Hypo- and hyper-perfusion in MCI and AD identified by different ASL MRI sequences

Arterial spin labeling (ASL) perfusion MRI has been increasingly used in Alzheimer's Disease (AD) research. However, ASL MRI sequences differ greatly in terms of arterial blood signal preparations and data acquisition strategies, both leading to a large difference of signal-to-noise ratio (SNR). It is of great translational importance to compare the several widely used ASL MRI sequences regarding sensitivity of ASL measured cerebral blood flow (CBF) for detecting the between-group difference across the AD continuum. To this end, this study compared three ASL MRI sequences in AD research, including the 2D Pulsed ASL (PASL), 3D Background Suppressed (BS) PASL, and 3D BS Pseudo-Continuous ASL (PCASL). We used data from 100 healthy and cognitively normal elderly control (NC) subjects, 75 patients with mild cognitive impairment (MCI), and 57 Alzheimer’s disease (AD) subjects from the AD neuroimaging initiative (ADNI). Both cross-sectional perfusion difference and perfusion versus clinical assessment correlations were examined. The major findings included: 3D PCASL sequence identified stronger patient versus control CBF/rCBF differences than 2D PASL and 3D PASL; MCI showed reduced CBF and CBF redistribution; CBF in orbito-frontal cortex presents a new U-shape change pattern from normal aging to MCI and to AD; 3D PCASL identified a negative rCBF to memory correlation while 2D PASL showed a positive correlation.


Introduction
Alzheimer's disease (AD) is the most common type of dementia and the 6 th leading cause of death in the United States (Solis et al., 2020). AD has been characterized by amyloid-beta deposition and tau pathology, but these pathological changes often do not lead to clinical symptoms and are often insensitive to disease progression. Functional brain measures such as cerebral blood flow (CBF) are necessary to understand the AD pathology versus symptom discrepancy. CBF is a key physiological index that reflects cerebral metabolism associated with regional functional activity (Baron et al., 1982;Detre et al., 2012;DeWitt et al., 1988;Furlow et al., 1983;Liu et al., 2004;Musiek et al., 2012;Raichle, 1998;Vestergaard et al., 2016). CBF has long been investigated as a regional brain marker for brain function (Detre et al., 2009 and has been increasingly studied in AD (Alsop et al., 2000Swinford et al., 2022;Wang, 2014;Xu et al., 2010;Ze Wang et al., 2013).
Arterial Spin Labeling (ASL) MRI is a completely noninvasive technique for measuring CBF (Detre et al., 1992). Using the magnetically labeled arterial blood as the endogenous tracer, ASL MRI measures the perfusion signal through the tissue signal changes induced by the labeled arterial blood water that exchanged with tissue water. To control the background tissue signal, the arterial spin labeling radiofrequency pulses are phase modulated to have zero net effects on the arterial blood water magnetization. The control image acquired using this arterial blood magnetization untagged sequence is considered the baseline Data used in the preparation of this article were obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database (adni.loni.usc.edu). As such, the investigators within the ADNI contributed to the design and implementation of ADNI and/or provided data but did not participate in the analysis or writing of this report. A complete listing of ADNI investigators can be found at: http://adni.loni.usc.edu/wp-content/uploads/how_to_apply/ ADNI_Acknowledgement_List.pdf.

3
image. The perfusion-weighted signal is then extracted from the difference between the control and label images Luis Hernandez-Garcia et al., 2022;Wang, 2022). The quantitative CBF can be calculated from the ratio between the weighted perfusion signal and the MR signal at the fully relaxed condition -the M0, using the singlecompartment model (Alsop et al., 2014). Over the past decades, a variety of ASL techniques have been proposed, which can be roughly divided into four classes: pulsed ASL, continuous ASL (CASL), pseudo-CASL (PCASL), and velocity-selective ASL (VS-ASL) (Luis . PASL uses a short-duration RF pulse to invert magnetization of arterial blood water, while CASL uses a long RF pulse with small flip angle to continuously drive the arterial blood water magnetization from the positive to the negative direction. PCASL represents the current state-of-art ASL technique. As an extension of CASL, PCASL replaces the long RF pulse in CASL with a train of small RF pulses which can be executed in most of MR scanners without violating the coil hardware limit (Dai et al., 2008). Another big advantage is that PCASL implements the control labeling through alternating the phase of the successive RF pulses without flipping the labeling plane to the other side of the imaging place, resulting in a better background signal control than CASL. VS-ASL can theoretically avoid the post-labeling delay time and may become a future standard in ASL  but thus far VS-ASL has been rarely used in AD research. Both PASL and PCASL have been implemented as commercial products by major MR vendors and are widely available in the clinical MR machine.
