The identification and cognitive correlation of perfusion patterns measured with arterial spin labeling MRI in Alzheimer’s disease

Although biomarker development, in particular focusing on amyloid-β (Aβ) and paired helical filament tau, is improving the diagnostic efficacy of Alzheimer’s disease (AD), the invasive procedure of cerebrospinal fluid (CSF) testing and the limited availability of positron emission tomography (PET) impede their wide clinical application. Moreover, the current biomarker system for AD diagnosis is not absolutely specific or comprehensive. Patients with other conditions, such as dementia with Lewy bodies (DLB) and cerebral amyloid angiopathy, or even cognitively healthy individuals could also exhibit positive results for Aβ and tau on PET scans [1]. On the other hand, the pathophysiological process of AD is quite complex and potentially heterogeneous and may not be limited to Aβ and tau. The National Institute on Aging and the Alzheimer’s Association (NIA-AA) Research Framework indicated that other biomarkers may be incorporated in the AT(N) system in the future, such as TDP-43, α-synuclein, biomarkers of neuroinflammation, and vascular biomarkers in particular [2].

There is abundant evidence that brain vasculature plays an important role in the pathogenesis and progression of AD, even independent of Aβ and tau and vascular risk factors [3]. Moreover, an updated hypothetical model of AD biomarkers has suggested that vascular dysfunction, such as changes in the cerebral blood flow (CBF) and blood–brain barrier, may contribute to the initial stage of the pathophysiological process in AD, even before Aβ and tau pathology [4]. As an indicator of not only neurovascular uncoupling but also brain dysfunction, CBF assessment is a promising tool for early diagnosis and disease monitoring in patients with AD [5].

Arterial spin labeling (ASL) is a non-invasive MRI technique to quantitatively measure CBF using arterial blood water as an endogenous tracer, which is less time-consuming, and more feasible and repeatable compared with traditional perfusion measurements, such as PET and single photon emission computed tomography (SPECT) [6]. Alsop et al. first reported significant global and regional CBF reduction measured with continuous ASL in patients with AD [7]. While not entirely consistent, subsequent ASL studies have observed robustly decreased CBF in the cingulate, precuneus, parietal lobes, and inferior frontal regions [6].

Although previous studies suggested that ASL is a reliable measure of neurovascular dysfunction in AD, they employed voxel-by-voxel univariate analytical methods that ignored intrinsic functional correlations between anatomical structures of interest. Scaled subprofile model (SSM) is a multivariate analysis method based on principal component analysis (PCA) to identify significant spatial covariance patterns in functional brain images [8]. Indeed, this method has been used to establish disease-related metabolic patterns of AD [9‐11] and other neurodegenerative diseases on the basis of ¹⁸F-fluorodeoxyglucose (FDG) PET.

A preliminary study including a small sample with clinically diagnosed AD identified an AD-related perfusion pattern (ADRP) using continuous ASL, which featured negative loading mainly in temporal and parietal brain areas [12]. However, there have been few subsequent investigations in larger and independent cohorts, in particular in patients with biomarker confirmed AD. We previously identified an age-related perfusion pattern from pseudo-continuous ASL (pCASL) data using the method of SSM/PCA and demonstrated its individual evaluation according to subject expression score [13]. In this study, we established an ADRP with SSM/PCA in patients with AD whose diagnosis was supported by amyloid PET scan. We then tested and cross-validated the expression of this ADRP as an MRI perfusion biomarker for AD diagnosis and objective assessment of clinical severity with cognitive function in different domains. Finally, we compared the characteristics of the ADRP and its subject expression scores with regionally specific CBF values and cognitive correlations using complementary univariate analysis, which was performed independently from the SSM/PCA procedure.

This study established an AD-related CBF covariance pattern measured with ASL MRI using SSM/PCA. Consistent with the results of voxel-wise univariate analysis with normalization for global CBF value, the ADRP indicates a relatively negative loading in the bilateral middle and posterior cingulate and precuneus, bilateral inferior parietal lobule, and bilateral middle frontal gyrus, and a relatively positive loading in the bilateral putamen and a few frontal areas in patients with AD. Both subject expression scores of the ADRP and relative regional CBF value could efficiently distinguish patients with AD from healthy individuals. Moreover, compared with relative regional CBF, the ADRP subject expression scores more strongly correlated with global cognition and various cognitive domains, including memory, attention and information processing speed, executive function, language, and visuospatial ability.

To the best of our knowledge, this is the first study to establish an ADRP with SSM/PCA using pCASL data in patients with amyloid biomarker supported AD. This perfusion ADRP accounted for 25.64% of the variance in the CBF imaging data and included both negative and positive loadings. However, a previous ASL MRI study with a small sample only reported a covariance perfusion pattern characterized by negative loadings mostly around the parahippocampal gyrus in the medial temporal and occipital lobes, as well as the thalamus in clinically diagnosed patients with AD; but no areas with positive loadings were found using SSM/PCA [12]. Our ADRP is consistent with the results of multivariate analysis from an early study using the gold standard perfusion measure of H₂O¹⁵ PET, which observed decreased concomitant flow with SSM/PCA in the bilateral inferior parietal lobule and cingulate, left middle and inferior frontal, and precentral and supramarginal gyri, although they found increased concomitant flow in more extensive areas than we observed, including the bilateral insula, lingual gyri and cuneus, left fusiform and superior occipital gyri, and right parahippocampal gyrus and pulvinar [25].

