Interaction of amyloid and tau on cortical microstructure in cognitively unimpaired adults

Abstract

INTRODUCTION:

Neurite orientation dispersion and density imaging (NODDI), a multi-compartment diffusion-weighted imaging (DWI) model, may be useful for detecting early cortical microstructural alterations in Alzheimer’s disease (AD) prior to cognitive impairment.

METHODS:

Using neuroimaging (NODDI and T1-weighted MRI) and CSF biomarker data (measured using Elecsys^® CSF immunoassays) from 219 cognitively unimpaired participants, we tested the main and interactive effects of CSF Aβ₄₂/Aβ₄₀ and pTau on cortical NODDI metrics and cortical thickness, controlling for age, sex, and APOE ε4.

RESULTS:

We observed a significant CSF Aβ₄₂/Aβ₄₀ × pTau interaction on cortical neurite density index (NDI), but not orientation dispersion index (ODI) or cortical thickness. The directionality of these interactive effects indicated: 1) among individuals with lower CSF pTau, greater amyloid burden was associated with higher cortical NDI; and 2) individuals with greater amyloid and pTau burden had lower cortical NDI, consistent with cortical neurodegenerative changes.

DISCUSSION:

NDI is a particularly sensitive marker for early cortical changes that occur prior to gross atrophy or development of cognitive impairment.

1. INTRODUCTION

Alzheimer’s disease (AD) is characterized by silent accumulation of amyloid-β (Aβ) and tau pathology that occurs for years prior to development of cognitive impairment [1]. Detecting the earliest neurodegenerative changes during the asymptomatic, preclinical stage of the disease is important not only for understanding disease mechanisms and progression, but also for improving diagnosis, staging, and monitoring response to therapeutic intervention.

By measuring the diffusion of water molecules within tissues, diffusion-weighted imaging (DWI) provides the ability to assess neuronal microstructure in vivo, making it potentially more sensitive than conventional T1-weighted structural imaging for detecting early cortical changes associated with AD pathology [2]. Recent advancements in multi-shell DWI acquisitions and multi-compartment modeling techniques have greatly improved the ability to study cortical microstructure [3]. Neurite orientation dispersion and density imaging (NODDI) [4] is one such multi-compartment technique in which the composite diffusion signal is modeled by three microstructural compartments: isotropic diffusion (i.e. free water), intracellular diffusion (highly restricted diffusion in neurites), and extracellular diffusion (hindered diffusion outside neurites). These subcomponents are used to calculate two quantitative microstructural metrics: 1) neurite density index (NDI), or the volume fraction of intracellular (neurite) diffusion, and 2) orientation dispersion index (ODI), which reflects the degree of neurite coherence. An important aspect of NODDI is the modeling of free water (including cerebrospinal fluid [CSF]), which helps account for partial volume effects and makes it particularly well-suited for investigating cortical microstructure, especially in neurodegenerative conditions.

Previous studies using the NODDI model to investigate cortical microstructural alterations in clinical AD have demonstrated lower NDI and ODI across widespread cortical regions in both young onset and sporadic AD participants [5,6]. Moreover, participants with mild cognitive impairment (MCI) had lower cortical NDI – but not cortical thickness – in several key regions affected early in AD, suggesting that NDI may be a sensitive marker of cortical microstructural alterations that occur prior to measurable cortical atrophy. However, whether cortical NODDI microstructure is associated with AD pathology prior to development of cognitive impairment remains to be determined.

In this study, we tested the relationship between AD pathology as indexed by cerebrospinal fluid (CSF) biomarkers and cortical NODDI microstructure in cognitively unimpaired late middle-aged adults. Specifically, we used whole-brain analyses to test the main and interactive effects of CSF Aβ and phosphorylated tau 181 on both cortical NODDI metrics and cortical thickness, followed up with targeted region of interest (ROI) analyses in an AD signature composite region to more directly compare the effects on cortical microstructure and cortical thickness.

