference region chosen to normalize tracer uptake is one of the most common sources of variability between scans.It is influenced by many factors, including the choice of tracer, the scanner's axial FoV, the precision of the alignment with the images defining the regions of interest (ROIs), the quality of the parcellation, atrophy, and its sensitivity to variations in blood flow. 95Overall, the choice of reference region is still a subject of debate and varies with the context of the study.The interpretation of SUVr values will also have to be adapted to the reference region used, as the resulting values have different ranges (see Figure 3).
The most common reference regions used in the literature are the whole cerebellum, the cerebellar gray matter, the pons, the brainstem, the (eroded) supratentorial white matter (sWM), and a composite reference typically combining some or all of these regions. 48Due to non-specific uptake near the falx cerebri, tau tracers often further restrict the reference region to inferior regions of cerebellar gray matter.
Several studies suggested the use of the sWM as a suitable reference region for longitudinal studies. 33,48,124,125,136,137This finding could be due to this reference region being large and more contiguous, and, therefore, less susceptible to noise, PVE, and registration errors.
In addition, compared to the more widely used cerebellar and pontine reference regions, the sWM might be less sensitive to differences in the patient's head positioning within the FoV across several timepoints, whereas SUVr normalized to the cerebellum can be overestimated in cases where the cerebellum is cropped or near the edge of the image due to small axial FoV in some scanners. 48,137 a result, compared to other reference regions, SUVr computed using sWM showed significantly less variability over time and greater power to detect change in protein deposition, as well as less implausible negative trends over time in studies acquired with 18 Fflorbetapir, 48,124,125,136,137 11 C-PiB, 33,138 and 18 F-flortaucipir. 54,139 alternative approach for 18 F-flortaucipir using a subject-specific WM reference region based on the Parametric Estimation of Reference Signal Intensity (PERSI) was also found to be more reproducible and stable over time than other common reference regions, 139 resulting in a tau metric more sensitive to change.
Despite these benefits favoring the use of the sWM over other reference regions, several concerns have been raised both for amyloid and tau regarding the stability of the sWM over time.The sWM uptake has been shown to increase with age as well as cortical amyloid load, 33,140 and the pharmacokinetic properties of sWM differ from the gray matter.A 18 F-florbetapir study showed that sWM is more susceptible to blood-flow alterations than the cerebellum, 140 and suggests that the stability of the sWM over time is mainly due to lower statistical noise compared to the cerebellum.Regarding studies showed that the sWM had low sensitivity to differentiate between Aβ-positive and Aβ-negative individuals and was less stable over time than the cerebellar gray matter. 141,142Moreover, in longitudinal studies, sWM quantification might also be impacted by changes in WM integrity, further increasing the variability of the SUVr.Reduced amyloid uptake has been linked to WM hyperintensities 143 ; however, a review paper from 2017 aggregating results from 34 studies suggests that there is no direct association between the two. 144Regarding tau, a study by Moscoso et al. demonstrated a decreased retention of 18 Fflortaucipir in regions with WM hyperintensities, and that up to 57% of the variance in longitudinal SUVr linked to variations in the reference region tau uptake. 145nally, the use of WM as reference region is not recommended for conversion to the Centiloid scale (CL) as it was shown to add significant measurement variability compared to cerebellar-based reference regions 146 and could lead to underestimation of the amyloid load. 147vel methodologies have been created to address issues caused by dependency on anatomically defined reference regions: the evaluation of brain amyloidosis (ELBA), 148 the Aβ-PET pathology accumulation index (Aβ index), 149 AMY Q, 148,150 and the amyloid pattern similarity score (AMPSS). 151These methods are independent of a reference region and are described in more detail later.Although these datadriven approaches show promising results, the advantage of the SUVr lies in its simplicity and therefore wide usage across studies with static acquisitions.

Regions of interest
Identifying an appropriate target region will also impact the performance of longitudinal analyses.Common choices are to use a global neocortical ROI, selecting individually defined anatomic regions that have been identified as being particularly vulnerable to AD pathology or composite regions designed to capture a summary of AD-related changes.On one hand, global regions will likely be more robust to noise and measurement error than smaller individual regions; however, they may combine areas of high change and low change together, reducing the ability to detect changes in the early phases of the disease.Creating a composite region attempts to find the balance between the twomore sensitive than a global measure but less susceptible to error than an individual region.For amyloid tracers, the composite region often comprises the frontal, anterior/posterior cingulate, lateral parietal, and lateral temporal regions. 48The regions commonly used in tau imaging tend to be grouped according to Braak staging 152,153 : amygdala, entorhinal cortex, fusiform, parahippocampal, and inferior temporal and middle temporal gyri.Due to spill-out effects from non-specific binding in the choroid plexus, accurately quantifying tau binding in the hippocampus using 18 F-flortaucipir can be challenging, 154,155 but it may be mitigated with effective application of PVC. 156Longitudinal trajectories at the regional level for amyloid 157,63,125,62,158,36,159 and tau 5,77,86,54 have been investigated but are beyond the scope of this review.
