Perilesional white matter gradients reveal microstructural differences in cerebral amyloid angiopathy versus Alzheimer's disease

Abstract

INTRODUCTION

White matter hyperintensities (WMHs) are common in both Alzheimer's disease (AD) and cerebral amyloid angiopathy (CAA), yet their spatial tissue characteristics and microstructural differences remain poorly understood.

METHODS

We analyzed 351 participants: 184 amyloid beta (Aβ)‐positive AD and mild cognitive impairment (MCI), 139 Aβ‐negative cognitively normal controls (CN), and 28 probable CAA. Multimodal magnetic resonance imaging metrics were used to estimate spatial gradient parameters for periventricular WMHs (pWMH) and deep WMHs (dWMH).

RESULTS

CAA demonstrated distinctive free‐water fraction (FWF), fractional anisotropy (FA), mean diffusivity (MD), and plasma volume within pWMH, as well as spatial gradient parameters of pWMH. These pWMH spatial gradient parameters produced area under the curve (AUC) values of 0.71 (FWF), 0.72 (MD), and 0.79 (FA) when distinguishing CAA from AD/MCI. We retested a subset of the cohort after 1 to 2 years (AUCs: FWF = 0.89, MD = 0.79, FA = 0.85).

DISCUSSION

Spatial gradient parameters reflect disease‐specific microstructural and vascular changes, providing insights into CAA and AD pathology.

Highlights

• We studied spatial gradients of white matter microstructure in CAA and AD.

• CAA showed distinctive microstructural characteristics and gradient parameters in pWMH.

• pWMH gradient parameters reflect distinct pathology in CAA versus AD.

1. BACKGROUND

White matter hyperintensities (WMHs) are lesions in the brain's white matter that appear hyperintense on T2‐weighted and fluid‐attenuated inversion recovery (FLAIR) magnetic resonance imaging (MRI) images. ¹ These WMHs are commonly observed in the aging brain ² , ³ and in various neurological conditions such as cerebral small vessel disease (cSVD) ⁴ and Alzheimer's disease (AD). ⁵ In cSVD, diffusion and perfusion MRI have revealed microstructural damage and microvascular dysfunction within the WMH core. ⁶ , ⁷ Moreover, the normal‐appearing white matter (NAWM) surrounding WMHs also exhibits reduced structural integrity and perfusion and has been associated with an increased risk of WMH progression. ⁸ , ⁹ , ¹⁰

Beyond cSVD, previous studies demonstrated that WMHs were also associated with AD pathology ¹¹ and that greater WMH burden in AD correlates with worse cognitive performance ¹² and increased neurodegeneration. ¹³ The revised National Institute on Aging–Alzheimer's Association (NIA–AA) research framework (2024) expanded AD biomarkers from the traditional AT(N) system to the AT(N)‐IVS (inflammation, vascular factor, synucleinopathy) model. The AT(N) system makes a diagnosis based on the presence or absence of amyloid, tau, and neurodegeneration, and the AT(N)‐IVS model integrates the vascular contribution to AD as well. Notably, WMHs are now explicitly recognized as the primary recommended biomarker for vascular injury (V) within this updated framework. ¹⁴ Currently, most clinical MRI reports simply describe “WMH burden” qualitatively or via Fazekas scoring. ¹⁵ A framework should be introduced to create a quantitative, region‐ and disease‐specific profiling method.

The Fazekas scoring system classifies WMHs according to their spatial distribution into periventricular and deep WMHs (pWMH and dWMH). Periventricular and deep WMHs exhibit distinct microstructural and microvascular characteristics in cSVD. ¹⁶ Spatial gradients in FLAIR intensity, microstructural metrics – such as mean diffusivity (MD), fractional anisotropy (FA), and free‐water fraction (FWF) – as well as perfusion parameters and blood–brain barrier (BBB) integrity, have been shown to differ between pWMH and dWMH regions in cSVD.

Cerebral amyloid angiopathy (CAA) is a type of cSVD, ¹⁷ characterized by the progressive deposition of amyloid beta (Aβ) within the walls of small cortical and leptomeningeal arteries. ¹⁸ Both CAA and AD are strongly associated with the presence of WMHs, with both conditions showing significant effects on white matter integrity. ¹⁹ However, current knowledge remains limited regarding how microstructure changes from WMH to surrounding NAWM in AD compared to CAA. Addressing this gap is clinically important, as CAA‐like WMHs may heighten the vascular risk for amyloid‐related imaging abnormalities with hemorrhage (ARIA‐H) during Aβ‐targeted immunotherapy. ²⁰

In this study, we aimed to investigate the distinct spatial trajectories of white matter microstructural and vascular integrity surrounding WMHs in CAA and AD. By quantifying perilesional gradients of diffusion and perfusion markers, we sought to reveal disease‐specific changes reflecting microvascular remodeling, BBB disruption, and tissue degeneration. The term “gradient” refers to the macroscale spatial trend of MRI‐derived tissue properties across concentric layers surrounding WMH lesions, rather than voxel‐level intensity gradients.

2. METHODS

2.1. Participants and study design

This study utilized data from the ongoing Ruijin NeuroBank of Alzheimer's Disease and Dementia (RJNB‐D) cohort, a prospective study enrolling participants between November 2021 and December 2024. The RJNB‐D cohort comprises individuals with all‐cause dementia recruited from memory clinics and cognitively normal controls (CN), with detailed characteristics described in our previous publications. ²¹ Cognitively normal volunteers were recruited from a neighboring community. The flow chart of this study is shown in Figure 1. Medical history included prior lacunar stroke, diabetes mellitus, coronary artery disease, hyperlipidemia, and current use of medications such as antiplatelet agents and statins. The study was also registered with ClinicalTrials.gov (NCT05623124). This study was approved by the institutional ethics committee. All participants provided written informed consent.