2. MATERIALS AND METHODS

2.1. Participants

We identified 219 cognitively unimpaired individuals from the Wisconsin Alzheimer’s Disease Research Center (ADRC) clinical core (n = 131) and the Wisconsin Registry for Alzheimer’s Prevention (WRAP) study (n = 88) who had undergone both MRI acquisition (multi-shell DWI and T1-weighted) as well as lumbar puncture for CSF biomarker quantification. All participants had previously undergone comprehensive neuropsychological testing to determine cognitive status. ADRC participants underwent the neuropsychological battery in the National Alzheimer’s Coordinating Center (NACC) Uniform Data Set (UDS) Version 3, and WRAP participants underwent a similar comprehensive cognitive battery as previously described [7]. Supplementary Table 1 provides a summary of neuropsychological testing scores for participants in the current study. All ADRC participants are reviewed by a multidisciplinary consensus review committee consisting of physicians, neuropsychologists, and nurse practitioners. WRAP participants are selectively reviewed by the consensus review committee when cognitive abnormalities are detected by algorithm on neuropsychological tests. General exclusion criteria for the ADRC and WRAP studies include any significant neurologic disease (other than AD dementia), history of alcohol/substance dependence, major psychiatric disorders (including untreated major depression), or other significant medical illness. APOE ε4 genotyping procedures have been previously described [8], and participants were categorized as noncarriers (zero ε4 alleles) or carriers (one or two ε4 alleles). The University of Wisconsin Health Sciences Institutional Review Board approved all study procedures and all participants provided written informed consent.

2.2. Lumbar Puncture and Cerebrospinal Fluid Analysis

CSF was collected via lumbar puncture in the morning after a minimum 4hr fast with a Sprotte 25-or 24-gauge spinal needle at the L3/4 or L4/5 interspace using gentle extraction into propylene syringes. CSF (~20 mL) was then combined, gently mixed and centrifuged at 2,000 x g for 10 minutes. Supernatants were frozen in 0.5 mL aliquots in polypropylene tubes and stored at −80 °C.

CSF Aβ₄₀, Aβ₄₂, and phosphorylated tau 181 (pTau) concentrations were measured using Elecsys^® CSF immunoassays as part of the NeuroToolKit [9,10]. CSF amyloid burden was quantified by the Aβ₄₂/Aβ₄₀ ratio (lower CSF levels indicate greater amyloid burden), which shows better correspondence with brain amyloid deposition than CSF Aβ₄₂ alone [11]. Aβ₄₂/Aβ₄₀ ratio values for all participants at our center (regardless of clinical status) are generally lower due to pre-analytical differences in sample collection and processing [12-14]. Cutoff values for CSF biomarkers were developed independently at our center using a larger cohort of individuals with Elecsys^® CSF immunoassay data. CSF Aβ₄₂/Aβ₄₀ cutoffs were derived using receiver operating characteristic (ROC) curve analysis and Youden’s J statistic, with [C-11]Pittsburgh compound B (PiB) PET imaging positivity as standard of comparison [15]. CSF pTau cutoffs were derived by using a positive threshold set at +2SD above the mean of a reference group of 223 CSF Aβ₄₂/Aβ₄₀ negative, unimpaired younger participants (ages 45-60 years). The cutoff for CSF Aβ₄₂/Aβ₄₀ positivity was <0.046, and the cutoff for CSF pTau positivity was >24.8 pg/mL.

2.3. MRI Acquisition and Processing

MRI data were acquired using a General Electric 3T MR750 scanner with a 32-channel head coil. Diffusion-weighted images were acquired using a multi-shell 2D spin-echo echo-planar imaging pulse sequence (6 x b = 0 s/mm², 9 x b = 500 s/mm², 18 x b = 800 s/mm², and 36 x b = 2000 s/mm²; TR/TE = 8575 ms/76.8 ms; 2x2x2 mm³ isotropic spatial resolution; 128 x 128 acquisition matrix). T1-weighted structural images were acquired using a 3D inversion recovery prepared fast spoiled gradient-echo (FSPGR) BRAVO sequence (TI = 450 ms; TR/TE = 8.1 ms/3.2 ms; flip angle = 12°; 1x1x1 mm³ isotropic spatial resolution).

Diffusion-weighted images underwent denoising [16], Gibbs ringing correction [17], and motion and eddy current correction [18]. Fractional anisotropy (FA) maps were fit using Diffusion Imaging in Python (DIPY) [19], and subsequently segmented into white matter (WM) fraction maps using Atropos in ANTs [20]. NODDI parameter maps (NDI, ODI, and isotropic volume fraction [V_ISO]) were fit using Accelerated Microstructure Imaging via Convex Optimization (AMICO) [21] with an optimized intracellular intrinsic parallel diffusivity parameter (d_∥) of 1.1 μm²/ms to improve fit in gray matter [3]. All diffusion images were visually inspected prior to further analyses.

