Perivascular Spaces as Determinants of Amyloid, Tau, and Vascular Biomarker Progression

Abstract

Objective

Magnetic resonance imaging (MRI)‐visible enlarged perivascular spaces (PVS) are markers of cerebral small vessel disease (SVD) and aging, processes implicated in both neurodegenerative and cerebrovascular pathologies. However, longitudinal positron emission tomography (PET) studies examining PVS as a mechanism underlying Alzheimer's disease (AD) pathophysiological progression are lacking. Our overall objective was to investigate the associations of baseline PVS burden with white matter hyperintensity (WMH) volume, amyloid‐PET, and tau‐PET, both cross‐sectionally and longitudinally.

Methods

We examined 2,229 participants across the aging‐AD spectrum from the Mayo Clinic Study of Aging (MCSA) and the Mayo Alzheimer's Disease Research Center (ADRC) (mean age = 68.59 ± 12.70 and 48.94% women). PVS and WMH were quantified using a novel in‐house algorithm. Amyloid on [¹¹C]Pittsburgh Compound B (PiB)‐PET and tau on AV1451‐PET were measured. Cross‐sectional regional and global associations were examined using canonical correlation analysis (CCA). Longitudinal associations were examined using age‐ and sex‐adjusted linear mixed effect (LME) models with natural splines and subject‐wise 10‐fold cross‐validation.

Result

Three distinct associations were observed: (1) PVS in the basal ganglia (PVS‐BG) was associated with greater WMH volume both cross‐sectionally and longitudinally; (2) PVS‐BG covaried with higher PiB standardized uptake value ratio (SUVR) cross‐sectionally, with longitudinal trajectory associations limited to amyloid‐negative individuals; (3) PVS in the centrum semiovale (PVS‐CSO) was not associated with amyloid or tau, but was associated with occipital WMH, a pattern characteristic of CAA.

Interpretation

Findings lend support for PVS‐BG as an early indicator of small vessel dysfunction and WMH pathogenesis. Individuals with higher PVS‐BG burden exhibited higher WMH burden and faster WMH progression, suggesting that PVS‐BG may identify individuals at increased risk of progressive SVD. Evidence linking PVS to downstream AD biomarker progression was weaker. Findings support the relevance of PVS‐CSO to CAA, suggesting that they may reflect changes occurring early in the disease process. ANN NEUROL 2026;100:695–708

Alzheimer's disease (AD) is increasingly understood not only as a disease of abnormal protein aggregation, but also as a condition intertwined with vascular dysfunction. Cerebral small vessel disease (SVD) is now recognized as a major contributor to cognitive impairment and dementia. Among the key imaging markers of SVD, enlarged perivascular spaces (PVS) have been poorly studied, although they are gaining increasing attention due to their role in brain waste clearance of toxic material like amyloid, a key AD pathology. ¹ , ²

PVS are fluid‐filled spaces surrounding the small penetrating vessels of the brain. Whereas typically microscopic and invisible to the eye on magnetic resonance imaging (MRI), PVS become visible when dilated. Although their pathophysiology remains unclear, it has been proposed that the dilation of PVS may occur when perivascular clearance is impaired, resulting in the accumulation of fluid within these spaces, or conversely, that PVS dilation itself may disrupt perivascular fluid dynamics and clearance. ³ , ⁴ The appearance of visible PVS on MRI represents an important marker of SVD (Fig 1), ⁵ , ⁶ , ⁷ , ⁸ with increasing evidence implicating PVS in altered cerebral hemodynamics and interstitial fluid drainage. ⁹ , ¹⁰ , ¹¹

FIGURE 1.

The PVS. (A) Graphical illustration of PVS anatomy, created with BioRender.com. (B) PVS as visualized on structural T2‐weighted MRI scan. (C) Quantification of PVS using multispectral image segmentation with outline of the 2 regions were PVS burden is observed–the centrum semiovale in blue (PVS‐CSO), and the basal ganglia in red (PVS‐BG). MRI = magnetic resonance imaging; PVS = Perivascular spaces; PVS‐BG = PVS in the in the basal ganglia; PVS‐CSO = PVS in the centrum semiovale. [Color figure can be viewed at www.annalsofneurology.org]

PVS burden has been linked to elevated risk of developing dementia, independent of vascular risk factors and other markers of SVD. ¹² Importantly, PVS can be detected on MRI earlier than other SVD markers, ¹³ , ¹⁴ and genome‐wide association studies have implicated PVS in early mechanisms of cerebrovascular dysfunction. ¹⁵ Together, these findings highlight the potential of PVS as an early biomarker and their potential involvement in exacerbating AD pathologies.

A growing body of evidence suggests that cerebrovascular dysfunction may accelerate the accumulation of hallmark AD pathologies like amyloid plaques and neurofibrillary tau tangles, with impaired perivascular and glymphatic clearance acting as key mechanistic pathways connecting AD and SVD. Glymphatic dysfunction has been associated with increased amyloid and tau pathology, and greater white matter injury, in both preclinical and symptomatic AD, ¹⁰ , ¹⁶ highlighting PVS as a potential upstream driver and early predictor of neurodegeneration.

