Aberrant pattern of regional cerebral blood flow in Alzheimer's disease: a voxel-wise meta-analysis of arterial spin labeling MR imaging studies

Many studies have applied arterial spin labeling (ASL) to characterize cerebral perfusion patterns of Alzheimer's disease (AD). However, findings across studies are not conclusive. A quantitatively voxel-wise meta-analysis to pool the resting-state ASL studies that measure regional cerebral blood flow (rCBF) alterations in AD was conducted to identify the most consistent and replicable perfusion pattern using seed-based d mapping. The meta-analysis, including 17 ASL studies encompassing 327 AD patients and 357 healthy controls, demonstrated that decreased rCBF in AD patients relative to healthy controls were consistently identified in the bilateral posterior cingulate cortices (PCC)/precuneus, bilateral inferior parietal lobules (IPLs), and left dorsolateral prefrontal cortex. The meta-regression analysis showed that more severe cognitive impairment in the AD samples correlated with greater decreases of rCBF in the bilateral PCC and left IPL. This study characterizes an aberrant ASL-rCBF perfusion pattern of AD involving the posterior default mode network and executive network, which are implicated in its pathophysiology and hold promise for developing imaging biomarkers.


INTRODUCTION
Alzheimer's disease (AD), the most prevalent type of dementia in the aging population [1][2][3], is featured by memory disturbance, attentional and executive deficits, and visuospatial and perceptual impairments [4]. The neuropathological hallmark of AD is the progressive accumulation of amyloid beta (Aβ) plaques and tau-related neurofibrillary tangles, and eventually accompanied by the damage and death of neurons in the brain [5]. The underlying neurobiology is far from being complete and no effective medications are available today for AD to slow or halt the damage and destruction of neurons [2]. This disorder has caused a substantial burden not only on the patients and their caregivers but also on the socioeconomic system [6]. Understanding the neural basis and early detection of AD are therefore very important.
Arterial spin labeling (ASL) MR imaging is a noninvasive technique that can quantitatively measure cerebral blood flow (CBF) by magnetically labelling the inflowing arterial blood water in vivo as an endogenous tracer [7][8][9]. ASL allows an automated voxel-by-voxel statistical analysis for regional CBF (rCBF) differences without any priori hypothesis [10][11][12]. Regional CBF is recognized as a reflection of intrinsic neural activity and brain physiology, which has been validated in normal aging and neuropsychiatric disorders [13][14][15].
in the AD brain is closely matched with the metabolism pattern measured using PET [10,[17][18][19][20]. In addition, ASL offers a similar diagnostic ability as PET in the detection of AD [18][19][20][21]. Owing to its ease of acquisition, non-invasiveness, non-radiation, and reliability, ASL is increasingly proposed as an alternative to PET and holds promise for developing imaging biomarkers [13,15,22].
Consequently, it would be of great interest for this study to conduct a timely meta-analysis of ASL studies to identify the most consistent and replicable perfusion pattern of AD. This voxel-wise meta-analysis utilized the anisotropic effect-size version of seed-based d mapping (SDM) software [43,44], which has been extensively applied in previous meta-analyses of neuroimaging studies for neuropsychiatric disorders [44][45][46][47][48][49].

Included studies
According to the strategy of literature search and study selection, a total of 17 ASL studies that investigated rCBF differences between 327 AD patients and 357 healthy controls were finally eligible for the meta-analysis [10, 11, 23-30, 37-42, 50]. Figure 1 presents a flow diagram of the studies that met the criteria for this metaanalysis. Of these included studies, thirteen were published in English and the other four were in Chinese [30,37,40,41]. Fourteen out of the 17 studies were performed on the 3.0T MRI scanning systems and the other three were on the 1.5T MRI systems. Regarding the techniques applied to measure resting-state CBF in these studies, ten studies used pulsed ASL (PASL); five used pseudocontinuous ASL (pCASL); and the remaining two used continuous ASL (CASL). The quality score of each included study was no less than 8.5 (total score = 10), which indicates that the quality is acceptable. Table 1 summarizes the demographic, clinical, and technical characteristics as well as the quality scores of the ASL studies included in the meta-analysis. Supplementary Table 1 presents the diagnostic criteria, disease stage, cognitive test, and vascular burden assessment for AD of the included studies in the meta-analysis. arterial spin labeling, rCBF, regional cerebral blood flow. www.impactjournals.com/oncotarget