Various atlases have been employed for the definition of the target region, and the corresponding labelling protocols for key structures may vary from atlas to atlas.In the studies included in this review, five atlases were commonly used: the Desikan-Killiany atlas, 160 implemented in FreeSurfer in 25% of studies 31,63 ; the Automated Anatomical Labeling atlas, 161 in around 15% of studies 27,35 ; the Hammers atlas, 162 available in PNEURO/PMOD, in around 13% of studies 36,59,124,125,163,164 ; the Mayo Clinic Adult Lifespan Template (MCALT) atlas, in 4% of studies 77,129,165 ; and the Harvard-Oxford atlas (2%). 83,166Neuroparc, a standardization method across brain atlases has been developed recently, 167 and could help with comparing data obtained with different atlases.
Apart from the choice of reference and target regions, the accurate co-registration between PET and atlas is critical to the precision of lon- ing clinical trials. 10,230][171][172] These limitations as well as the increased number of longitudinal PET studies (see Figure 2) triggered the development of novel, alternatives methods to quantify the protein load.
Data-driven PET analysis methods rely on the assumption of a multivariate pattern of amyloid deposition that can be captured across the population.Several metrics use image decomposition to extract the specific tracer binding.The amyloid load (Aβ L ) and tau load (Tau L ) by Whittington and Gunn describe the SUVr trajectory over time using a logistic growth model, assuming that each brain region has a maximum capacity to carry amyloid. 52,173,174Many methods use dimensionality reduction techniques such as principal component analysis (PCA) 149,150 or non-negative matrix factorization (CL NMF ) 169 to compactly represent this multivariate pattern, often optimizing the weighting of the largest component to create an adaptive template to obtain more accurate registrations. 149,175Further improvements have explored how to make these methods more robust to different tracers 169 and not require a reference region. 149,150Also based on image decomposition, the specific Aβ load (SAβL) uses deep learning to generate a map of non-specific binding from MR sequences.This map is then subtracted from the original SUVr image to obtain a map of specific binding. 176,177The ELBA takes a very different approach and tries to translate visual interpretation of images mathematically.
It does so by capturing the variation in the image intensity using two criteria: a metric of iso-intensity surface complexity and one that evaluates the histogram propensity toward higher values.These metrics are optimized and combined to form the ELBA. 148ong all these methods, the Aβ L , Tau L , ELBA and CL NMF have been evaluated against longitudinal data.The Aβ L showed increased sensitivity assessed by the effect size of the change in amyloid burden (Hedges g = 0.49 compared to 0.36 for SUVr) and increased ability to separate scans between clinical groups.Another study using AβL also found lower longitudinal variability in AβL compared to SUVr, but only for amyloid-negative subjects. 178The equivalent approach for tau tracers, Tau L , showed greater effect size than SUVr for the change in tau deposits between six clinical groups, which suggested that smaller sample sizes would be required to show a 25% reduction in tau accumulation over 1 year (power = 80%, α = 0.05).The longitudinal consistency was assessed for ELBA by using the standard deviation of the residuals from least square linear regression with respect to time between scans.The ELBA technique performed better than SUVr, with a variability normalized by the interquartile range of ±2.3%.Finally, compared to CL, the CL NMF was shown to have greater longitudinal consistency with fewer abnormal changes over time, to be more robust to change in tracer between timepoints, and to have a stronger association with baseline value (Spearman ρ of 0.39 vs 0.21 for CL).
Although promising, these novel metrics should be evaluated further in longitudinal settings.

Toward simultaneous image reconstruction, joint PET image processing, and analysis
The most common approach to analyzing longitudinal PET data is to obtain outcome measures from each timepoint independently.This is a practical solution, particularly if different timepoints have been acquired on different scanners.However, images acquired from the same subject represent repeated measures, which contain large amounts of shared information that could increase the precision of longitudinal protein change by reducing the (differential) image noise.