Despite their potential clinical relevance, PVS remains an understudied phenomenon in humans. Prior studies have largely relied on visual ratings of PVS burden on selected MRI slices, which are subjective, less precise, and lack detailed morphological information. More recently, quantitative PVS measures have emerged, including automated segmentation tools that enable detection and quantification of PVS on brain MRI, providing estimates of both global and regional PVS burden. Accordingly, the use of quantitative assessments of lesion burden and their regional distribution has been highlighted as a research priority for the field. ¹⁷

Additionally, the extent to which PVS burden predicts the longitudinal progression of key pathologies and cognitive decline remains an understudied area. Longitudinal studies examining baseline PVS in relation to subsequent pathological changes are crucial for clarifying the roles of vascular and glymphatic mechanisms in the development of AD. Such research may also help to identify early targets for preventative strategies and biomarker‐driven interventions. Therefore, we aimed to examine participants from 2 well‐characterized longitudinal cohorts, a community‐based cohort, and a clinical cohort recruited from memory clinics. Leveraging multi‐modal MRI and positron emission tomography (PET) imaging, we investigated associations between PVS burden and future biomarker trajectories of amyloid, tau, and white matter hyperintensity (WMH) processes, with the aim of understanding underlying mechanisms reflected by PVS.

Methods

Participants

Participants were drawn from the Mayo Clinic Study of Aging (MCSA) and the Mayo Alzheimer's Disease Research Center (ADRC). ¹⁸ , ¹⁹ The MCSA is a population‐based cohort in Olmsted County, Minnesota, designed to investigate the prevalence, incidence, risk factors, and biomarkers of MCI and dementia. ¹⁸ The Mayo Clinic ADRC is a longitudinal cohort comprising participants referred to the Mayo Clinic. Together, these cohorts include individuals across the lifespan (ages 30–100), encompassing cognitively unimpaired individuals as well as those with mild cognitive impairment (MCI) and dementia, offering a unique opportunity to study the vascular contributions to neurodegeneration across the spectrum of cognitive aging. Both the MCSA and ADRC protocols have been approved by the Institutional Review Boards of Mayo Clinic and Olmsted Medical Center. Written informed consent was obtained from all participants. Participants were included if they had at least one timepoint of PVS data, and at least one measurement of WMH volume, amyloid‐PET, or tau‐PET. Individuals were excluded if they were receiving anti‐amyloid treatment, had a diagnosis of Lewy body dementia, progressive supranuclear palsy, frontotemporal dementia, or presented with a stroke or any anatomic abnormalities on structural MRI that significantly impacted imaging segmentation (eg, encephalomalacia).

Image Acquisition

All images were acquired on 3T Siemens Prisma scanners using a 64‐channel head coil, with identical protocols between the MCSA and ADRC cohort. Structural MRI acquisitions spanned 2 protocol epochs. The earlier epoch, through early 2020, included a 3D T1‐weighted magnetization‐prepared rapid gradient echo (MPRAGE) images (0.8 mm isotropic; TR/TE/TI ≈ 2,300/3.14/945 ms, flip angle = 9 degrees), a 3D T2‐weighted sequence (0.8 mm isotropic; 3D T2‐SPACE, TR/TE ≈ 3,200/409 ms), and a sagittal 3D fluid‐attenuated inversion recovery (FLAIR) with slightly anisotropic resolution (≈1.0 × 1.0 × 1.2 mm; TR/TE/TI ≈ 4,800/441/1,650 ms). The later epoch followed the ADNI‐4 structural MRI protocol which specifies 1.0 mm isotropic 3D acquisitions for all core structural contrasts: a 3D T1‐weighted MPRAGE (TR/TE/TI ≈ 2,300/2.98/900 ms, flip angle = 9 degrees), a sagittal 3D FLAIR (TR/TE/TI ≈ 4,800/441/1,550 ms), and a 3D T2‐weighted sequence (SPACE; TR/TE ≈ 3,200/409 ms).

Image Processing

A 3‐step approach was taken in the image processing, as previously described. ²⁰ First, images were affine registered and resampled into the T1‐weighted image coordinate system (subject space). The SPM12 unified segmentation algorithm is used as follows: (1) segmentation is performed with the traditional 6 classes (grey matter [GM], white matter [WM], cerebrospinal fluid [CSF], skull, scalp, and air); (2) the contrast model is expanded to include deep GM, a subsequent parameter optimization is performed including prior probability maps with a deep GM channel; (3) the contrast model is expanded to include WMH and a final parameter optimization using further expanded prior maps that include a WMH channel. At this point, estimates of the contrast model parameters, bias‐correction fields, and deformation field between subject and Mayo Clinic Adult Lifespan Template (MCALT; https://www.nitrc.org/projects/mcalt/) space have been estimated and class‐wise probability maps generated. A lobar atlas map from the MCALT was applied to extract PVS and WMH volume separately in the frontal, parietal, occipital, and temporal lobes.