FIGURE 1.

Flow chart of present study: participant inclusion pipeline of study.

For the primary analysis (main dataset), we included 351 participants at their baseline visit, consisting of (1) Aβ‐positive (Aβ+) AD/MCI (n = 184), (2) Aβ‐negative (Aβ−) CN (n = 139), and (3) probable CAA (n = 28) diagnosed according to the modified Boston criteria version 2.0. ¹⁸ The criteria for diagnosing AD and MCI adhered to the research criteria established by the NIA–AA workgroups. ²² Aβ positivity was determined by visual check of ¹⁸F Florbetapir (FBP) positron emission tomography (PET) scans by two senior PET experts (F. X. and Q. H.).

To assess the reproducibility of findings, a retest analysis was conducted in a longitudinal subset (retest dataset) of the primary cohort based on imaging data acquired at 12‐ or 24‐month follow‐up, including 29 visits from AD/MCI participants and 12 from probable CAA participants.

2.2. Image acquisition and processing

The multimodal MRI utilized in the study included: three‐dimensional (3D) T1‐weighted, 3D T2 FLAIR, arterial spin labeling (ASL), dynamic contrast‐enhanced (DCE), and multishell diffusion MRI. The detailed acquisition and processing protocols are provided in the Supplementary Materials.

The following metrics were extracted from WMH regions of interest (ROIs) after multimodal image co‐registration into the individual FLAIR image space: FA, MD, and FWF calculated from diffusion MRI, cerebral blood flow (CBF, mL/min/100 g) calculated from ASL‐MRI, Ktrans (volume transfer constant between blood plasma and the extravascular extracellular space), and Vp (blood plasma volume fraction) computed from DCE‐MRI data.

RESEARCH IN CONTEXT

1. Systematic review: Existing literature highlights WMHs in AD and CAA, focusing on their association with imaging markers. However, the spatial tissue characteristics of these markers from lesion core to surrounding NAWM remains insufficiently understood, particularly in differentiating AD from CAA.

2. Interpretation: Our study provides new insights by quantifying spatial gradients of microstructural, perfusion, and BBB metrics around WMHs, revealing distinct differences between AD and CAA. These spatially resolved features provide mechanistic insights into disease‐specific injury processes.

3. Future directions: Future research should explore longitudinal changes in these spatial gradients and their correlation with clinical outcomes. Investigating the underlying molecular mechanisms and validating these findings in larger, multicenter cohorts will be crucial for developing biomarkers for early diagnosis and treatment monitoring in AD and CAA.

2.3. Spatial tissue gradient analysis

To investigate the spatial gradient of MRI measures across WMH borders, we generated concentric ROIs extending both inward and outward from the WMH boundary (Figure 2, left). We defined 0 mm as the WMH boundary. Inward concentric ROIs were defined as −2 to −1 mm and −1 to 0 mm from the WMH boundary. For dWMH, outer ROIs were limited to 0 to 1 mm, 1 to 2, mm and 2 to 3 mm in the perilesional NAWM, whereas for pWMHs, outer ROIs extended to 0 to 2 mm, 2 to 4 mm, 4 to 6 mm, and 6 to 8 mm. This asymmetric design was chosen because dWMH clusters could locate near the cortical surface and are closely spaced, making larger outer ROIs more susceptible to overlap or contamination by adjacent lesions or other tissue types. Overlapping ROIs were excluded from further analysis.

FIGURE 2.

T2‐fluid‐attenuated inversion recovery (FLAIR) signal intensity in periventricular and deep white matter hyperintensities in cognitively normal controls. Left: normalized T2‐FLAIR signal intensity profiles across concentric regions of interest (ROIs) in periventricular white matter hyperintensities (pWMH) and deep WMH (dWMH). Right: border slope (−1 to 0 mm to 0 to 2 mm, −0.218) was steeper than both the inner slope (−2 to −1 mm to −1 to 0 mm, −0.116, p < 0.001) and outer slope (4 to 6 mm to 6 to 8 mm, −0.006, p < 0.001) in pWMH. Corresponding profiles in dWMH: The border slope (−0.267) similarly exceeded inner (−0.128, p < 0.001) and outer (−0.026, p < 0.001) slopes. Solid lines mark the visual border of WMHs. Error bands represent 95% confidence intervals.

We examined normalized FLAIR intensity profiles in CN and confirmed that signal intensity in the outermost ROIs approached levels typical of NAWM, supporting the spatial definition boundaries used for both pWMHs and dWMHs. Specifically, to assess the spatial specificity of these boundaries, we computed three distinct slopes across concentric ROIs: the inner WMH slope, the WMH boundary slope, and the outer NAWM slope (Figure 2, right).

For each subject and imaging metric, we applied piece‐wise linear regression to the spatial profile across the WMHs and its surrounding NAWM. The model divides the signal trajectory into two linear segments and identifies the optimal breakpoint that minimizes the total mean squared error (MSE). Continuity was not enforced in the model, allowing for a potential step change at the breakpoint to flexibly capture abrupt transitions in tissue properties.