Since the inception of ASL MRI, it has been applied to a large body of translational research including aging and AD studies (Detre et al., 2009Haller et al., 2016;Telischak et al., 2015). ASL MRI is particularly appealing to aging and AD research because it is non-invasive and non-radioactive and can be repeated many times (Wang, 2014;. More importantly, the regional CBF maps measured by ASL MRI can be used as a marker of regional brain function and its alterations by disease pathology. In fact, low CBF may even represent a major cause of AD pathology and subsequent cognitive decline (Zlokovic, 2005). Over the past decades, ASL MRI has been increasingly studied in the AD continuum, and the most consistent CBF change pattern in AD is the hypoperfusion in the temporoparietal cortex, which correlates with disease severity (Hu et al., 2010;Wang, 2014;Ze Wang et al., 2013) In addition to these cross-sectional findings, two recent studies have reported the longitudinal CBF reductions in AD and patients with mild cognitive impairment (MCI) (Camargo et al., 2021;Duan et al., 2021).
While the findings shown in current ASL MRI-based AD research literature are highly encouraging, a challenge is that AD ASL MRI data have been acquired with different ASL MRI sequences, especially in some multi-site longitudinal studies such as the Alzheimer's Disease Neuroimaging Initiative (ADNI)(http:// adni. loni. usc. edu/). It is unknown how these acquisition variabilities (sequences and scanner vendors) affect the CBF alteration detections in the AD continuum. Thus far, only one paper has been published assessing the effects of three ASL MRI sequences on CBF difference detection between MCI and normal elderly control (NC) (Dolui et al., 2017). The data were acquired using a Siemens MR machine with a 2D PASL, a 2D PCASL (without background suppression (BS)) and a 3D PCASL (with BS) sequence developed at the University of Pennsylvania (Vidorreta et al., 2012(Vidorreta et al., , 2014Wu et al., 2007). The current paper differed from the previous study by comparing three product ASL MRI sequences from two major MR vendors for assessing CBF difference between NC and MCI and AD. The sequences included the 2D Gradient Echo-Planar Imaging-based PASL MRI sequence from Siemens Healthineers (abbreviated as 2D PASL), the 3D Turbo Gradient Spin Echo-based PASL MRI sequence from Siemens Healthineers (3D PASL), and the 3D Fast Spin Echo-based PCASL MRI sequence from GE Healthcare (3D PCASL). The MR vendors included Siemens and GE. 2D PASL and 3D PASL are both based on Cartesian readout but the former is without BS and the latter is with BS; 3D PCASL is based on a stack of spiral readout and BS (Dai et al., 2008). Our first aim was to compare the sequences for differentiating patients with MCI from NC and patients with AD from MCI, and AD from NC. The second aim was to examine whether CBF calculated from the different ASL MRI data shows different sensitivity to the potential associations between regional CBF in NC, MCI, and AD subjects and memory as measured by the immediate recall (LIMM) total score (Chelune et al., 1990). We focused on memory because memory loss is a hallmark symptom of AD (Jahn, 2013) and LIMM is a typical measure of memory function with lower LIMM scores corresponding to severer memory decline. Previous studies have shown a positive correlation between CBF (measured by ASL MRI) and the LIMM score (Leeuwis et al., 2017(Leeuwis et al., , 2018. Comparing the CBF versus memory correlations revealed by different ASL sequences can provide additional information for evaluating the sequences for AD research.