This advantage has been clearly demonstrated in research on structural MRI. 14,16,179There are many points within the reconstruction and pre-processing pipeline where this shared information may be leveraged.

Joint image reconstruction
Ideally, the optimal place to incorporate shared information would be the earliest point of analysis, during the image reconstruction.
Ellis and Reader 180 developed a joint reconstruction method for two images that incorporate priors based on shared characteristics of the images. 181,182Joint reconstruction was also implemented in a framework using linear temporal dependency and spatial alignment between scans to build a difference image with greater contrast, which resulted in reduced sample sizes to detect change in SUVr. 183These methodologies have been tested primarily for a situation with two scans; the feasibility of these approaches with multiple scans remains unclear.
F I G U R E 4 Overview of registration strategies used in longitudinal PET pipelines.
Implementation within a clinical setting could be challenging due to the increased storage and computational requirements of retaining raw list mode data from all timepoints, as this format of the data needs orders of magnitude more storage than the resulting reconstructed images.

Joint image processing
Besides reconstruction, there are still other steps later in the processing stage where shared information can be leveraged.Defining a common space for image analysis is particularly relevant in the case of longitudinal studies where two or more scans are compared at the individual level.Here we focus on registration strategies in two contexts: PET-only pipelines and PET-MR or PET-CT pipelines.The most common registration strategies from the reviewed literature are described in Figure 4.
The final analysis space varies between studies; the most used ones are the native PET and native MR space, as well as a standard (usually MNI) space.Many studies opt for MNI space to normalize their data and help compare scans both at the individual and at the group level.However, Kolinger et al. compared the impact of the reference space for analysis on the SUVr and found no significant difference between native and standard space, but any image feature using volumetric information should be performed in native space. 184th the increasing availability of longitudinal registration pipelines for structural MRI, a few studies have begun to implement them into longitudinal analysis of amyloid and tau PET data. 26,39,54,66,185Longitudinal registration pipelines are available in SPM, 179 FreeSurfer, 16 and ANTs. 186These pipelines create unbiased, within-subject templates that tend to be equidistant from each timepoint included in the registration.In short, PET images are first rigidly registered to the MR closest to it in time.The serial MR scans are then co-registered to form the within-subject template.Next, the parcellation can be derived from the within-subject template or the parcellations are first obtained in native MR space and transformed into the midpoint space.By overlapping the parcellations from all timepoints, the final selection of ROIs can be cropped to include only voxels belonging to all timepoints.This more conservative approach of excluding voxels within a region could be particularly useful when substantial atrophy over the course of the study is likely. 185ing these longitudinal techniques from high-resolution MRI could improve the precision of longitudinal PET measurements by reducing the noise introduced by distinct processing pipelines, as well as reducing potential sources of bias.Early longitudinal pipelines of structural MR, which transferred follow-up scans onto the baseline for analysis, have been shown to cause a bias arising from the asymmetry of transformations and subsequent resamplings applied to the follow-up scans in comparison to the baseline. 187,188Although evaluation of longitu- Further approaches that use all of the timepoints to produce longitudinally consistent outcomes should be examined.For instance, besides generating a consistent parcellation, the longitudinal registration pipeline could be applied to generate a shared attenuation correction map, independently of which modality it is derived from, or use the same one for both timepoints.Future studies could also investigate incorporating shared information to dynamic acquisitions, for instance by performing a joint kinetic modelling of all timepoints.

DISCUSSION
In this review we present an overview of the state-of-the-art of lon-  by the Radiological Society of North America QIBA 101 should be followed.Regarding tau, an international consensus for the use of 18 F-flortaucipir was also published. 191arting with the acquisition of the scans, several aspects of the processing can be harmonized such as the effective spatial resolution of PET images obtained from different reconstruction protocols 192,193 and the data acquired in multiple centers. 194,195andardization of the processing and analysis can then be considered, including standardization of the PET outcome metrics across amyloid, 150,169,196,197 and tau tracers or using open-source processing pipelines such as the PET Unified Pipeline PUP, 198 the robust PET-only processing rPOP, 199 or the Computational analysis of PET by AIBL CapAIBL. 200Establishing standards of reporting ARC measures would also increase the comparability and reproducibility across studies.