The breakpoint is the spatial location (in millimeters [mm]) along the WMH gradient where the slope of tissue changes shifts most significantly. It marks the transition between WMH‐core influence and surrounding NAWM‐like tissue and thus may indicate the spatial extent of lesion‐induced tissue alteration. Accordingly, “Slope1” refers to the rate of change within the inner segment (core to breakpoint), while “Slope2” refers to the rate of change in the outer segment (breakpoint to outermost NAWM). In the spatial gradient analysis across different disease groups, we introduced the breakpoint and derived Slope1 and Slope2.

2.4. Statistical analysis

Prior to group comparison, all metrics were tested for right‐skewed distribution using the Fisher–Pearson coefficient of skewness. Variables with skewness > 1 were log‐transformed. Group comparisons of demographic and clinical variables across diagnostic groups were performed using one‐way analysis of variance (ANOVA), followed by post hoc Tukey's honestly significant difference (HSD) test.

Ordinary least‐squares regression was used for group comparison of each metric (FA, MD, FWF, CBF, Ktrans, Vp) within periventricular or deep WMH ROIs, adjusting for age, sex, and atherosclerosis‐related factors (hypertension, diabetes mellitus, hyperlipidemia, and smoking status). After piece‐wise linear regression, we compared Slope1, Slope2, and breakpoint values across diagnostic groups. One‐way ANOVA and post hoc method were used to assess group differences. The WMH burden, age, sex, and atherosclerosis‐related factors (hypertension, diabetes mellitus, hyperlipidemia, and smoking status) were included as covariates in the analysis of covariance. p values were corrected for multiple comparisons using the Benjamini–Hochberg false discovery rate (FDR) procedure.

Logistic regression models were applied to assess the diagnostic performance of WMH gradient features in differentiating CAA from the AD continuum (MCI and AD), with receiver operating characteristic (ROC) analysis to calculate area under curve (AUC) values. In the main dataset, AUC values were computed in‐sample, based on logistic regression models trained and tested on the full dataset. In the retest dataset, we used a fully independent validation strategy: Models were trained on the main dataset and evaluated on the retest cohort, which consisted of follow‐up scans obtained 12 to 24 months later. Importantly, the retest subjects were not used in any part of model training or standardization, ensuring complete independence between training and testing phases.

The confidence interval (CIs) was obtained using Tukey's HSD test after a one‐way ANOVA or computed separately using pooled standard deviation from each pair with standardized effect sizes (Cohen's d). The CIs for AUC were estimated using the bootstrap method with 10,000 resamples. All modeling and statistical analyses were implemented in Python 3.9 using scikit‐learn (for piece‐wise linear regression), statsmodels (for covariate‐adjusted comparisons, ANOVA, and ANCOVA), and SciPy (for skewness testing), with visualization generated via seaborn and matplotlib.

3. RESULTS

The demographic characteristics and vascular risk factors of the participants are summarized in Table 1. The CAA had the highest WMH burden, followed by AD, MCI, and CN (p < 0.001). Lacunar stroke prevalence was highest in CAA (32.1%), exceeding rates in AD, MCI, and CN (all p < 0.05), which exhibited comparable frequencies.

TABLE 1.

Demographic and clinical characteristics of participants by diagnostic group.

	CN (n = 139)	MCI (n = 77)	AD (n = 107)	CAA (n = 28)	p value (AD vs CAA)	p value (AD vs CN)	p value (AD vs MCI)	p value (CAA vs CN)	p value (CAA vs MCI)	p value (CN vs MCI)
Age	66.9 ± 8.1	71.4 ± 7.3	70.6 ± 7.7	73.7 ± 6.3	0.044	<0.001	0.456	<0.001	0.123	<0.001
Education (years)	12.2 ± 3.4	12.4 ± 3.7	11.0 ± 4.3	11.6 ± 3.3	0.361	0.017	0.016	0.184	0.158	0.694
MMSE	28.7 ± 1.1	26.4 ± 1.7	16.9 ± 5.1	21.5 ± 5.7	<0.001	<0.001	<0.001	<0.001	<0.001	<0.001
Sex (female, %)	86 (61.9%)	46 (59.7%)	67 (62.6%)	7 (25.0%)	0.011	0.939	0.768	0.008	0.033	0.706
Smoking	29 (20.9%)	17 (22.1%)	22 (20.6%)	11 (39.3%)	0.071	0.920	0.948	0.066	0.130	0.972
Hypertension	67 (48.2%)	35 (45.5%)	46 (43.0%)	18 (64.3%)	0.072	0.494	0.856	0.178	0.137	0.806
Diabetes	29 (20.9%)	9 (11.7%)	10 (9.3%)	9 (32.1%)	0.005	0.023	0.787	0.293	0.030	0.131
Hyperlipidemia	64 (46.0%)	22 (28.6%)	26 (24.3%)	12 (42.9%)	0.088	<0.001	0.631	0.920	0.251	0.018
CAD	18 (12.9%)	10 (13.0%)	11 (10.3%)	8 (28.6%)	0.030	0.657	0.738	0.073	0.114	0.839
Statin use	61 (43.9%)	24 (31.2%)	22 (20.6%)	14 (50.0%)	0.004	<0.001	0.142	0.700	0.122	0.092
Antiplatelet use	29 (20.9%)	11 (14.3%)	14 (13.1%)	5 (17.9%)	0.733	0.155	0.987	0.918	0.886	0.313
Lacunar stroke	15 (10.8%)	1 (1.3%)	5 (4.7%)	9 (32.1%)	<0.001	0.132	0.403	0.008	<0.001	0.012
WMH burden	0.008 ± .010	0.010 ± .010	0.013 ± .0100	0.034 ± .019	<0.001	<0.001	0.049	<0.001	<0.001	0.115

Note: Continuous variables are reported as mean ± standard deviation; categorical variables as n (%).

p values for post hoc pair‐wise comparisons (age, education, MMSE, WMH burden) were obtained by Tukey's Honestly Significant Difference tests; for categorical measures (sex, smoking, hypertension, diabetes, hyperlipidemia, CAD, statin use, antiplatelet use, lacunar stroke) by χ ² test or Fisher's exact test as appropriate. Unadjusted p values are reported for descriptive comparison of baseline characteristics here only here.