PVS maps were then built on the segmentation model output. A bank of 3D Frangi filters were passed over bias‐corrected T1‐weighted and T2‐weighted images to highlight vessel‐like structures. ²¹ , ²² The filter outputs were summed, smoothed, and clipped between 0 and 1. A mask containing a rough PVS search region was warped from the template space and cleaned up based on segmentation output to avoid areas immediately adjacent to the ventricles and cortical GM. The resulting mask was then applied to the Frangi filter output to create a PVS prior. The full set of probability maps including the WMH channel were resampled into subject space using the deformation field. The CSF prior channel was then replaced with the PVS prior and the resulting set of priors were normalized to unity on a per‐voxel basis. A call to the internal SPM unified segmentation likelihood function was made using the new subject space priors and the previously determined contrast parameters. The resulting likelihood maps were converted to probability estimates using the standard SPM ratio approach. The resulting PVS channel highlights voxels with contrasts similar to CSF and a vessel‐like presentation in the PVS search region. Notably, the inclusion of the FLAIR image and WMH probability class serves to differentiate PVS from WMH. Binary PVS maps are determined by thresholding the PVS probability maps. Clustering is performed and estimates of lengths and volumes were generated. Given the etiological and anatomic differences between PVS locations, we focused on 2 regions where PVS are commonly observed: PVS in basal ganglia (PVS‐BG), which follow the course of deep perforating arterioles, and PVS in centrum semiovale (PVS‐CSO), which follow long cortical penetrating arterioles.

PET Imaging

Amyloid PET imaging was performed using [¹¹C]Pittsburgh Compound B (PiB) PET imaging, ²³ and tau PET was acquired using [¹⁸F]AV1451 flortaucipir PET, synthesized on site (Avid Radiopharmaceuticals), ²⁴ as previously detailed. ¹⁹ Late‐uptake amyloid PET images were acquired from 40 to 60 minutes and tau PET from 80 to 100 minutes post‐injection. Computed tomography scans were obtained for attenuation correction. Amyloid and tau PET standardized uptake value ratios (SUVRs) were derived by normalizing composite multi‐region target regions of interest (ROIs) to the cerebellar crus gray median using an in‐house pipeline implemented in SPM12. Amyloid PET values are expressed in both SUVR units and Centiloid units. The conversion from SUVR to Centiloid was performed as recommended in Klunk et al. ²⁵

Statistical Analysis

Standard statistical techniques were used for descriptive analyses. Parametric data were analyzed with either independent t tests or analysis of variance (ANOVA). Nonparametric data were analyzed with the Mann–Whitney U test. Ordinal data were analyzed using the Kruskal‐Wallis test. Group comparisons of categorical variables were analyzed using χ ² tests of independence. Variables were log‐transformed for normality. In all models, the primary predictor was baseline PVS volume, and outcome variables were WMH volume, amyloid PiB‐PET, and tau‐PET for imaging, as well as cognition. Separate models were fitted for each outcome variable, with age, sex, and time from baseline as covariates.

First, we examined cross‐sectional associations between PVS and imaging variables of interest (WMH, amyloid‐PET, and tau‐PET) to disentangle the mechanisms through which PVS relates to AD and cardiovascular disease (CVD) processes. We performed canonical correlation analysis (CCA) using a 2‐level analytical approach. ²⁶ , ²⁷ , ²⁸ , ²⁹ CCA evaluated multivariate associations between sets of PVS and outcome variables, identifying regional associations by extracting maximally correlated linear combinations of variables. In the first‐level analysis, permutation‐based CCA was conducted to assess the relationship between the 2 constructs, each represented as a separate multivariate dataset. The first dataset was comprised of PVS ROIs, whereas the second dataset was comprised of ROIs of the respective outcome variable (ie, WMH, amyloid, or tau), and these were analyzed in parallel CCA models. Missing data were imputed using default settings in the multivariate imputation by chained equations (mice) package in R, and all variables were standardized into z‐scores prior to CCA. Significance of the CCA model was tested via permutation testing with 5,000 iterations. To assess the stability of the canonical correlations, we additionally performed 10‐fold cross‐validation. The dataset was randomly partitioned into 10‐folds of approximately equal size. For each fold, the CCA model was trained on 90% of the data and tested on the remaining 10%. To test whether the cross‐validation folds differed systematically in sample characteristics, we compared fold assignments by age, sex, and study cohort. The cross‐validation splits did not differ significantly by age (ANOVA: F = 0.06, p = 0.808), sex (χ ² = 7.01, p = 0.636), or study cohort (χ ² = 7.72, p = 0.563). We report canonical variates that were significant at p < 0.05 (2‐tailed) with shared variance above 5% (Rc² > 0.05). In the second‐level analysis, we extracted subject‐specific scores for each pair of canonical components to examine whether associations between the constructs remained significant in a simple linear regression analysis adjusting for sex and age. Canonical weights were examined to determine the contribution of different variables within the canonical variate.