Subject information
All data included in this study were from the ADNI including both the phase 2 and 3 (ADNI 2, ADNI 3). ADNI initially was launched in 2003 by the National Institute on Aging (NIA), the National Institute of Biomedical Imaging and Bioengineering (NIBIB), the US Food and Drug Administration, private pharmaceutical companies, and nonprofit organizations, as a 60 million, five-year public-private partnership. Michael W. Weiner, MD, from San Francisco Veterans Affairs Medical Center and the University of California-San Francisco, is the principal investigator of ADNI. Subjects included in this study were separated into three groups (NC, MCI, and AD) according to their diagnostic status at the time the ASL MRI images were acquired. The 2D PASL data contained 25 NC, 25 MCI, and 25 AD subjects; the 3D PCASL data had 40 NC, 25 MCI, and 16 AD subjects; the 3D PASL data contained 35 NC, 25 MCI, and 16 AD. For this study, the inclusion criteria were that all the subjects must have ASL MRI (2D or 3D) and the structural MPRAGE images. Age, sex, years of education, and LIMM scores were extracted from the files PTDEMOG.csv, and NEUROBAT.csv, respectively. The scores were matched according to the ASL MRI data acquisition dates. Note that the 2D PASL data had more MCI patients than the other two data types, we randomly selected 25 to be included in this study to match the number of MCI patients in the other two ASL datasets. Tables 1, 2 and 3 show the demographic and clinical information for 2D PASL, 3D PCASL, and 3D PASL.
Absolute CBF value can be calculated from both the 2D PASL data and the 3D PCASL data but not the 3D PASL data because of the lack of M0 images in the 3D PASL data. To allow comparisons among the three types of ASL data, we calculated relative CBF for all three datasets.

2D PASL data processing
Similar to our previous studies (Camargo et al., 2021;Ze Wang et al., 2013), ASLtbx (Wang, 2022;Wang et al., 2008) was used for preprocessing all MR images. The processing steps included motion correction (Wang, 2012), temporal denoising through a high pass filtering and temporal nuisance regression, spatial smoothing, CBF quantification, outlier cleaning (Li et al., 2018), partial volume correction, and spatial registration to the Montreal Neurology Institute (MNI) standard brain space. Temporal nuisances including head motion time courses (3 translations and 3 rotations), and the cerebrospinal fluid (CSF) mean signal time course were regressed out from ASL image series at each voxel. CSF mask was defined during the structural MR image segmentations (see below). Spatial smoothing was performed with an isotropic Gaussian kernel with a full-width-at-half-maximum of 6 mm. The preprocessed ASL label and control image pairs were successively subtracted. The difference was subsequently converted into a quantitative CBF map using the one-compartment model included in ASLtbx.
Structural images were segmented into grey matter (GM), white matter (WM), and CSF using the segmentation tool provided in SPM12. The GM/WM/CSF images were projected into the native ASL image space based on the spatial transform obtained from the mean ASL control image versus structural MRI registration. The GM/WM/ CSF probability was used to correct partial volume effect (PVE) at each GM voxel using a previously described approach (Du et al., 2006). The PVE corrected CBF map was then registered into the structural image space using the same mean ASL control image versus structural MRI registration transform. We calculated rCBF maps by dividing the absolute CBF at each voxel by the whole brain mean CBF value (mean CBF of the GM and WM). GM and WM masks were created during structural MRI segmentation and registration (see above). For the CBF or rCBF versus the LIMM score correlation analysis, we only focused on regions defined by the meta-region-ofinterest (meta-ROI) (Landau et al., 2012) which consists of seven spherical regions-of-interest (ROIs) in the left and right hippocampus, left and right angular gyrus, left and right temporal lobe, and precuneus. CBF and rCBF were extracted from the five ROIs: the composite meta-ROI, the bilateral hippocampus, bilateral angular gyrus, bilateral temporal lobes, and precuneus. Correlation between the LIMM score and mean rCBF of each ROI was assessed separately using simple regression. Age and sex were included as covariates. These correlation analyses were performed for NC, MCI, and AD separately for each type of ASL CBF image.

3D PASL and 3D PCASL data processing
The processing steps for 3D ASL data were very similar to the 2D ASL processing with a few changes depending on the labeling approach. For 3D PASL data, rCBF was computed at each voxel by dividing the perfusion signal by the whole brain mean (mean value of the GM and WM). For 3D PCASL, rCBF images was calculated by dividing the absolute CBF map by the whole brain mean (mean CBF of the GM and WM. GM and WM masks were obtained through structural MRI based brain segmentations). Structural MRI segmentations for both 3D PCASL and 3D PASL were performed using FreeSurfer (surfer. nmr.mgh.harvard.edu/). The corresponding GM/WM/CSF masks were registered into the native ASL MRI space to be used for calculating rCBF and for extracting the temporal nuisances for data preprocessing. CBF and rCBF were extracted from the five ROIs and correlated to LIMM using the approach described in previous section.