Abbreviations: AD, Alzheimer's disease; CAA, cerebral amyloid angiopathy; CAD,  coronary artery disease; CN, cognitively normal; MCI, mild cognitive impairment; MMSE, Mini‐Mental State Examination; WMH, white matter hyperintensity.

3.1. WMH segmentation validation

To validate the spatial specificity of WMH segmentation, we analyzed the FLAIR intensity gradients across concentric ROIs in CN (Figure 2). In the pWMH region, the slope across the WMH boundary (from the −1 to 0 mm eroded ROI to the 0 to 2 mm outer margin) was significantly steeper than both the inner WMH slope (from −2 to ‐1 mm to −1 to 0 mm; Cohen's d = −1.052; 95% CI: −0.120 to −0.086; p < 0.001) and the outer WMH slope (from 4 to 6 mm to 6 to 8 mm; p < 0.001). Similarly, in the dWMH region, the boundary slope also exceeded both the inner slope (Cohen's d = −1.440; 95% CI: −0.163 to −0.123; p < 0.001) and the outer slope (Cohen's d = −2.897; 95% CI: −0.226 to −0.202; p < 0.001) (Figure 2).

As expected, periventricular and deep WMHs exhibited distinct biological profiles in CN. Compared to pWMHs, dWMHs showed higher perfusion, lower BBB permeability, and more preserved microstructural integrity in CN (Table S1).

3.2. Differences in metrics within WMH

We further compared MRI imaging metrics across diagnostic groups within the pWMH and dWMH regions, adjusting for age, sex, and atherosclerosis‐related factors.

The mean FWF in the pWMH region showed a decreasing trend from CN to MCI, AD, and CAA groups (Figure 3). FWF was lower in the CAA group compared to CN (95% CI: −0.073 to −0.012; FDR‐corrected p < 0.05; Table S2). No FWF differences were found between groups in the dWMHs.

FIGURE 3.

Group differences in microstructural and perfusion metrics within white matter hyperintensity (WMH) regions. Boxplots (mean ± SD) of free‐water fraction (FWF), fractional anisotropy (FA), mean diffusivity (MD), plasma volume (Vp), cerebral blood flow (CBF), and the volume transfer constant (Ktrans) across cognitively normal controls, mild cognitive impairment, Alzheimer's disease, and cerebral amyloid angiopathy groups in periventricular WMH (pWMH) and deep WMH (dWMH). Boxplots show median (line), interquartile range (boxes), and outliers (dots). Adjusted for age, sex, and atherosclerosis‐related factors. *p < 0.05; **p < 0.01.

In the pWMH region, FA was increased in AD compared to CN (95% CI: 0.015 to 0.042; p = 0.001). In the same region, FA was also markedly elevated in CAA compared to both CN (95% CI: 0.018 to 0.062; p = 0.004) and MCI (95% CI: 0.007 to 0.048; p = 0.043). No difference in pWMH FA was found between AD and CAA (95% CI: −0.012 to 0.036; p = 0.454). In contrast, a trend‐level reduction in dWMH FA was also observed in CAA compared to AD (95% CI: −0.048 to −0.005; p = 0.077). CAA showed lower FA in dWMH than MCI (95% CI: −0.056 to −0.012; p = 0.023) and CN (95% CI: 0.041 to 0.114; p = 0.025, Figure 3, Table S2).

Similarly, after adjusting for FWF, pWMH FA in both AD and CAA remained elevated relative to CN (p = 0.004 and 0.041), and dWMH FA was significantly reduced in CAA compared with MCI and CN (p = 0.039 and 0.038).

In the pWMH region, the MD was elevated in the CAA compared to CN (95% CI: 0.041 to 0.114; p = 0.001) and MCI (95% CI: 0.034 to 0.132; p = 0.012), and also higher in CAA than in AD (95% CI: 0.230 to 0.098; p = 0.017). A marginal increase in MD was also observed in AD compared to CN (95% CI: 0.004 to 0.043; p = 0.082). No significant difference was found in the dWMHs (Table S2). After controlling for the FWF, pWMH MD in CAA was also higher than AD (p = 0.035), MCI (p = 0.033), and CN (p = 0.019).

In the pWMH region, Vp was lower in CAA relative to MCI (95% CI: −1.627 to −0.583; p = 0.003) and AD (95% CI: −1.232 to −0.236; p = 0.032); a similar pattern was observed in the dWMH region (CAA vs MCI: 95% CI: −1.125 to −0.379; p = 0.004). A trend‐level reduction in Vp was also found in CAA relative to AD (95% CI: −0.908 to −0.063; p = 0.093). In contrast, no group differences in CBF or Ktrans were observed in both pWMHs and dWMHs (Figure 3).