In the absence of ground-truth, the evaluation of comprehensive reproducible pipelines is challenging.However, several strategies can be implemented to determine how to generate the most precise and clinically valuable ARC estimates (see more details in Supplementary Material).Overall, there is a need for unified frameworks specifically designed for longitudinal quantitative PET studies.
sources of variability; (2) increased use of longitudinal registration pipelines; (3) guidelines defining the context of use for motion and partial volume correction; and (4) critical role of harmonization and standardization strategies.
relevant peer-reviewed articles were identified by conducting a search on the 22/11/2021 using the PubMed database.The following query was used: ((longitudinal) OR (followup) OR (follow up) OR (serial)) AND ((PET) OR (Positron Emission Tomography [MeSH])) AND ((amyloid) OR (tau)) AND ((dementia) OR (Alzheimer*)) AND (brain).In conjunction, further relevant papers were searched by cross-checking pertinent references from the publications screened.No date restriction was applied to the search; only papers in English language were considered.F I G U R E 1 PRISMA flow diagram-PubMed search on November 22, 2021.OSEM: ordered subset expectation maximization algorithm; ROI: region of interest.
F-GTP1), or a combination thereof.Exclusion criteria were as follows: (1) cross-sectional PET studies; (2) studies without PET data; (3) studies not related to dementia; (4) reviews with no or very little mention of the methodology of the articles screened; (5) animal studies.This initial selection process yielded 152 relevant articles that were suited for further evaluation.In the second step, we reviewed the methodology and results using the following additional inclusion criteria: (1) the outcome measure involves change between timepoints or a rate of change in AD-protein deposition over time and (2) the change or rate of change in AD-protein deposition must be one of the dependant variables of analysis in the paper.Finally, we excluded: (1) articles available PET time frames to create adjusted images representing a consistent post-injection time window.Dual time-window protocols might be beneficial to implement in longitudinal studies, as reducing the total acquisition time required to obtain suitable data for kinetic modelling may lead to more accurate quantification by accounting for confounds such as tracer delivery compared to SUVr approaches, while the shorter acquisition time should decrease the level of motion.
Partial volume effect (PVE) is one of the main physical factors limiting the quantitative accuracy of PET.It can be seen as the combination of two effects.First, due to the discrete sampling, a single voxel in the reconstructed image can contain a mixture of tissues with different levels of uptake (gray matter, white matter [WM], dura, meninges, cerebrospinal fluid [CSF]).Second, the relatively poor PET spatial resolution (≈3 mm to ≈7-8 mm depending on the scanner), modelled via the point spread function, causes spill-out/spill-in of one region onto another.
accurate quantification of the PET signal is heavily dependent on the accuracy of the attenuation correction.The original attenuation correction methods relied on a PET transmission scan using Germanium-68.With the advent of PET-CT systems, a low-dose CT was acquired to generate an attenuation coefficient map by converting Hounsfield units into linear attenuation coefficient via a piecewise linear transformation. 133PET-MR presents different challenges for attenuation correction, as standard MR-imaging techniques cannot provide an accurate estimate of an object's electron density, leading to difficulties distinguishing bone from air and other tissues, and the smaller bore and strong magnetic field do not allow for a PET transmission scan to be performed.MR-based attenuationcorrection sequences are available for the two currently available PET-MR scanners: the Biograph mMR (Siemens Healthcare GmbH,

F I G U R E 3
18 F-florbetaben, Example of an SUVr parametric map obtained from a 18 F-flutemetamol PET scan.The color bars illustrate the different ranges of SUVr depending on the choice of reference region.
gitudinal PET measurements.Even in PET-MR scanners, motion can introduce a misalignment between the atlas derived from the MR and the PET data.According to Schwartz et al., who quantified the impact of this misalignment, motion accounts for ≈1% of the test-retest variability of amyloid PET, which in turn results in a rate of change with a variability of ≈7%. 107The choice of registration software is also crucial as shown by Markiewicz et al., who compared four common registration softwares and found that Statistical Parametric Mapping (SPM, Wellcome Trust Centre for Neuroimaging, London, UK) 168 resulted in the lowest measurement uncertainty.Novel "data-driven" metrics of amyloid and tau deposition The more established protein deposition metrics-including SUVr, BP ND , the CL-all present limitations.Although SUVr is simple to compute and can be derived from a short static acquisition, it is dependent on many factors (e.g., choice of reference region [see Figure 3]; length of the FoV; quality of the parcellation; precision of the alignment between the MR, the CT, or the atlas; stability of that region over time; variations in blood flow).The BP ND is a more accurate measure of the density of available receptors as it considers the effect of blood flow, but it requires long, dynamic acquisitions that cannot be implemented easily in a clinical setting or standardized across multiple centers.CL offers an opportunity to harmonize SUVr-based metrics, using post hoc linear transformations to a scale calibrated by a reference 11 C-PIB-based SUVr data set.The SUVr remains the most established metric, whereas the use of Centiloid has steadily increased since its development in 2015 and is now used in many studies includ- dinal pipelines has focused predominantly inwards around accurately measuring change in various structural metrics of neurodegeneration, initial evaluations to the impact of PET have begun.Both the MR study of Tustison et al. and the tau PET study by Schwartz et al. found that longitudinal pipelines were able to produce more stable measures of change over time and increase cognitive group separation 186,189 ; however, what elements of longitudinal pipelines have the most impact on improving precision requires further work.