Regional CBF differences by pooling all included studies
The voxel-wise SDM analysis demonstrated that decreased rCBF in AD patients compared to healthy controls were mainly located in the bilateral posterior cingulate cortices (PCC)/precuneus, bilateral inferior parietal lobules (IPLs), and left dorsolateral prefrontal cortex (DLPFC). In contrast, no regions showed significant increases of rCBF in AD patients relative to healthy controls. The SDM results are described in Table 2 and illustrated in Figure 2.

Reliability analysis
The whole-brain jackknife sensitivity analysis revealed that decreases of rCBF in the bilateral PCC/ precuneus and bilateral IPLs were the most robust findings, replicable in all 17 combinations. Decreases of rCBF in the left DLPFC remained highly replicable, as it was preserved in all but four in combinations of studies (Table 3).

Publication bias analysis
No publication biases for the identified brain regions with rCBF differences between AD patients and healthy controls were observed, which was revealed by the approximately symmetric funnel plots (Supplementary Figure 1) and the non-significant Egger's tests ( Table 2).

DISCUSSION
To the best of our knowledge, this is the first quantitative meta-analysis to pool the resting-state ASL studies to identify the most consistent pattern of rCBF and to explore its clinical correlation in AD. A total of 17 studies as measured with ASL in 327 AD patients and 357 healthy controls were included in this voxel-wise meta-analysis. Decreases of rCBF in AD patients relative to healthy controls are consistently identified in the bilateral PCC/precuneus, bilateral IPLs, and left DLPFC. Furthermore, the meta-regression analysis indicates that more severe cognitive impairment in the AD samples was associated with greater decreases of rCBF in the bilateral PCC and left IPL.
The bilateral PCC/precuneus and bilateral IPLs, where this study observed the most consistent brain regions of decreased rCBF in AD, are key nodes of the posterior default mode network (DMN) [51,52]. The DMN, which is comprised of highly interconnected brain areas involved in higher cognitive functions, is the most active brain system at rest in healthy subjects [51]. There is compelling evidence suggesting that the DMN, especially the posterior part is functionally and structurally impaired in AD patients [53][54][55][56][57][58][59] and atrisk subjects [12,[59][60][61]. Dysfunction of the DMN, a hallmark of AD, has been linked to core memory, attentional, and visuospatial deficits [62,63]. The DMN nodes, such as the PCC/precuneus, parietal and temporal lobes were confirmed to be more vulnerable to early amyloid deposition [54,55,64]. High amyloidbeta deposition in the DMN was demonstrated in AD patients [54,[65][66][67][68]. In addition, some studies suggest that decreases of rCBF in these regions, especially in the PCC/precuneus, may reflect a remote functional deficits caused by the neuronal damage in the medial temporal structures [10,28,[69][70][71][72]. As such, decreases of rCBF in the bilateral PCC/precuneus and bilateral IPLs observed in this study may be a reflection of pathophysiological process that involves vascular dysfunction and neuronal degeneration underlying AD [14]. The perfusion pattern identified in the current metaanalysis is largely in line with the metabolism pattern from the voxel-wise meta-analyses of FDG-PET studies in AD patients [49,73]. In addition, the perfusion pattern using ASL in AD are concordant with the perfusion pattern using other MRI modalities, such as H 2 15 O-PET, 99m Tc-hexamethylpropyleneamine oxime (HMPAO)-single photon emission computed tomography (SPECT), and 99m Tc-ethyl cysteinate dimmer (ECD)-SPECT [73]. These data suggest that ASL provides information comparable to above invasive imaging techniques and shows potential as a reliable and safe alternative. The findings identified in the present voxel-wise meta-analysis are also consistent with the ASL studies using ROI analyses [10,27,34,74,75]. Decreases of rCBF in these areas could serve as specific regions of interest for a further diagnostic utility. Decreases of normalized rCBF in the bilateral PCC/ precuneus had an accuracy of 86.0% (91.3% sensitivity and 80.0% specificity) in distinguishing AD patients from healthy controls [26]. Brain activity changes in the DMN may be an early marker for AD. Alexopoulos et al identified hypometabolism/hypoperfusion consistently in the DMN, especially the posterior part, such as the PCC/ precuneus and IPLs in amnestic mild cognitive impairment (aMCI) converters [73]. They further demonstrated that hypometabolism/hypoperfusion in the IPLs was the most reliable predictor of the progression from aMCI to AD [73]. A recent study observed a continuing decrease of CBF in the PCC/precuneus and other related regions in the continuum of AD [76]. Moreover, our meta-regression analysis shows that the severity of cognitive impairment in the AD samples was associated with the rCBF changes in the bilateral PCC and left IPL, which suggests that altered rCBF in these regions may act as an imaging marker for tracking disease progression.
Besides the areas of rCBF changes observed in the DMN in AD patients compared with healthy controls, we also identified the regions of decreased rCBF belonging to the central executive network (CEN), such as the DLPFC and posterior parietal areas [77,78]. The CEN is another intrinsic brain network that is known to be involved in executive functioning, particularly important for maintaining highlevel cognition [78,79]. Aberrant functional connectivity in the CEN was observed in AD [80][81][82][83][84][85]. Decreases of rCBF in the regions of the CEN probably account for the executive deficits in AD patients [79].  Atrophy of gray matter [86] and white matter [87] in the medial temporal lobe (MTL) is a characteristic and could serve as a neurostructural predictor of AD [88]. Interestingly, we did not observe consistent rCBF changes in MTL in AD patients relative to healthy controls in the present meta-analysis. This structural-functional discordance has been frequently detected in previous studies [10,28,[69][70][71]89], which is interpreted as a compensatory response to morphologic alterations [71]. Therefore, rCBF changes in the MTL may not be sensitive enough to distinguish healthy elders from AD patients, as aging-related tau pathology, hypometabolism, and hypoperfusion in the MTL were also observed in normalaged individuals [68,[90][91][92].