3.3. Differences in WMH microstructure gradient

In the pWMH region, Slope2 of FWF was steeper in CN compared to CAA (Cohen's d = −0.850; 95% CI: −1.269 to −0.431; p = 0.001) and AD (Cohen's d = −0.543; 95% CI: −0.969 to −0.119; p = 0.018). MCI also exhibited a faster FWF decay than CAA (p = 0.010). Furthermore, the FWF breakpoint was more distal from the WMH lesion in CAA relative to CN (Cohen's d = 0.663; 95% CI:  0.248 to 1.078; p = 0.011), MCI (Cohen's d = 0.770; 95% CI:  0.320 to 1.220; p = 0.006), and AD (Cohen's d = 0.602; 95% CI:  0.176 to 1.028; p = 0.042; Figure 4, Table 2). No significant group differences were observed for Slope1 in either the pWMH or dWMH regions. After adjusting for age, sex, WMH burden, and atherosclerosis‐related factors, the differences in FWF Slope2 between groups were not significant (Supplementary Table 3). No differences in slope or breakpoint were detected for FWF in the dWMH region (Table 2, Figure S1).

FIGURE 4.

Spatial gradients of free‐water fraction (FWF), fractional anisotropy (FA), and mean diffusivity (MD) in periventricular white matter hyperintensities (pWMHs) across diagnostic groups. Piece‐wise linear regression analyses demonstrate distinct spatial patterns of FWF, FA, and MD in pWMHs among cognitively normal controls (CN), mild cognitive impairment (MCI), Alzheimer's disease (AD), and cerebral amyloid angiopathy (CAA) groups. Dashed vertical lines indicate breakpoint locations. Blue and red dashed lines indicate the linear regression lines before and after the breakpoint.

TABLE 2.

Group‐wise comparisons of two‐segment linear fit metrics in periventricular (pWMH) and deep (dWMH) white matter hyperintensities across diagnostic groups.

Region	Metric	CN	MCI	AD	CAA	FDR p value	Significant pair‐wise (Tukey)
pWMH
FWF	Slope1	−0.0397	−0.0322	−0.0326	−0.0444	0.630	—
	Slope2	−0.0102	−0.00933	−0.00817	−0.0056	0.012	AD versus CN (p = 0.018)
							CAA versus CN (p = 0.001)
							CAA versus MCI (p = 0.010)
	Breakpoint	2.25	2.17	2.33	2.86	0.089	AD versus CAA (p = 0.042)
							CAA versus CN (p = 0.011)
							CAA versus MCI (p = 0.006)
CBF	Slope1	3.3	2.83	3.27	3.42	0.783	—
	Slope2	1.32	1.23	1.41	1.69	0.456	—
	Breakpoint	1.72	1.57	1.49	2.08	0.050	AD versus CAA (p = 0.022)
FA	Slope1	0.0514	0.0545	0.0452	0.0429	0.063	—
	Slope2	0.00191	−0.00311	−0.0064	−0.0145	<0.001	AD versus CAA (p = 0.003)
							AD versus CN (p = 0.001)
							CAA versus CN (p = 0.001)
							CAA versus MCI (p = 0.001)
							CN versus MCI (p = 0.007)
	Breakpoint	2.87	2.84	2.83	2.79	0.880	—
MD	Slope1	−3.38e‐05	−3.76e‐05	−3.55e‐05	−3.81e‐05	0.775	—
	Slope2	−4.29e‐06	−3.03e‐06	−1.06e‐06	1.37e‐06	<0.001	AD versus CAA (p = 0.049)
							AD versus CN (p = 0.001)
							AD versus MCI (p = 0.016)
							CAA versus CN (p = 0.001)
							CAA versus MCI (p = 0.001)
	Breakpoint	2.76	2.71	2.89	3	0.042	—
Ktrans	Slope1	9.3e‐05	0.000279	2.95e‐05	6.73e‐05	0.145	—
	Slope2	−3.5e‐06	−3.49e‐05	−1.8e‐06	2.88e‐05	0.007	CAA versus MCI (p = 0.041)
	Breakpoint	1.42	1.43	1.63	2.43	0.062	CAA versus CN (p = 0.045)
							CAA versus MCI (p = 0.044)
Vp	Slope1	0.00236	0.00131	0.00127	0.000403	0.241	—
	Slope2	−0.000102	−5.54e‐05	8.14e‐05	0.000279	0.155	—
	Breakpoint	1.5	1.5	1.63	2.14	0.509	—
dWMH
FWF	Slope1	−0.0456	−0.0415	−0.0597	−0.0573	0.155	—
	Slope2	−0.0285	−0.0254	−0.0277	−0.0306	0.257	—
	Breakpoint	0.975	0.974	0.921	1.04	0.889	—
CBF	Slope1	2.34	2.11	1.68	4.06	0.177	—
	Slope2	1.19	1.11	0.922	1.81	0.072	—
	Breakpoint	0.887	0.804	0.775	0.692	0.833	—
FA	Slope1	0.0425	0.047	0.0369	0.0435	0.427	—
	Slope2	0.00462	−0.00212	−0.000709	−0.000526	0.051	AD versus CN (p = 0.029)
							CN versus MCI (p = 0.009)
	Breakpoint	1.08	0.961	0.939	1.18	0.164	—
MD	Slope1	−3.36e‐05	−3.18e‐05	−3.12e‐05	−4.36e‐05	0.106	—
	Slope2	−1.39e‐05	−1.03e‐05	−8.42e‐06	−9.64e‐06	0.011	AD versus CN (p = 0.001)
	Breakpoint	1.06	1.03	0.986	1.11	0.930	—
Ktrans	Slope1	9.19e‐06	3.06e‐05	−9.97e‐06	−1.73e‐05	0.791	—
	Slope2	9.96e‐06	2.66e‐05	−5.82e‐06	6.11e‐05	0.167	—
	Breakpoint	0.833	0.75	0.789	0.929	0.850	—
Vp	Slope1	−0.000352	0.000216	0.000477	0.000117	0.212	—
	Slope2	0.000452	0.000418	0.000306	0.000553	0.065	—
	Breakpoint	0.958	0.929	0.868	1.21	0.397	—