As a result, addressing the sources of variability summarized here is not always feasible because of time, cost, skills required, or software availability constraints.Several questions about how and when to best address those sources of variability remain to be explored.First regarding biological factors, although the impact of atrophy and blood flow fluctuations can be mitigated using PVC and dynamic imaging, both of those solutions can introduce additional noise.Their appropriate use in a longitudinal setting is dependent on the advancement of AD at the individual level, and on the true amyloid accumulation within the time window inbetween PET scans.For instance, for subjects in the later stages of the disease more likely to have pronounced atrophy, PVC is necessary to avoid underestimating the AD-protein accumulation.Defining the level of atrophy above which PVC should be applied, taking into account the duration of the study, investigating the best PVC method to use longitudinally for all available tracers, or presenting results both with and without atrophy could help address some of the debates around PVC usage.How to balance precision and accuracy should also be explored to evaluate the acquisition of longitudinal static versus dynamic acquisitions.Indeed, although the latter is less susceptible to variations in blood flow, the BP ND or DVR derived from these acquisitions might be noisier, have a narrower dynamic range and are less feasible, in terms of time and cost, for many clinical research scenarios than SUVr and CL outcomes based on static acquisition.Another factor that could play a role in AD-protein accumulation is resilience in the form of cognitive reserve and brain maintenance.Investigating how those mechanisms influence AD-protein accumulation could possibly enhance disease-prediction models.Despite the numerous questions around how to control or mitigate biological and physiological variability, there are many areas where technical factors could be improved further and reduce measurement error.Although the impact of motion on quantification has been investigated, and many motion correction approaches exist, it has yet to be consistently implemented in processing pipelines and reported in the methodology of articles.For motion correction, guidelines defining acceptable levels of motion based on whether the gain in accuracy and the gain in signal is greater than the uncertainty introduced by the registration and resampling/interpolation processes could be developed.As motion is not always easily detectable and data sets continue to grow in size, systematic motion detection and quantification techniques will be particularly important to ensure that automated pipelines produce accurate AD-protein quantification.Further research could focus the characterization of the ARC (1) at the individual level and (2) across the AD continuum, considering the non-linearity of amyloid and tau accumulation over time.In addition, consistent strategies to define accumulators versus stable trajectories could increase comparability of different studies.On the other hand, a varying percentage of negative ARC is also observed, especially in individuals with low amyloid burden.Better understanding of the potential technical factors leading to those results could help either improve the performance of the ARC metric or better define the amyloid hypothesis in the earliest stages of the disease.The next crucial step to improve the interpretability and reproducibility of longitudinal studies is to generalize the use of standardization and harmonization strategies.Multinational and international standards developed for the use of amyloid PET by the Society of Nuclear Medicine and Molecular Imaging and the European Association of Nuclear Medicine 190 and one of the biggest sources of uncertainty, and this could be addressed through longitudinal registration pipelines.Finally, the low temporal precision of the ARC relative to the small magnitude of change detectable means that studies should include always include an estimation the ARC variability.compartmentor three-compartment PVC for 11 C-PiB).The studies selected and reviewed depict the state of the art of the literature focused on longitudinal amyloid and tau PET.One limitation, however, is that numerous studies use longitudinal PET without ARC as one of the main dependent outcomes analyzed.We also acknowledge that the level of details in reporting methodologies is guided by the focus of the paper and limited by journal editorial constraints, and that more comprehensive information could be available upon request.Finally, translating technological progress into methodological advances achievable in a research and clinical setting can be difficult.
implemented using robust and accurate registration techniques, as it has been shown to lead to up to 7% imprecision in the measure of change in SUVr.Next, many studies suggest that PVC should be used, especially in the case of longitudinal tau studies, where spill-in from off-target binding and substantial evidence of correlation between tau deposition and atrophy patterns could introduce errors in measurements of change.Quantification errors, or inconsistencies between timepoints, is