Limitations and future perspectives
Several limitations in this study should be acknowledged. First, a huge body of ASL studies in AD was excluded because of the chosen voxel-wise approach and this approach was based on summarized coordinates and their effect sizes rather than on raw imaging data or statistical brain maps, which might limit its accuracy [93]. In addition, the patients in the included studies were the clinical samples, who compared with community-based normal control volunteers. In this context, although they were matched or adjusted regarding age, sex ratio and education, some other critical factors, such as socioeconomic status, vascular risk burden, cognitive reserve, racial/ethnic make-up, and genetic vulnerability were not addressed in most of the original studies, which might lead to some heterogeneity in the conclusions and remains to be further addressed. Further analysis of multicenter raw imaging data in large homogeneous samples, like ASL-MRI scans from the Alzheimer's disease Neuroimaging Initiative (ADNI) subjects [94], would confirm the present findings. Second, our meta-analysis that synthesized the findings from the cross-sectional studies could not determine whether decreases of rCBF in the identified brain areas are the cause or consequence of AD [14]. Longitudinal studies could provide further insights. Third, ASL acquisition parameters, pre-and post-processing steps, such as scanner field-strength, inversion time, labeling duration, post label time delay, volume coverage, readout approaches, partial volume correction, and GM correction, may bias the results that warrant careful consideration by investigators. Further investigations in high field-strength MRI scanners with optimized and standardized imaging acquisition and analytical methodology are recommended [15].