Note: Data are mean values (slope in units per millimeter, breakpoint in millimeters) derived from piece‐wise linear fits of each MRI metric in ROIs around or in WMH lesions. ANOVA p values are from one‐way ANOVA across four diagnostic groups (CN, MCI, AD, CAA). “Significant pair‐wise (Tukey)” lists only those post hoc contrasts with p < 0.05. FDR p values were corrected for multiple comparisons using the Benjamini–Hochberg false discovery rate procedure.

Abbreviations: AD, Alzheimer's disease; breakpoint, distance at which gradient changes; CAA, cerebral amyloid angiopathy; CBF, cerebral blood flow; CN, cognitively normal; dWMH, deep white matter hyperintensity; FA, fractional anisotropy; FLAIR, normalized FLAIR intensity, FWF, free‐water fraction; Ktrans, volume transfer constant; MCI, mild cognitive impairment; MD, mean diffusivity; pWMH, periventricular white matter hyperintensity; Vp, plasma volume.

In terms of MD, CN showed a steeper decline in the outer pWMH region compared to both AD and CAA (CN vs AD: Cohen's d = −0.716; 95% CI: −0.978 to −0.455; p = 0.001; CN vs CAA: Cohen's d = −1.270; 95% CI: −1.702 to −0.839; p = 0.001). MCI also exhibited a faster MD decay than AD (Cohen's d = −0.448; 95% CI: −0.746 to −0.149; p = 0.016). In contrast, CAA (Cohen's d = −1.045; 95% CI: −1.505 to −0.585; p = 0.001) demonstrated a positive MD slope in this region (Table 2). The results remained generally consistent after adjusting for age, sex, WMH burden, and atherosclerosis‐related factors (Table S3). In the outer dWMH region, CN exhibited a steeper MD gradient (Slope2) than MCI and AD, whereas CAA showed the most flattened gradient following covariates adjustment (Table S3).

FA Slope2 also distinguished between diagnostic groups. In outer pWMH, CAA exhibited a steeper decline compared to MCI (Cohen's d = 1.180; 95% CI: 0.713 to 1.646; p = 0.001) and AD (Cohen's d = −0.803; 95% CI: −1.234 to −0.373; p = 0.003), and the results remained unchanged after covariate adjustment.

Breakpoint positions for FA did not differ across groups (p > 0.05). A similar pattern was observed in the dWMH region, where CN maintained a rising FA profile, whereas CAA, AD, and MCI displayed declining slopes (Table 2, Figure S1).

To balance the sample size in AD/MCI and CAA, we performed an age‐ and MMSE‐matched selection of AD individuals using a nearest‐neighbor approach. A large neighborhood size was chosen to identify a broad candidate pool and finally 35 AD individuals with the smallest distances were selected to form the final matched AD subgroup (Table S4). In the pWMH region, the FWF, MD, and FA breakpoints were more distal from the WMH lesions in CAA relative to the subgroup AD (FWF: 95% CI: 0.208 to 0.487; p < 0.001; MD: 95% CI: 0.145 to 0.325; p < 0.001; FA: 95% CI: 0.057 to 0.232; p = 0.002), while these metrics in the dWMH region were not significantly different between CAA and AD subgroup.

To clarify any effect from Aβ positivity in CAA, we compared these microstructure gradients between Aβ+ CAA (n = 14) and Aβ− CAA (n = 14). The two CAA subgroups were matched in age and sex (Table S5). Within the CAA group, Aβ+ CAA (n = 14) did not have differences from Aβ− CAA in the aforementioned diffusion MRI‐derived gradient metrics.

3.4. Difference in WMH perfusion and BBB gradient

In the pWMH region, the CBF breakpoint was located farther from the lesion core in CAA compared to AD (Cohen's d = 0.654; 95% CI:  0.212 to 1.096; p = 0.022, Table 2). After adjusting for the covariates, the CBF breakpoint remained most distal in CAA with marginal significance (p = 0.095; Table S3). In the dWMH region, no significant group differences in breakpoint position were observed.

Regarding BBB permeability, Slope2 of Ktrans in pWMH was positive in the CAA group, and the breakpoint was located farther from the lesion core compared to CN (Cohen's d = 1.174; 95% CI:  0.244 to 2.104; p = 0.045) and MCI (Cohen's d = 1.159; 95% CI:  0.254 to 2.063; p = 0.044). After adjusting for covariates, a similar trend was observed (p = 0.084). No differences in Ktrans slopes or breakpoint were found in the dWMH region. For Vp, no group‐level differences in Slope2 or breakpoint were observed in either pWMH or dWMH regions (Figures S2 and S3; all p > 0.05).