The metabolic covariance pattern measured with FDG PET was also characterized by negative loading in the posterior temporo-parietal and prefrontal regions and relatively positive loading in the subcortical, cerebellum, and sensorimotor regions in patients with AD [11]. The regions with negative loading in our ASL established ADRP are highly consistent with those in the signature FDG pattern for patients with typical AD [26]. The consistency of findings from ASL MRI and FDG PET supports a coupling between perfusion and metabolism. Since FDG PET is considered as a biomarker of neurodegeneration or neuronal injury in the pathological process of AD, our results suggest that ASL could be an alternative to FDG PET in the diagnosis and evaluation of patients with AD, at least at the mild to moderate stage.

Moreover, our ability to discriminate between AD and NCs on the basis of the perfusion ADRP subject expression scores is consistent with previous metabolic and perfusion studies with PET. Metabolic pattern measured with FDG PET exhibited a sensitivity of 82.0% and a specificity of 94.0% for discriminating AD from NCs using subject expression score [27]. Another study [28] found that the expression of a previously established AD-related metabolic pattern was significantly different between stable subjects and AD converters at baseline in a mild cognitive impairment (MCI) cohort, with an area under the ROC curve of 0.80. Furthermore, the abovementioned H₂O¹⁵ PET study [25] showed that the expression of the perfusion covariance pattern had good diagnostic accuracy (sensitivity of 76% and specificity of 81%) in discriminating AD from NCs and could predict the decline of memory and cognitive performance in subjects with minimal to mild cognitive impairment. A recent study used PCA to identify a metabolic AD conversion-related pattern on FDG PET and found that the resulting pattern expression score was superior to clinical variables in predicting conversion from MCI to AD over a follow-up period of 5 years [29]. Our perfusion ADRP showed a modest sensitivity of 68.75% and a very high specificity of 96.88% for the identification cohort and good diagnostic performance (both sensitivity and specificity above 78%) for the validation cohort in distinguishing AD patients from the NCs. Of note, the mean ADRP score and optimal cutoff value were lower in the validation cohort because those AD patients had lower severity than those in the identification cohort based on the MMSE scores and scores of language and visuospatial function (Table 1), given the negative correlations (Fig. 6) between cognitive function and ADRP score in AD patients across both cohorts. Large samples with comparable disease severity and cross-validation would be needed to establish the reproducibility of ADRP topographies and its expression for disease discrimination.

Consistent with previous studies [6], we demonstrated robust global hypoperfusion, and extensive regional CBF reduction and corresponding cognitive correlations using univariate analysis in patients with AD at the mild to moderate stage, in particular the middle and posterior cingulate, precuneus, and temporo-parietal association cortex; and relative CBF value in these regions, in particular the left posterior cingulate, showed high sensitivity and high specificity in distinguishing AD patients from NCs. It is well-established that vascular dysfunction, including changes in the blood–brain barrier integrity and CBF, is a prominent and early feature in AD pathophysiology. Cerebral hypoperfusion could give rise to reduction in Aβ clearance and subsequent accumulation of amyloid plaques and neurofibrillary tangles; in turn, the core AD pathology might exacerbate vascular injury and CBF decline [30]. Moreover, measures of brain demyelination, which have been recently demonstrated to be important biomarkers for the pathology of AD and MCI [31], are also associated with CBF reduction. Since myelin homeostasis based on oligodendrocyte metabolism is an energy-consuming process, it is particularly sensitive to hypoxia, hypoperfusion, or ischemia [32].

In addition, we also found relatively increased regional CBF in a few brain areas of AD patients after adjusting for global value, such as frontal supplementary motor area, primary motor and sensory cortex, and subcortical deep GM. These areas are usually preserved in brain structure and function at the early stage of the neurodegenerative process in AD according to previous studies [33]. Interestingly, increased CBF was observed in various brain regions in patients with AD across different studies with or without normalizing CBF by a reference value such as global perfusion, including the insular cortex, temporal cortex, frontal cortex, anterior cingulate, hippocampus, amygdala, and basal ganglia [34‐36], which was usually presumed to be a compensatory mechanism, especially at the MCI or mild dementia stage.