T1-weighted structural images were processed using the Computational Anatomy Toolbox (CAT12, http://www.neuro.uni-jena.de/cat/) in SPM12 (https://www.fil.ion.ucl.ac.uk/spm/). Cortical thickness was estimated using the projection-based thickness (PBT) method [22], followed by topological correction and spherical mapping and registration. Prior to vertex-wise analyses, surface data were resampled into template space (32k mesh resolution; ~2 mm average vertex spacing) and smoothed using the default 15 mm FWHM filter size. All CAT12 segmentations and cortical surface meshes were visually inspected for surface defects and cortical thickness estimation errors prior to further analyses.

2.4. Gray Matter-Based Spatial Statistics (GBSS) Processing

GBSS uses the tract-based spatial statistics (TBSS) framework [23] to allow for gray matter-specific voxel-wise statistical analysis of cortical microstructure [24]. Processing steps for GBSS have been previously described [6,24]. Briefly, for each participant, a gray matter (GM) fraction map was generated by subtracting the WM fraction map (estimated from the participant’s FA map using 2 tissue segmentation in Atropos) and the CSF fraction map (estimated from the participant’s NODDI V_ISO parameter map) from 1. The GM, WM, and CSF fraction maps were then multiplied by respective tissue weightings (CSF = 0, WM = 1, GM = 2) and combined to generate a “pseudo T1-weighted” image. Pseudo T1-weighted images from all participants were then used to create a common voxel space population template via iterative nonlinear registration using the antsMultivariateTemplateConstruction2.sh script in ANTs. The warp fields generated from this step were then used to nonlinearly warp NDI, ODI, and GM fraction images for each participant into the population template space. GM fraction images in population template space were averaged to generate a mean GM image, which was then skeletonized using FSL’s tbss_skeleton tool. Finally, NDI and ODI were projected onto the GM skeleton from the local GM fraction maxima, and the GM skeleton was thresholded to only include voxels with GM fraction > 0.65 in > 70% of participants.

2.5. AD Signature Composite Region Processing

In order to compare effects for both cortical microstructure and cortical thickness within the same brain regions, an AD signature composite ROI was constructed using the Desikan atlas [25] and included bilateral inferior parietal, middle temporal, inferior temporal, precuneus, fusiform, and entorhinal subregions [26]. For cortical microstructure, a T1-weighted atlas-space image was nonlinearly registered to each participant’s pseudo T1 image, and the resulting warp fields were used to transform binarized subregion ROIs into native diffusion space. Subregion ROIs were masked by the participant’s GM fraction map (thresholded at 0.7 and binarized) to reduce the contribution of non-GM voxels within the ROIs. Volume (estimated by number of voxels) and microstructural metrics were extracted within each GM-masked subregion and used to calculate a weighted averaged for the AD signature composite region (weighted by relative subregion volume to account for differences in ROI subregion size). For cortical thickness, mean cortical thickness within each subregion was extracted using ROI tools in CAT12 and averaged to calculate mean cortical thickness within the overall AD composite region.

2.6. Statistical Analysis

We first performed whole-brain analyses to determine the main and interactive effects of CSF Aβ₄₂/Aβ₄₀ and pTau on cortical NODDI metrics and cortical thickness. CSF Aβ₄₂/Aβ₄₀ and pTau values were treated as continuous variables (pTau was log₁₀-transformed due to non-normality), and all statistical models included age, sex, and APOE ε4 as covariates. The main effects linear regression model included Aβ₄₂/Aβ₄₀ and pTau as separate variables, according to the formula:

$$\text{NDI},\text{ODI},\text{or CT}={\beta }_0+{{\beta }_1}^{\ast }\text{Age}+{{\beta }_2}^{\ast }\text{Sex}+{{\beta }_3}^{\ast }APOE\epsilon 4+{{\beta }_4}^{\ast }\mathrm{A}{\beta }_{42}∕\mathrm{A}{\beta }_{40}+{{\beta }_5}^{\ast }\text{pTau}+\epsilon$$ (1)

and the interaction linear regression model included an additional Aβ₄₂/Aβ₄₀ × pTau interaction term, according to the formula:

$$\text{NDI},\text{ODI},\text{or CT}={\beta }_0+{{\beta }_1}^{\ast }\text{Age}+{{\beta }_2}^{\ast }\text{Sex}+{{\beta }_3}^{\ast }APOE\epsilon 4+{{\beta }_4}^{\ast }\mathrm{A}{\beta }_{42}∕\mathrm{A}{\beta }_{40}+{{\beta }_5}^{\ast }\text{pTau}+{{\beta }_6}^{\ast }(\mathrm{A}{\beta }_{42}∕\mathrm{A}{\beta }_{40}\times \text{pTau})+\epsilon$$ (2)

Voxel-wise GBSS analyses were performed on the final skeletonized population-space NDI and ODI images using nonparametric permutation inference (n = 10,000 permutations) with threshold-free cluster enhancement (TFCE) [27] in FSL’s randomise [28]. Vertex-wise cortical thickness analyses were performed using the resampled and smoothed surface data and FSL’s Permutation Analysis of Linear Models (PALM) tool (n = 10,000 permutations) with TFCE. Resulting statistical maps were family-wise error (FWE)-corrected at P_FWE < 0.05 and displayed as surfaces using Surf Ice (https://www.nitrc.org/projects/surfice/).