For the matched CAA and AD subgroup, the CBF, Ktrans, and Vp breakpoints were more distal from the pWMH lesion in CAA (CBF: 95% CI: 0.453 to 0.805; p < 0.001; Ktrans: 95% CI: 0.786 to 1.270; p < 0.001; Vp: 95% CI: 0.481 to 1.004; p < 0.001). In the dWMH region, these metrics were not significantly different between CAA and AD subgroup. Similarly to dMRI‐derived metrics, Aβ+ CAA (n = 14) had no differences from Aβ− CAA in CBF and BBB‐related gradient metrics.

3.5. ROC analysis of WMH gradients for differentiating CAA from AD

In the pWMH region, FA‐based gradients yielded an AUC of 0.79 (95% CI: 0.719 to 0.860). The MD (AUC = 0.72, 95% CI: 0.622 to 0.798) and FWF (AUC = 0.71, 95% CI: 0.620 to 0.802) also demonstrated moderate discriminative ability. The normalized FLAIR signal gradients reached an AUC of 0.70 (95% CI: 0.600 to 0.781), while CBF yielded an AUC of 0.68 (95% CI: 0.579 to 0.794). In the dWMH region, the best classification came from FLAIR (AUC = 0.72, 95% CI: 0.600 to 0.814), followed by CBF (AUC = 0.70, 95% CI: 0.597 to 0.798), while FWF (AUC = 0.61, 95% CI: 0.506 to 0.700), FA (AUC = 0.63, 95% CI: 0.527 to 0.733), and MD (AUC = 0.64, 95% CI: 0.556 to 0.732) showed relatively lower discriminatory ability (Figure 5).

FIGURE 5.

Receiver operating characteristic (ROC) curves for white matter hyperintensity (WMH) gradient features distinguishing cerebral amyloid angiopathy (CAA) from the Alzheimer's disease (AD) continuum. ROC curves were generated for five imaging‐derived gradient metrics around periventricular WMHs (pWMHs) (upper panel) and deep WMHs (dWMHs) (lower panel). In the main pWMH dataset, fractional anisotropy (FA) gradient metrics achieved an area under the curve (AUC) of 0.79, mean diffusivity (MD) 0.72, free‐water fraction (FWF) 0.71, normalized fluid‐attenuated inversion recovery (FLAIR) intensity 0.70, and cerebral blood flow (CBF) 0.68. In the retest dataset, AUCs were 0.85 for FA, 0.79 for MD, 0.89 for FWF, 0.82 for FLAIR, and 0.59 for CBF. The dWMH ROC analyses show: FA AUC = 0.63 and 0.69; MD = 0.64 and 0.68; FWF = 0.61 and 0.69; FLAIR = 0.72 and 0.79; CBF = 0.70 and 0.77, in the main and retest dataset respectively. The diagonal line (AUC = 0.50) is shown for reference.

3.6. Validation of WMH metrics for CAA differentiation

The ROC models trained in the main dataset were applied to the retest dataset (Table S6). In the pWMH region, the FWF achieved the highest diagnostic accuracy with an AUC of 0.89 (95% CI: 0.778 to 0.977; Figure 5), followed closely by FA (AUC = 0.85, 95% CI: 0.714 to 0.957) and FLAIR signal intensity (AUC = 0.82, 95% CI: 0.675 to 0.938). The MD also showed substantial predictive value (AUC = 0.79, 95% CI: 0.638 to 0.913), while CBF exhibited limited classification utility in this region (AUC = 0.59, 95% CI: 0.340 to 0.844).

The dWMH region showed generally lower discriminative performance: FLAIR intensity and CBF demonstrated the predictive power with AUCs of 0.79 and 0.77 (95% CI: 0.645 to 0.920 and 0.613 to 0.913), respectively. Other parameters such as FWF, FA, and MD yielded more modest performance (AUCs ranging from 0.68 to 0.69; Figure 5).

4. DISCUSSION

In this study, we demonstrated that AD and CAA exhibit distinct white matter microstructural and vascular profiles, as captured by spatial gradients and inflection points derived from concentric ROI modeling. Specifically, pWMH in CAA exhibited more distal breakpoints in metrics such as FWF, Ktrans, and CBF, reflecting broader perilesional tissue disruption compared to the AD/MCI continuum. These alterations were less pronounced in dWMH, underscoring a region‐specific vulnerability in CAA. Moreover, differences in FWF, MD, and FA gradients between CAA and AD continuum provide additional insight into the heterogeneity of pWMHs.

We compared microstructural and microvascular markers across WMH subtypes in different disease groups and found more severe damage in CAA. In the pWMH region, we observed a step‐wise increase in MD across diagnostic categories, from CN to MCI, AD, and CAA. The dWMH region showed a trend‐level reduction in FA specifically in the CAA group. These results suggest that changes in tissue water mobility are more pronounced in CAA. This supports the hypothesis that amyloid deposition in the blood vessels of CAA may disrupt the BBB and impair drainage, potentially leading to increased extracellular fluid. ²³ , ²⁴ In contrast, white matter changes in AD are often thought to be primarily driven by downstream degeneration associated with cortical neurodegeneration. ²⁵ , ²⁶ We observed an increasing trend of FA in pWMHs across different disease groups. These regional differences may reflect disparate pathophysiological mechanisms such as gliosis or perivascular remodeling, ²⁷ , ²⁸ which warrants further investigation. Consistent reductions in Vp across both WMH regions in CAA suggest impaired microvascular integrity, aligning with pathological evidence of microvascular damage due to vascular Aβ deposition. ²⁹ , ³⁰ In contrast, vascular alterations in AD appear to be milder, reflecting differences in the underlying pathophysiological mechanisms of WMH between AD and CAA. CBF and Ktrans showed limited ability to discriminate between diagnostic groups, possibly reflecting multifactorial influences and lower disease specificity.