Longitudinal analysis focused on key associations emerging from the cross‐sectional analysis. Specifically, we examined the relationships between baseline PVS‐BG burden and longitudinal progression of key imaging biomarkers of AD (amyloid and tau) and SVD (WMH) using linear mixed effects models fitted with the lme4 package in R software. Separate models were fitted for each outcome variable, with independent variables being baseline PVS burden (predictor), baseline age, sex, and time between visits (fixed effects and covariates). Time was modeled as years because baseline PVS measurement using natural splines (df = 2) to allow for nonlinear trajectories. Each model included an interaction between time and baseline PVS to assess whether PVS at baseline modified longitudinal trajectories of imaging outcome variables (amyloid, tau, and WMH). Mixed effects models included the random intercept and slope for each participant to account for individual variations in baseline pathology and rates of change over time. Results are reported based on the omnibus test of the spline*PVS interaction terms using likelihood‐ratio tests (LRTs) comparing nested models. Grouped 10‐fold cross‐validation was conducted, with folds at the participant level to ensure that repeated measures from a single subject do not leak into both training and test sets. To assess internal consistency of model estimations, we assessed fold‐to‐fold variability in root mean squared error (RMSE) based on original outcome scales.

All statistical analyses were performed using R software version 4.4.0. (https://www.r-project.org/). To reduce collinearity, all independent variables in mixed effects models and general linear models were mean‐centered. All p values were 2‐sided, and p < 0.05 were considered statistically significant.

Results

Participant Characteristics

Sample characteristics are summarized in Table 1. Of 2,229 participants, 48.94% were women. The mean age was 68.59 years (SD = 12.70), and participants had an average of 15.45 years of education (SD = 2.50).

TABLE 1.

Sample Characteristics

	n = 2,229
Demographics
Sex (N, % F)	1,091 (48.94)
Age, in yr	68.59 ± 12.70
Education, in yr	15.45 ± 2.50
APOE4	564 (32.62%)
Cohort
MCSA	1,868 (83.80%)
ADRC	361 (16.19%)
Cognition
MMSE	28.01 ± 2.74
AVLT	59.90 ± 19.72
Imaging
PVS volume, % TIV	0.36 ± 0.20
WMH volume, % TIV	0.64 ± 0.75
Amyloid PiB, Centiloids	35.49 ± 41.57
Tau SUVR	1.25 ± 0.24

Of the full sample, 2,118 had baseline WMH measurements, 2,110 had baseline PiB measurements, and 2099 had baseline tau measurements. Eight hundred fifty had repeated WMH data (mean 3.43 years of follow‐up), 819 had repeated PiB measurements (mean 3.55 years of follow‐up), and 812 had repeated tau measurements (mean 3.54 years of follow‐up). Five hundred participants were missing APOE4 data.

ADRC = Alzheimer's Disease Research Center; APOE = apolipoprotein E; AVLT = Auditory Verbal Learning Test; MCSA = Mayo Clinic Study of Aging; MMSE = mini mental state examination; PiB = [¹¹C]Pittsburgh Compound B; PVS = perivascular spaces; SUVR = Standardized Uptake Value Ratio; TIV = total intracranial volume; WMH = white matter hyperintensities.

Age‐Related PVS Increase Differs by Region

Older age related to an increase in PVS‐BG at a rate of 90.24 mm³ per decade (95% confidence interval [CI] = 83.33, 97.18, p < 0.001) and PVS‐CSO at a rate of 528.58 mm³ (95% CI = 449.09, 608.07, p < 0.001) per decade. This corresponded with an increase in PVS of 22.22% per decade (95% CI = 20.63, 23.82) in the BG, and 14.51% per decade (95% CI = 12.50, 16.55) in the CSO (Fig 2). PVS was similarly related to age across all 4 lobar regions, with the greatest age‐related increase observed in the frontal lobe (396.22 mm³ [350.00, 442.43], or 20.15% [18.04, 22.30] increase per decade, p < 0.001). These patterns were consistent in subgroup analyses of the population‐based MCSA cohort alone (ie, excluding ADRC participants; Supplementary Fig S1).

FIGURE 2.

Density plot of the distribution of baseline PVS volume as a function of age. PVS volumes are represented as a percentage of total intracranial volume. Colors represent the number of participants within each bin, where warmer colors indicate higher density, that is, higher numbers of individuals in the cell. Dashed lines represent GAM fits for each PVS region, with shaded areas indicating 95% CIs. The CSO comprises the frontal, parietal, and occipital lobe. CIs = confidence intervals; CSO = centrum semiovale; GAM = generalized additive model; PVS = perivascular spaces. [Color figure can be viewed at www.annalsofneurology.org]

Cross‐Sectional Analysis

The CCA results are summarized in Table 2 and Figure 3.

TABLE 2.