Statistical analysis
One-way analysis of variance (ANOVA) was used to assess the group differences for continuous variables and X 2 for the categorical variables (sex). The results were presented in the form of mean ± standarddeviation . Partial correlation analysis was used to assess the correlations between regional CBF and the LIMM score; age, sex, and years of education were used as nuisance variables and the results were corrected for multiple comparisons using Bonferroni correction. These neurobehavior scores were assessed on dates closest to that of the ASL-MRI scans.
Site effects were controlled with ComBat (Fortin et al., 2017). Two-sample t-test as implemented in SPM12 was used to assess the voxel-wise CBF difference between NC and MCI, MCI and AD, and NC and AD. Sex, age, and years of education were included as covariates. The same voxelwise statistical analyses were performed for both the absolute CBF and rCBF separately. A voxel-wise statistical significance threshold was set to p-value < 0.001. The Monte Carlo simulation-based cluster size estimation was used for correcting the multiple comparisons. The cluster size threshold was found to be < = 200, then used to threshold the suprathreshold clusters for all the analysis. BSPVIEW (bobspunt.com/software/bspmview) was used to visualize the image wise statistical analysis results.

Subject information
Significant differences of MMSE and LIMM were found between NC, MCI, and AD subjects in the three cases: 2D PASL, 3D PCASL, and 3D PASL. Lower MMSE or LIMM scores in aging and AD often indicate severer cognitive decline or impairment. For the cohort included in this study, both MMSE and LIMM scores decrease with disease severity in the direction of NC > MCI > AD (see Tables 1, 2 and 3). Based on 2D PASL, mean GM CBF in the NC/MCI/AD were: 26.61/19.55/18.13 ml/100 g/min. Mean GM CBF was not significantly different between NC and MCI (p = 0.086), but was significantly different between MCI and AD (p = 1.00e-6) and between NC and AD (p = 1.14e-7). For 3D PCASL data, the mean GM CBF of NC/MCI/AD was 29.59/27.51/26.77 ml/100 g/min, respectively. Mean GM CBF did not significantly differ between NC and MCI (p = 0.9487), but significantly differed between MCI and AD (p = 3.40e-6) and between NC and AD (p = 5.13e-5). Sex distributions in each of the NC, MCI, and AD group were significantly different among the three ASL data types (p < 0.0113, F test). Age, years of educations, and LIMM significantly differed among the three ASL MRI sequences for the MCI group (p < 7.5e-4, F test) only. They did not show significant difference among the three ASL data types for both the NC and AD group (p > 0.15). Figure 1 shows the results of the voxel-wise rCBF comparisons (two-sample t-test) between NC and MCI subjects for 2D PASL, 3D PCASL, and 3D PASL. Figure 2 shows the voxel-wise NC versus MCI CBF comparison results for 2D PASL and 3D PCASL. Regardless of CBF and rCBF, 3D PCASL showed the highest sensitivity in terms of the group level CBF difference cluster size and peak t-values in the fronto-parietal regions. Using 2D PASL, MCI showed statistically significant lower rCBF compared with NC subjects in the right superior orbital gyrus, right rectal gyrus, and right inferior temporal gyrus. For 3D PCASL, MCI patients showed significant lower rCBF compared to NC subjects in the right middle frontal gyrus; and a significant higher rCBF in MCI compared to NC subjects in the right superior medial gyrus, left precuneus, and right thalamus. For 3D PASL, significant lower rCBF in MCI as compared to NC was found in the left and right temporal gyrus, right middle temporal gyrus, right amygdala, and left caudate nucleus; significant higher rCBF in MCI compared with NC were found in the left superior parietal lobule, right superior frontal gyrus, right precuneus, and right cerebellum. Figure 3 shows the voxel-wise MCI versus AD rCBF comparison (two-sampled t-test) results for 2D PASL, 3D PCASL, and 3D PASL. Compared to the 2D PASL sequence, the 3D sequences showed higher sensitivity for the MCI minus AD rCBF difference detection. For 2D PASL, compared to MCI, AD showed significant higher rCBF in the medial