We noted the topographic similarity between regional loadings provided by SSM/PCA and perfusion differences revealed by SPM (cf. Figures 1 and 3). This is expected given that subject expression score of the ADRP is simply a summed product between the ADRP topography and an individual CBF image such that the greater the similarity between the ADRP and the CBF map, the higher the subject expression score in each AD patient [24]. Although relative CBF value in various regions had a high accuracy in distinguishing AD patients from NCs, subject expression score of the ADRP showed stronger and more extensive correlations with clinical variables, such as cognitive performance, in AD patients in this study. Of note, relative CBF values in this study were measured in spherical volumes of interest (4 mm radius) centered at the peak voxel of each significant cluster where group differences were maximal. The group differences would be smaller in conventional analytical methods owing to signal dilution from lower-intensity voxels when CBF values were measured over the entire cluster detected by SPM or the neuroanatomic structure defined on MRI. Because computation of pattern expression score is performed automatically without clinical information, this approach is more objective than diagnostic categorization achieved by visual interpretation or predefined region of interest (ROI) analysis [24]. Network analysis generally recovers more disease-specific, widely distributed brain regions that may not have direct biological correlates, whereas regional analysis is closer to the diagnostic process based on visual reading of clinical images [37]. According to our present results, it might be beneficial to combine the imaging markers obtained with these two complementary approaches to further improve the diagnosis and evaluation for AD.

All data used in this study were obtained from our longitudinal MRI research, in which the diagnosis of AD is supported by the core pathological biomarker of amyloid PET, and comprehensive neuropsychological testing and multimodal MRI were conducted for all participants. However, this study still has several limitations. First, like the majority of previous ASL studies, we reported CBF changes without correcting for partial volume effects (PVEs) given previous studies suggesting that correction for PVEs may not be necessary to improve the radiologic differentiation between patients with AD and subjects with subjective complaints [38], or to monitor disease severity in AD patients [39]. To examine any effect of brain atrophy on perfusion ADRP or CBF, we repeated the same SSM/PCA and SPM analysis with structural data in the identification cohort. Structural covariance pattern (ADRP-GM) was identified from a linear combination of the first 3 PC accounting for 14% of the total voxel × subject variance (Supplementary Fig. 2). We found that ADRP perfusion had small overlaps with a few atrophic regions in the ADRP-GM but only at the lowest level of reliability (P = 0.05) (Supplementary Fig. 3). Although there was some overlap between t-maps from CBF and GM tissue images without adjusting global value (Supplementary Fig. 4), it became minimal after adjusting for global difference in perfusion and GM tissue volume by ANCOVA even at the most liberal threshold of P = 0.05 (Supplementary Fig. 5). Hence, AD-related perfusion patterns in our study are truly vascular/perfusion-driven rather than a proxy measure for atrophy. Second, both SSM/PCA and SPM analyses were conducted within the same brain mask created from the high-resolution probabilistic GM map in SPM with a liberal threshold of pGM ≥ 0.3. To rule out any potential issues, we repeated SPM analysis in the identification cohort using the brain mask at a moderate threshold (pGM ≥ 0.5). The results showed the same regions of relative hypo- and hyper-perfusion as described above (see the comparison in Supplementary Fig. 6), indicating that the use of pGM ≥ 0.3 was adequate in this study. Third, repeated analyses in increasingly larger subsamples should be performed in a future study of SSM/PCA in conjunction with SPM analysis to determine the smallest sample sizes necessary for robust spatial covariance analysis. Fourth, although participants who had obvious large vessel vascular disease measured by carotid duplex ultrasound, TCD, or cranial MR angiography were excluded, the CBF maps might not reflect the effects of pure functional/microvascular physiology in the brain when blood flowing towards the brain is slowed (e.g., prolonged arterial transit time (ATT)). Nonetheless, the regional differences in CBF persisted in our AD sample after global normalization, suggesting that these findings were not entirely vascular. Recognizing the heterogeneous distributions of ATT across the brain, advanced protocols that include multiple post-labeling delays to account for effects of spatial variation in ATT [40] could be considered in future investigations. Finally, other factors that might influence cerebral perfusion, such as apolipoprotein E (APOE) genotype and WMH, were not taken into account in our data analysis, although we excluded obvious cerebral large and small vessel disease by multimodal MRI and duplex ultrasound.

While there was a significant correlation between subject expression score of the pattern and performance in global cognition and various cognitive domains, we could not prove whether this ADRP predicts disease progression because of cross-sectional analysis in the current study. In addition, unique brain networks associated with different cognitive disabilities could be respectively identified by multivariate analysis, with stronger correlations than those between this ADRP and all cognitive domains in the present study. For instance, several metabolic PCs measured with FDG PET were previously established in AD patients using SSM/PCA [10], namely, PC1 (posterior cortices) correlated with naming and visuospatial abilities, PC2 (limbic structures from the Papez circuit, e.g., medial temporal regions, posterior and anterior cingulate cortex, thalamus) correlated with episodic memory, and PC3 (frontal, parietal, temporal, and posterior medial association cortices) correlated with executive and global cognitive functions.

An AD-related perfusion covariance pattern featuring negative loading mainly in the bilateral temporo-parietal cortex, cingulate, and precuneus and relatively positive loading marginally in the subcortical deep GM and frontal areas has been identified by multivariate analysis based on ASL data, and its expression is characterized in relation to cognitive impairment in various domains. The diagnostic utility of this ADRP was replicated in an independent validation cohort. This ADRP may provide a promising MRI biomarker for AD diagnosis and monitoring on a prospective single-subject basis.