Results of Canonical Correlation Analysis

	WMH		Amyloid‐PET	Tau‐PET
	CV1	CV2	CV1	CV1
Canonical loadings
PVS
Basal ganglia	0.82	0.07	0.72	0.77
Frontal	0.28	0.63	0.48	0.71
Parietal	−0.02	0.77	0.14	0.25
Temporal	−0.04	0.57	0.10	0.28
Occipital	0.20	0.88	0.41	0.31
Outcome, WMH/Amyloid/Tau
Basal ganglia	0.72	−0.36	0.66	0.38
Frontal	0.92	−0.13	0.82	−0.06
Parietal	0.78	0.15	0.84	−0.02
Temporal	0.60	0.08	0.80	0.27
Occipital	0.68	0.63	0.90	0.13
Significance testing of canonical functions
Wilks’ Lambda, λ	0.71	0.90	0.83	0.87
F	26.14	12.02	15.62	11.71
Canonical correlation, Rc	0.46	0.28	0.37	0.30
Shared variance, Rc2	20.87	7.73	14.01	9.09

Each column represents separate canonical variates that emerged significant in canonical correlation analysis, that is, p < 0.05, and Rc² > 5%.

CV = canonical variate; PET = positron emission tomography; PVS = perivascular spaces; WMH = white matter hyperintensities.

FIGURE 3.

CCA between PVS burden and imaging outcomes. Each row represents a significant canonical variate; PVS‐WMH analysis revealed two significant canonical variates, while PVS‐amyloid and PVS‐tau produced one significant canonical variate each. Left: Heliographs of canonical variate loadings. Each heliograph displays the canonical loadings (structural coefficient) for the canonical variate pair. Canonical correlation analysis identifies combinations of PVS regions and imaging outcome regions (eg, WMH regions) that covary most strongly with one another. Each bar represents the strength of correlation between an individual variable and its respective canonical variate, such that longer bars indicate stronger contribution to the shared pattern of associations. Blue bars extending outward represent positive weights (ie, greater pathology); red bars extending inward represent negative weights (ie, lower pathology); uncolored bars indicate |r| < 0.30. Half‐maximum strength of correlation is indicated by the innermost (r = −0.50) and outermost (r = 0.50) circles. Right: Bivariate correlation plots of canonical correlations (residuals adjusted for sex and age). BG = basal ganglia; CCA = canonical correlation analysis; CV = canonical variate; Fron = frontal lobe; Occ = occipital lobe; Par = parietal lobe; PVS = perivascular spaces; Temp = temporal lobe; WMH, white matter hyperintensity. [Color figure can be viewed at www.annalsofneurology.org]

WMH Globally Related to PVS‐BG, Whereas Higher Occipital WMH Related to PVS‐CSO

For PVS‐WMH associations, CCA identified 2 significant canonical variates (CVs; CV1: Wilks’ λ = 0.71, F = 26.14, p < 0.001 and CV2: λ = 0.90, F = 12.02, p < 0.001) with shared variances above 5%. The first CV (r = 0.46, p < 0.001) explained 20.87% of the variance between PVS and WMH, whereas the second CV (r = 0.28, p < 0.001) explained 7.73% of the variance (see Table 2). In 10‐fold cross‐validation, the mean canonical correlation for CV1 was 0.45 (SD = 0.06), and 0.26 (SD = 0.09) for CV2. Results were consistent in permutation testing with 5,000 bootstrap samples, as well as adjusted analysis using simple linear regression analysis of composite variate scores controlling for sex and age (CV1: β = 0.27, t = 12.32, p < 0.001 and CV2: β = 0.28, t = 12.69, p < 0.001).

Regionally, canonical variate loadings of CV1 identified a positive association between PVS‐BG (canonical loading/structure coefficient, r = 0.82) and WMH across the brain (r _range = 0.60 to 0.92), particularly WMH in the frontal (r = 0.92) and parietal (r = 0.78) lobes (see Table 2 and Fig 3). CV2 revealed a second regional pattern driven by PVS in the occipital (r = 0.88) and parietal (r = 0.77) regions of the CSO, which corresponded with a positive association with occipital WMH (r = 0.63).

Amyloid PiB‐PET Across the Brain Associated With PVS‐BG

For PVS‐amyloid associations, one significant canonical variate was identified (λ = 0.83, F = 15.62, p < 0.001), which captured a shared variance of 14.01%. The canonical correlation was r = 0.37 in the full sample, and a mean of 0.36 (SD = 0.06) in 10‐fold cross‐validation. The association between PVS and amyloid was observed similarly in permutation testing (p < 0.001) and adjusted models accounting for sex and age (β = 0.17, t = 8.26, p < 0.001). This canonical variate was driven by PVS‐BG (r = 0.72), which was positively associated with greater amyloid PiB‐PET signal in all regions, particularly in the occipital lobe (r = 0.90; see Table 2 and Fig 3).