orbito-frontal cortex, lateral orbitofrontal cortex, and caudate, but lower rCBF in the cerebellum. For 3D PCASL, compared to MCI, AD had significant higher rCBF in medial orbitofrontal cortex, lateral orbito-frontal cortex, striatum, temporal pole, middle temporal cortex including hippocampus, insula, and superior temporal cortex, dorsal anterior cingulate cortex, and cerebellum, but showed significant lower rCBF the superior part of the prefrontal cortex, the superior part of the caudate, posterior cingulate cortex, and parietal cortex. For 3D PASL, significant higher rCBF in AD compared to MCI was found in the orbito-frontal cortex, insula, striatum, temporal cortex, anterior cingulate cortex, fusiform, visual cortex, and cerebellum. Figure 4 shows the voxel-wise MCI-AD absolute CBF comparison results for the 2D PASL and 3D PCASL data. For both types of data, the MCI-AD absolute CBF difference patterns (Fig. 4A for 2D PASL, Fig. 4B for 3D PCASL) were quite similar to those identified by the corresponding rCBF ( Fig. 3A and B for 2D PASL and 3D PCASL, respectively). Figure 5 shows results of the voxel-wise NC versus AD rCBF comparisons (two-sample t-test). 3D sequences showed higher sensitivity than 2D. 2D PASL did not show  1 3 any significant rCBF difference between NC and AD. Using 3D PCASL, AD had significant lower rCBF in left and right inferior and superior parietal lobule, left and right middle occipital gyrus but higher rCBF in the left and right cerebellum, left and right superior orbital gyrus, left and right hippocampus, left amygdala, and left putamen. Based on 3D PASL, AD showed significant lower rCBF in the right thalamus, left and right caudate nucleus, left putamen, left superior frontal gyrus, and left and right precentral gyrus, but higher rCBF in the left and right rectal gyrus, left and right temporal lobe, and left orbito-frontal gyrus. Figure 6 shows the voxel-wise NC-AD absolute CBF comparison results for the 2D PASL and 3D PCASL. 2D PASL revealed significant lower CBF in AD compared to NC in the left precuneus. 3D PCASL showed significant CBF reductions in AD compared to NC in the left and right cerebellum, left putamen, left insula lobe, left and right temporal lobe, and left and right superior orbital gyrus.

Discussion
Using existing data from ADNI, we compared three commercial ASL MRI sequences: 2D PASL, 3D PCASL, and 3D PASL for the efficacy of detecting the CBF and rCBF differences between NC and patients with MCI or AD and between MCI and AD. Our results showed that 3D PCASL showed the highest sensitivity for the between-group CBF or rCBF differences than the other two ASL sequences; 2D PASL showed the lowest sensitivity for detecting those differences. It did not even show any statistically significant CBF or rCBF difference between NC and AD. For neurocognitive correlation analyses, 2D PASL showed positive rCBF versus memory correlations in angular gyrus and the composite meta-ROI; while 3D PCASL revealed negative correlations between memory and rCBF in angular gyrus, the composite meta-ROI, and hippocampus. This study provided the following novel contributions and findings: first, we assessed cross-sectional CBF and rCBF difference among three types of cohorts: NC, MCI, and AD using three types of ASL MRI in a single study; second, the cross-sectional CBF/rCBF difference identified by 3D PCASL are spatially more distributed than those reported in the literature; third, compared to NC, we found mainly lower absolute Fig. 6 CBF difference between NC and AD identified by two types of ASL data. A) 2D PASL, B) 3D PCASL. Hot color means higher CBF in NC; blue color means lower in NC. The color bar represents statis-tical significance (t) values. The number underneath each image slice indicate the slice location in the MNI standard brain space CBF in MCI but both higher and lower CBF in AD; fourth, the rCBF based group comparisons showed lower frontal rCBF and higher parietal rCBF using the 3D PCASL and 3D PASL sequences; fifth, absolute CBF in orbitofrontal cortex appeared to be consistently affected by disease status change: it first reduced in MCI compared to NC and then increased in AD as compared to MCI.