Literature search and study selection
A comprehensive literature search was performed in the PubMed, Web of Science, and Embase databases up until 16 December, 2016, using the keywords "Alzheimer's Disease" AND "arterial spin labeling". The China National Knowledge Infrastructure (CNKI) database was searched for additional studies published in Chinese. Reference lists from relevant studies were manually reviewed for further eligible studies. Studies were included in the meta-analysis if they met the following criteria: 1) the study used the standard diagnostic criteria for Alzheimer's Disease [95][96][97][98][99]; 2) the study utilized resting-state ASL MR imaging to measure rCBF differences between Alzheimer's Disease and healthy controls; 3) the study applied a voxel-based statistical analysis; 4) the study reported three-dimensional coordinates in Talairach or MNI space; and 5) the study was published as an original article (not as a letter or a meeting abstract or a comment) in a peer-reviewed journal. Studies were excluded if they specifically used region of interest (ROI) approaches. Studies were also excluded if they did not report significant results with three-dimensional peak coordinates. To avoid duplication, only the study with a larger sample size was included in case that the patient populations overlapped. The quality of each eligible study was evaluated with a 10-point checklist (Supplementary Table 2) based on previous neuroimaging meta-analyses [100,101]. This study followed the MOOSE guidelines for the meta-analyses of observational studies [102].

Data analysis Main voxel-wise meta-analysis
A meta-analysis of rCBF differences between AD patients and healthy controls was performed using the SDM software package (www.sdmproject.com). The SDM approaches have been described in detail previously [44-47, 93, 103-105]. We briefly summarized here. Peak coordinates and effect sizes (e.g., t-values) of rCBF differences between AD patients and healthy controls were firstly extracted from each study [103,105]. The SDM software then separately recreated a standard MNI map of rCBF for each study applying an anisotropic Gaussian kernel (full width at half maximum [FWHM] = 20 mm) [44,103,105]. The mean map was generated by voxelwise calculation of the mean of the study maps with a standard random-effects model, taking into account the sample size, the intra-study variability, and the betweenstudy heterogeneity [44,103,105]. To generate significant results and the final map of rCBF, we applied the default SDM kernel size and threshold (uncorrected voxel p = 0.005, peak height Z = 1, cluster extent = 10 voxels), which is equivalent to a corrected p value of 0.05 and is found to optimally balance false positives and negatives [103,105]. Results were visualized with the BrainNet Viewer [106].

Reliability analysis
To test the reliability of the findings identified in the main voxel-wise meta-analysis, a whole-brain voxel-based jackknife sensitivity analysis was performed by iteratively repeating the same analysis, excluding one study at a time [44,46,93,103].

Analysis of publication bias
Possible publication bias was examined with a standard meta-analysis using the Stata 12.0 software (Stata Corp LP, College Station, TX, USA) by extracting the values from the relevant peaks from the main voxel-wise meta-analysis. An asymmetric funnel plot and a p-value less than 0.05 for Egger's test were considered significant.

Meta-regression analysis
A meta-regression analysis was conducted to assess the severity of cognitive impairment examined by MMSE scores that correlate with the ASL results. A stringent threshold of p = 0.0005 and a cluster extent of 10 voxels were used for the meta-regression analysis [103,104].

CONCLUSIONS
Our study shows the most consistent and replicable decreases of rCBF in the bilateral PCC/precuneus, bilateral IPLs, and left DLPFC in AD patients compared with healthy controls via the voxel-wise meta-analysis of ASL studies. These aberrant regions predominantly involve in the default mode and central executive networks that are implicated in the AD pathophysiology. This study further demonstrates that reduced rCBF in the bilateral PCC/precuneus and left IPL was related to the severity of cognitive impairment in AD patients, which suggests that altered rCBF in these regions may act as an objective imaging marker for tracking AD progression.

Author contributions
HRM, PLP and SHC designed the protocol. HRM and HCS wrote the main manuscript. LQS, JHC and ZYD obtained the data. PLP, GDW and RL analyzed the results. PRX and JGZ revised the manuscript. All authors reviewed the manuscript.