Tau‐PET Signal in Basal Ganglia Positively Associated with PVS‐BG

For PVS and tau, one canonical variate emerged significant (λ = 0.87, F = 11.71, p < 0.001) with a shared variance of 9.09%. The canonical correlation was 0.30 in the full sample and 0.29 (SD = 0.03) in 10‐fold cross‐validation. Results were replicated in permutation testing (p < 0.001) but were not significant after adjusting for sex and age in separate simple regression models (β = 0.03, t = 1.38, p = 0.168). This canonical variate was driven by PVS in the basal ganglia (r = 0.77) and frontal lobe (r = 0.71) and higher AV1451 PET binding in the basal ganglia (r = 0.38; see Table 2 and Fig 3) which is typically regarded as off‐target binding, most often to inflammatory binding sites like monoamine oxidase (MAO) B.

Longitudinal Analysis

Given the significance of PVS‐BG over PVS‐CSO, and the lack of lobar‐specific effects in cross‐sectional analysis, longitudinal analyses focused on PVS‐BG burden.

Baseline PVS‐BG Associated With Differential Longitudinal Trajectories of WMH

Higher baseline PVS‐BG was significantly associated with faster nonlinear progression of WMH (LRT χ ² (2) = 13.60, p = 0.001) and amyloid (LRT χ ² (2) = 8.67, p = 0.013), but not longitudinal tau accumulation (LRT χ ² (2) = 0.39, p = 0.824; see Table 3 and Fig 4). Ten‐fold cross‐validation demonstrated stability across all models, as reflected by small fold to fold variability in RMSE, which is reported on the original outcome scale (see Table 3). This supports internal consistency of model estimation. To facilitate clinical interpretation, model‐predicted WMH trajectories showed that individuals in the high PVS group had 30.19% more WMH at baseline and 33.78% more WMH after 4 years, compared to the low PVS group.

TABLE 3.

Longitudinal Analysis of Baseline PVS‐BG on Progression of WMH, Amyloid, and Tau

	LRT		10‐fold cross‐validation
	LRT χ 2 (2)	p	RMSE mean	RMSE SD
WMH	13.60	0.001**	0.279	0.046
Amyloid PET	8.67	0.013*	0.162	0.017
Tau PET	0.39	0.824	0.105	0.012

LRT comparing the full model (spline of time in years × baseline PVS‐BG interaction) versus the reduced model (no interaction). RMSE is the out‐of‐sample root mean squared error from 10‐fold participant‐level cross‐validation adjusted for the fixed effects of baseline age and sex, accounting for random intercept for participants, modelled using a natural spline with 2 degrees of freedom.

^* p < 0.01.

^** p < 0.001.

LRT, likelihood ratio test; PET = positron emission tomography; PVS‐BG = perivascular spaces in the basal ganglia; RMSE, root mean squared error; WMH, white matter hyperintensities.

FIGURE 4.

Baseline PVS‐BG burden on the predicted trajectories of WMH, amyloid, and tau. Predicted trajectories were modeled using mixed effects regressions with natural spline terms for time and interactions with baseline PVS burden. Models were adjusted for baseline age and sex and included random slopes and intercepts for each participant. Outcome variables were log transformed and standardized using z‐transformation. Blue lines represent participants in the lowest tertile of PVS burden, green lines represent the middle tertile, and red lines represent the highest tertile of PVS burden. Tri‐panel plots at the base of each main panel display smoothed individual trajectories, faceted by baseline PVS‐BG tertiles (low, mid, and high) to illustrate within‐group variability and individual‐level trends. PVS‐BG = perivascular spaces in the basal ganglia; WMH = white matter hyperintensity. [Color figure can be viewed at www.annalsofneurology.org]

PVS‐BG as Predictor of Disease Progression Amplified in Amyloid‐Negative Individuals

To determine the pathways through which PVS may be involved in downstream disease processes, we analyzed A+ and A− individuals separately and found that PVS‐BG burden predicted accelerated progression of WMH and amyloid in A− individuals, but not in A+ individuals (Supplementary Table S1).

Discussion

This study examined the cross‐sectional and longitudinal associations between baseline PVS burden and imaging markers of Alzheimer's pathology in a sample spanning the AD spectrum. Using multivariate approaches, we observed that PVS‐BG was most consistently associated with WMH progression, supporting its role as an early marker of vascular pathology. Whereas PVS‐BG showed cross‐sectional associations with amyloid PiB PET and off‐target AV1451 binding in the basal ganglia, PVS‐BG did not predict longitudinal accumulation of amyloid or tau on PET imaging. PVS‐CSO was related to occipital WMH.