The sensitivity differences between the three commercial ASL MRI sequences are consistent with the ASL technique development and evaluation literature (Fernandez-Seara et al., 2005;Gunther et al., 2005;Mutsaerts et al., 2014Mutsaerts et al., , 2015Nanjappa et al., 2021;Steketee et al., 2015;Vidorreta et al., 2012). Higher sensitivity of the 3D sequences may be contributed by BS, which significantly increases the signal-to-noise-ratio (SNR) of the ASL MRI signal by nulling the strong background tissue signal. The relatively higher sensitivity of PCASL compared to PASL may be contributed by the stronger signal from the labeled arterial blood during the longer arterial spin labeling duration in PCASL compared to PASL. It may also be contributed by the better background signal control in PCASL (Alsop et al., 2014;Jezzard et al., 2018;Luis Hernandez-Garcia et al., 2022).
The cross-sectional comparison results were consistent with the AD/MCI ASL imaging research literature. For both 3D PASL and 3D PCASL, a common lower CBF and rCBF pattern in MCI relative to NC was found in the left and right middle orbital gyrus and left and right temporal lobe. Temporal lobe is pivotal in memory and is associated with auditory information processing, language, emotion, and part of visual perception. Temporal lobe CBF reductions in both MCI and AD have been observed using ASL MRI in several studies Dai et al., 2009;. Hypoperfusion in MCI as compared to NC was found in the left and right side of precuneus, angular gyrus, and cerebellum. Hypoperfusion in precuneus was consistent with previous findings reported by other groups and us (Binnewijzend et al., 2013;Camargo et al., 2021;Chao et al., 2009). Hypoperfusion in cerebellum was consistent with our recent ADNI data based longitudinal CBF study (Camargo et al., 2021), and may reflect a decline of cerebellum functions in MCI (Jacobs et al., 2018). 3D PCASL revealed a statistically significant global CBF reduction in MCI patients relative to NC, which is supported by our previous work (Ze Wang et al., 2013) and by the CBF reduction trend in MCI compared to NC reported in (Lövblad et al., 2015). After controlling this global CBF change in the rCBFbased analyses, 3D PCASL revealed hyperperfusion (higher rCBF in MCI than NC) patterns in the limbic area, prefrontal cortex, temporal cortex, posterior cingulate cortex/precuneus while the hypoperfusion (lower rCBF in MCI than NC) patterns in orbito-frontal cortex, temporal cortex, and temporal cortex remained. The presence of both rCBF-derived hypoperfusion and hyperperfusion in MCI (relative to NC) indicates a redistribution of CBF in MCI, which might represent a mechanism for compensating the functional impairment in MCI. Our data did not show global CBF difference between MCI and AD. Both 2D PASL and 3D PCASL found higher CBF in medial orbito-frontal cortex in AD relative to MCI. 3D PCASL showed spatially more distributed hyper-perfusion (higher in AD than MCI) and hypoperfusion (lower in AD than MCI) in AD compared to MCI. The rCBF-based MCI-AD differences in 2D PASL and 3D PCASL are quite similar to those found in the absolute CBF images. 3D PASL only showed hyperperfusion in AD compared to MCI (higher rCBF in AD than MCI), which is not consistent with the aforementioned AD CBF literature. The NC-AD differences are approximately the summation of the NC-MCI and MCI-AD differences. 2D PASL did not find any statistically significant CBF/rCBF differences between NC and AD. This mis-detection suggests a low sensitivity of 2D PASL, which may be mainly caused by the low SNR. 3D PCASL revealed reduced CBF in most of the cortex except for the prefrontal part. Using rCBF, 3D PCASL and 3D PASL both showed hyperperfusion (higher rCBF in AD than NC) in prefrontal cortex and the anterior and lateral part of the limbic system but hypo-perfusion (lower rCBF in AD than NC) in the back of the brain. This bidirectional rCBF change pattern: frontal hyper-perfusion and posterior hypoperfusion in AD matched the findings in our previous study (Hu et al., 2010).
Interestingly, the seemingly least sensitive 2D PASL showed significant rCBF versus memory correlations. 3D PASL did not find any significant CBF/rCBF versus memory correlations. 3D PCASL showed negative rCBF versus memory correlations. Because the ROIs are defined from the regions with hypo-metabolism or hypo-perfusion in AD, the positive rCBF versus memory correlations in these regions identified by 2D PASL means that lower CBF (relative to the whole brain mean) reflects the severity of clinical symptoms of the AD patients. The seemingly contradictory findings in 3D PCASL might be contributed by the relatively larger CBF variations (3D PCASL yielded higher CBF value than 2D PASL) and the smaller sample size. Regressing out age and sex may have disproportional effects on the CBF/rCBF versus memory associations. For example, we observed positive correlation between the meta-ROI and hippocampus ROI CBF and the LIMM score in 3D PCASL but not in the 2D PASL before regressing out age and sex. But we did not find any significant correlation after regressing out age and sex.