Distinguishing between AD‐ and cSVD‐related WMHs has remained a challenge due to overlapping radiological patterns. ³¹ , ³² Most previous studies relied on total or qualitative WMH assessments, which lack the spatial resolution to detect region‐specific differences between AD and CAA. ¹² , ³³ This study expands the current understanding of WMH‐associated tissue gradients in cSVD by quantitatively modeling the penumbral transition zone. ¹⁶ We introduced a piece‐wise linear model, rather than a single gradient or polynomial fit, allows for the identification of distinct inner and outer regions of tissue change. The breakpoint represents the point along the spatial profile at which the tissue signal trajectory changes most significantly, separating two distinct linear segments. Biologically, it reflects the spatial transition from the WMH lesion core to the surrounding NAWM and may approximate the extent of tissue damage or remodeling due to underlying pathological processes.

Our spatial analysis revealed significant heterogeneity in the microstructural and microvascular gradients surrounding WMH across different disease groups. Research indicates that cSVD and AD share microstructural and microvascular alterations in the NAWM, which are associated with both the severity and progression of WMHs. ⁸ , ¹⁰ , ²⁵ , ³⁴ However, the spatial distribution differences of microvascular and microstructural changes between CAA and AD in NAWM remain unexplored. The increase in extracellular FWF likely indicates microvascular degeneration, demyelination, and fiber loss, which are key features of cSVD pathology. ⁸ , ¹⁶ , ³⁵ Additionally, the FWF in NAWM varies according to AD diagnosis, the presence of MCI, and the severity of dementia. ³⁴ , ³⁶ Our study further compared the spatial distribution of the FWF between CAA and AD. Compared to MCI and CN, the FWF breakpoint in CAA was more distant. This suggests that diffuse white matter microstructural damage is exacerbated by vascular dysfunction. ³⁷ We also found that CAA showed a broader range of Ktrans in the penumbra with a steeper external slope (Slope2) compared to MCI. In addition, slowing BBB leakage and enhancing metabolic waste clearance may help mitigate WMH expansion. ³⁸ However, due to the smaller sample size in the CAA group, further validation in larger cohorts is needed. Previous studies showed that reduced CBF in the penumbra could predict WMH progression. ¹⁰ In our study, CAA exhibited a more extensive CBF deficit, distinguishing it from AD. These findings further emphasize the role of vascular pathology in CAA. No significant differences were found between CAA and AD in the dWMH region across all microstructural and vascular markers. Overall, microvascular and microstructural changes in CAA occur over a broader range before macroscopic white matter degeneration.

Our study used concentric ROIs and segmented regression to analyze local tissue gradients in the WMH microenvironment. Previous studies have found that the FWF in NAWM can distinguish subtle vascular damage from microstructural changes in white matter in the early stages of AD. ³⁴ Our study showed that in the pWMH region, gradient metrics such as FWF and FA may help distinguish CAA from the AD continuum. In contrast, CBF showed limited discriminative capacity, indicating that perfusion‐based parameters may be less sensitive to CAA‐related changes. The higher AUC values observed in the retest dataset may reflect improved signal contrast due to subtle disease progression, smaller sample size with less variability, or sampling bias inherent in the longitudinal subset. While encouraging, these findings should be interpreted cautiously and require validation in larger external cohorts.

The recognition of CAA‐like WMHs in AD is also critical for AD therapeutic decision‐making. The Aβ‐targeted immunotherapy for AD has demonstrated promising efficacy but is associated with an increased risk of ARIA‐H, particularly in individuals with CAA pathology. ²⁰ , ³⁹ While susceptibility‐weighted imaging is valuable for detecting cerebral microbleeds, its diagnostic utility in CAA is limited to hemorrhagic manifestations and does not capture the microstructural or BBB‐related tissue alterations that may precede overt bleeding. Our approach leverages diffusion and perfusion metrics to quantify perilesional microstructure changes, offering an insight into CAA‐related WMH.

A few limitations should be considered when interpreting the results. First, the sample size of the CAA group was small, which may limit the generalizability of some findings. Second, although our findings provide important cross‐sectional insights, longitudinal data are needed to clarify the temporal relationship and potential causal link between WMH gradient parameters and clinical progression. Third, the spatial resolution of the diffusion and perfusion MRI sequences differed substantially. Although all images were registered to a common space and resampled for ROI extraction, the native resolution may still limit the ability to reliably distinguish concentric perilesional layers at 1‐ to 2‐mm intervals, particularly in perfusion‐based parameters. This potential partial volume effect should be taken into account when interpreting spatial gradients and breakpoints. Lastly, while we validated the results within our center as proof of concept, external validation in larger, independent, multicenter cohorts is necessary to confirm the robustness of these findings.

In conclusion, our study provides insights into the distinct pathophysiological processes underlying AD‐ and CAA‐related white matter injury by introducing a spatially resolved framework for quantification and characterizing WMH gradients.

CONFLICT OF INTEREST STATEMENT

The authors declare no conflicts of interest. Author disclosures are available in the Supporting Information.

CONSENT STATEMENT

All human subjects provided informed consent.

Supporting information

Supporting information

Supporting information