Age‐Related PVS Increase Was Most Pronounced in Basal Ganglia and Frontal Lobe

Advancing age was associated with increasing PVS burden across the brain, with notable regional differences in rates of change. The most pronounced age‐related increase was observed in the BG, with a 22% increase in volume per decade. This aligns with our understanding of the BG as being particularly susceptible to age‐related cerebrovascular alterations like hypertension and arterial stiffening, owing to its microvascular architecture and reliance on small penetrating end‐arteries with limited collateral supply. ³⁰

PVS‐CSO increased at a lower rate of 15% per decade, despite a larger absolute increase in volume due to the larger size of the CSO compared to the BG. In contrast to PVS‐BG, mounting evidence suggests that PVS‐CSO relates to WM changes or amyloid pathology, ³¹ , ³² and less directly linked to vascular aging. For instance, in healthy older adults, PVS‐BG, but not PVS‐CSO, relate to lower cerebrovascular reactivity ³³ and deep medullary vein changes. ³⁴

Age‐related increases in PVS burden were also observed across all lobar regions, with the frontal lobe exhibiting the steepest rate of change, increasing by 20% per decade. The greater burden of PVS in the frontal lobe compared with posterior regions aligns with well‐documented vulnerability of frontal brain regions to SVD and age‐related cerebrovascular alterations, including vascular stiffening and endothelial dysfunction. ³⁵

PVS‐BG as an Early Marker of Small Vessel Dysfunction and WMH Progression

PVS‐BG was positively associated with greater baseline WMH burden cross‐sectionally and WMH progression in longitudinal analysis. Model‐derived trajectories indicated that individuals in the high PVS group exhibited approximately one‐third greater WMH burden at baseline and after 4 years. This finding is aligned with the vascular interpretation of PVS‐BG. The basal ganglia are supplied by deep perforating arteries, which are particularly susceptible to hypertensive SVD. As such, PVS‐BG are thought to represent hypertensive arteriopathy (HA‐SVD; type 1 SVD), and considered to be linked more closely to cerebrovascular health, implicating pathogenic processes like blood brain barrier dysfunction and endothelial dysfunction. ³⁶ , ³⁷ , ³⁸ , ³⁹ , ⁴⁰ Accordingly, the link between PVS‐BG and WMH across multiple brain regions may potentially be attributed to diffuse microvascular dysfunction expected in HA‐SVD. The observation that PVS‐BG predicts future WMH progression, particularly in amyloid‐negative individuals, underscores its potential utility as an early marker of cerebrovascular injury, and as a prognostic marker to identify individuals at heightened risk of vascular‐driven disease processes and cognitive decline. That being said, existing studies on the association between PVS and cognition have been mixed, which could be attributed to vast heterogeneity in PVS measurement methods, and the reliance on subjective visual ratings which tend to be less sensitive. ⁴¹ , ⁴² Another plausible explanation is that worsening PVS may be more upstream than cognitive decline. Future work on the associations of quantitative PVS volume measures are warranted to understand the role of PVS.

Weak Links Between Baseline PVS and Longitudinal Progression of Amyloid Deposition

Baseline PVS‐BG was related to amyloid PiB cross‐sectionally, but not with longitudinal amyloid accumulation. Canonical correlation analysis revealed a multivariate pattern in which PVS burden in the basal ganglia, frontal lobe, and occipital lobe covaried most strongly with amyloid burden across all ROIs in cross‐sectional analysis.

Longitudinal analysis focused specifically on PVS‐BG, which predicted accelerated progression of WMH and amyloid in individuals who were amyloid‐negative, but not those who were amyloid‐positive. Although this suggests that PVS‐BG may be relevant in early pathogenic processes, there was weaker evidence for its involvement with amyloid deposition or tau deposition. This may plausibly represent early glymphatic dysfunction at a stage when amyloid accumulation is not yet fully established, in alignment with previous evidence that glymphatic dysfunction precedes and exacerbates amyloid deposition. ¹¹ , ⁴³ Conversely, the absence of associations in amyloid‐positive individuals may highlight the contribution of downstream neurodegenerative or inflammatory cascades at later disease stages, and/or plateauing of vascular/glymphatic processes. The lack of significant associations between PVS burden and PiB‐PET were also reported by others in patients living with AD and vascular cognitive impairment. ⁴⁴

Lack of Evidence for Baseline PVS Association With Longitudinal Tau Accumulation

Longitudinal analysis did not show an association between baseline PVS and longitudinal change in AV1451 binding. The absence of a longitudinal tau association aligns with the idea that tau accumulation is less closely linked to perivascular drainage than amyloid and WMH, and may be dominated by network‐level propagation mechanisms. ⁴⁵ , ⁴⁶

PVS‐BG was also associated with AV1451 binding in the basal ganglia, although this should be interpreted with caution due to well‐documented off‐target binding of AV1451, for example, to neuromelanin, MAO, iron‐rich structures, microhemorrhages, and other non‐tau proteins like TDP43 which may co‐localize in these regions. ⁴⁷ , ⁴⁸ , ⁴⁹ , ⁵⁰