Several studies have compared different ASL sequences in healthy subjects using the same or different vendor machine (Mutsaerts et al., 2014Nanjappa et al., 2021;Steketee 1 3 et al., 2015;Vidorreta et al., 2012). The value of this study is the cross-sectional comparison of ASL MRI for normal aging, MCI, and AD, which are often very difficult to be recruited to a sequence comparison project using the within-subject design. Another new contribution is that the sequences are commercial products from two different vendors, while previous studies focused on laboratory-use ones.
We should note some limitations of this study. First, the sample size was modest. Particularly, there were fewer AD patients in the 3D PASL and 3D PCASL cohorts. The small sample size plus the multi-site data acquisition make it difficult to fully attribute the observed betweensequence discrepancy of the patient versus control or MCI versus AD CBF change patterns. Data harmonization through ComBat can not control other unknown factors such as environment or diet or life style, which can all contribute to CBF/rCBF changes (Joris et al., 2018;Stillman et al., 2021). These factors may also be part of the reason for the contradictory rCBF versus memory associations found in 2D PASL and 3D PCASL. Second, 3D PASL data did not contain an M0 scan for calculating the absolute CBF value. This technical issue should be solved in future data acquisitions so the 3D PASL and 3D PCASL sequences can be directly compared based on the quantitative CBF. A third limitation is that all ASL data were acquired with a single post-labeling delay (PLD) time, which might be shorter than the real arterial transit time (time for the labeled arterial blood to transit from the labeling plane or labeling location to the imaging site). However, as the PLD used in ADNI (1.9 s for 2D PASL and 2 s for 3D PASL and 3D PCASL) was longer than or similar to the arterial transit time (ATT) of NC, MCI, and AD reported in the literature (Liu et al., 2012;Sun et al., 2022;Tsujikawa et al., 2016), the possibility of having a shorter than expected PLD should be low for most of brain regions. For brain regions with an ATT > PLD, the observed group level CBF/rCBF difference may be partially contributed by the potential group level ATT difference. To fully delineate these two components, we will need to measure ATT using the multi-PLD ASL MRI or the Hadamard encoded ASL (Gunther, 2007;MacIntosh et al., 2010;Wells et al., 2010) or the slice timing based global ATT estimation method (Camargo & Wang, 2022). The fourth limitation is that the several key variables (age, sex, years of educations) were not matched in the MCI group for the three ASL data types. Sex was significantly different among the three ASL data types for NC, MCI, and AD. Although these variables were included as nuisances in the corresponding analyses, residual effects may still exist. In that sense, part of the observed across-ASL data type differences might be contributed by these variables.

Conclusion
In terms of group level CBF/rCBF changes among the different stages of the AD continuum, 3D PCASL (with BS) identified more widespread CBF difference than 3D PASL (with BS) and 2D PASL. CBF in orbito-frontal cortex presents a U-shaped change pattern from normal aging to MCI and to AD. The value of 3D PCASL for assessing the neurocognitive correlates of CBF/rCBF needs more data to be fully evaluated.
Acknowledgements Data collection and sharing for this project were funded by the Alzheimer's Disease Neuroimaging Initiative (ADNI) (National Institutes of Health Grant U01 AG024904) and DOD ADNI (Department of Defense award number W81XWH-12-2-0012). ADNI is funded by the National Institute on Aging, the National Institute of Biomedical Imaging and Bioengineering, and through generous contributions from the following: AbbVie, Alzheimer's Association; Author contributions AC downloaded the data, analyzed, and interpreted the data, and wrote the initial version of the manuscript. ZW designed the study, interpreted the results, wrote the manuscript.

Competing interests
The authors declare no competing interests.
Ethical approval Subject recruitment and data acquisition were approved by the Internal Review Boards for the parent ADNI project. All human subjects provided written consent forms before participating the ADNI study. Data re-analysis for this study was approved by Internal Review Board of University of Maryland Baltimore.