PVS in CSO as a Potential Marker of Early CAA

PVS‐CSO did not exhibit significant associations with amyloid or tau pathology on PET imaging but demonstrated a regionally specific association with WMH in the occipital lobe. This is noteworthy, given that CAA preferentially involves the occipital lobes, ⁵¹ , ⁵² , ⁵³ with prior evidence that posterior‐dominant WMH are associated with strictly lobar CMB, a hallmark feature of CAA. ⁵² In CAA, PVS dilation is thought to represent chronic impairment of perivascular drainage of leptomeningeal arteries, leading to reduced amyloid removal and the development of CAA. ⁵⁴ , ⁵⁵ The link between PVS‐CSO and posterior WMH may be attributed to PVS‐CSO being anatomically and functionally linked to superficial WM tracts and posterior circulation territories. ⁵⁶ Furthermore, the CSO receives its blood supply via cortical and leptomeningeal vessels, which are the vessel categories typically affected by CAA through the build‐up of Aβ within these vessel walls, ⁵⁷ , ⁵⁸ , ⁵⁹ , ⁶⁰ reinforcing the biological plausibility of the observed association. The link between PVS‐CSO and CAA has also been evidenced in clinical studies, in which individuals with severe PVS‐CSO have an increased odds of lobar microbleeds and superficial siderosis, two hallmark features of CAA. ³⁸ , ⁵⁶ , ⁶¹ Such evidence has contributed to the addition of PVS‐CSO in the updated Boston Criteria version 2.0 for the diagnosis of CAA, ⁶¹ which is further supported in our data. Importantly, the lack of association with amyloid PET does not negate PVS as a relevant biomarker of CAA, given autopsy studies demonstrating that CAA does not relate to PiB‐PET which measures cortical parenchymal Aβ plaques. ⁶²

Taken together, our findings emphasize the importance of PVS‐CSO as a possible early imaging marker for identifying individuals at heightened risk of CAA and amyloid‐related imaging abnormalities (ARIA). Given the increasing prevalence of anti‐amyloid monoclonal antibody treatments for AD, incorporating regionally specific PVS assessments into safety monitoring protocols could improve early detection of vascular amyloid burden and ARIA risk prediction, potentially as dynamic indicators of treatment‐related microvascular complications. Altogether, further research is warranted to assess the utility of regional PVS assessment, particularly in the occipital lobe, in safety monitoring for anti‐amyloid treatment.

Implications

The distinct pathophysiological pathways highlighted in this study underscore the importance of considering PVS in a region‐specific manner. Results of this study have important implications for both research and clinical practice. The regional distinctiveness of PVS patterns could enhance early prognostic stratification for neurodegenerative and CVD progression. Importantly, this is also the timeframe when intervention can be most effective. Early risk assessment is particularly valuable because, at later stages, multiple concomitant pathologies often coexist, making it challenging to disentangle underlying etiology and thereby limiting the effectiveness of therapeutic interventions. By understanding the spatial PVS patterns relevant to disease progression (ie, PVS‐BG), this helps clinicians incorporate these markers into routine assessments to improve risk stratification and clinical care planning. PVS being an early marker of pathological processes represents one of its key strengths as a biomarker for AD.

Strengths and Limitations

The longitudinal design is a key strength of this study, allowing us to examine the predictive value of baseline PVS on longitudinal progression of AD markers. Other strengths include the sizable cohort with multimodal imaging and the wide age range of participants. Additionally, the combined assessment of global and regional PVS burden and multivariate analytical approaches allowed us to derive insights into etiological regional differences. Finally, quantitative PVS volumetry provided more precise and objective measurements compared with manual visual rating scales. Limitations include follow‐up duration; longer follow‐up periods may provide greater clarity to our mechanistic understanding. Furthermore, our sample composition of predominantly White individuals limits the generalizability of findings to more heterogeneous samples. Finally, discrepancies between cross‐sectional and longitudinal findings may be partially attributed to differences in sample composition across analyses, given that longitudinal analysis was conducted on a smaller subset of participants.

Conclusions

Our findings suggest that PVS‐BG could serve as an early indicator of WMH progression, but less so far AD biomarker progression, reinforcing the potential role of PVS‐BG as an early marker of SVD, rather than an AD‐specific biomarker. Conversely, PVS‐CSO, particularly in the parietal‐occipital region, was associated with WMH in the occipital lobe, and therefore may be reflective of CAA‐related processes.

Author Contributions

A.L. and P.V. contributed to the conception and design of the study; A.L., J.L.G., M.H., S.R., S.A.P., C.G.S., K.K., V.J.L., R.C.P., J.G‐R., C.R.J.J., and P.V. contributed to the acquisition and analysis of data; A.L. and J.L.G. contributed to drafting the text and preparing the figures.

Potential Conflicts of Interest

The authors do not have any pertinent disclosures relevant to this study.

Supporting information

Supplementary Figure S1. Density plot of the distribution of baseline PVS volume as a function of age in MCSA sample.

Supplementary Table S1. Longitudinal analysis of baseline perivascular spaces in the basal ganglia (PVS‐BG) on progression of white matter hyperintensity (WMH), amyloid, and tau, split by amyloid positivity.

Data Availability

MCSA data is shared widely on GAAIN: https://www.gaaindata.org/partner/MCSA.