In order to determine the directionality of Aβ₄₂/Aβ₄₀ × pTau interaction effects, significant voxels from GBSS statistical maps were de-projected from the mean GM skeleton and subsequently warped back into each participant’s native diffusion space. In native diffusion space, mean parameter values (averaged across all voxels with a significant Aβ₄₂/Aβ₄₀ × pTau interaction effect) were then extracted from NODDI parameter maps for each participant. These values were used for simple slopes analyses using the interactions package (version 1.1.1) in R to examine continuous by continuous interactions. In simple slopes analyses, conditional slopes are estimated to test whether the relationship between predictor and outcome variables is significant at a given level of the moderator variable. Separate tests evaluated both Aβ₄₂/Aβ₄₀ and pTau as moderators. To facilitate interpretation of simple slope results, two moderator levels were chosen based on CSF biomarker cutoff groupings (e.g. Aβ− and Aβ+; pTau− and pTau+). Specifically, simple slopes analyses tested: 1) the relationship between Aβ₄₂/Aβ₄₀ and NODDI metrics at mean pTau levels of both the pTau− and pTau+ groups; and 2) the relationship between pTau and NODDI metrics at mean Aβ₄₂/Aβ₄₀ levels of both the Aβ+ and Aβ− groups. Finally, participants were classified into four biomarker groups (A−/T−, A−/T+, A+/T−, A+/T+) based on dichotomous CSF Aβ₄₂/Aβ₄₀ and pTau cutoffs, and differences in NODDI metrics between groups were tested using ANCOVA models (age, sex, and APOE ε4 included as covariates) followed by pairwise post hoc tests with false discovery rate (FDR) P value correction for multiple comparisons.

To more directly compare microstructure and cortical thickness, we tested the main and interactive effects of Aβ₄₂/Aβ₄₀ and pTau on NODDI metrics or cortical thickness within the AD signature composite region [26]. Linear regression models were identical to those used in whole-brain analyses, and once again, scatter plots and simple slopes analyses were used to determine the directionality of the continuous by continuous Aβ₄₂/Aβ₄₀ × pTau interaction effects. Finally, we examined the regional patterns of Aβ₄₂/Aβ₄₀ × pTau interaction effects on NODDI microstructure by performing simple slopes analyses on each AD composite subregion separately (using FDR correction for multiple comparisons). All statistical analyses were performed in R (version 3.6.3) using the interactions (version 1.1.1) package for simple slopes analysis, ggplot2 (version 3.3.0) for scatterplots and data visualization, and ggseg (version 1.5.5) for brain atlas visualization.

3. RESULTS

3.1. Participant Characteristics

Participant characteristics are presented in Table 1. All participants were cognitively unimpaired based on comprehensive neuropsychological testing (see Supplemental Table 1 for summary of cognitive testing results). Based on CSF biomarker cutoffs, 8% (18/212) of participants were CSF amyloid positive while an additional 10% (22/212) were both CSF amyloid and tau positive, which is consistent with larger samples of cognitively unimpaired participants from our center [14].

Table 1.

Participant characteristics

Characteristic	Value
N	219
Age at MRI, years (mean ± SD)	66.8 ± 7.5
Age at LP, years (mean ± SD)	64.9 ± 7.5
Age difference between MRI and LP, years (median [IQR])	1.21 [0-4.03]
Sex, % female (n)	58.9% (129/219)
APOE ε4 allele, % positive (n)	34.9% (76/219)
Primary race/ethnicity, n (Caucasian/African American/Asian/Native American/Other)	213 / 3 / 1 / 1 / 1
Education, years (mean ± SD)	16.6 ± 2.3
Montreal Cognitive Assessment (MoCA) (ADRC participants, n = 131)	27.1 ± 2.3
Mini-Mental State Examination (MMSE) (WRAP participants, n = 88)	29.4 ± 0.9
CSF values
Aβ40, ng/mL (mean ± SD)	14.8 ± 4.6
Aβ42, pg/mL (mean ± SD)	908 ± 389
Aβ42/Aβ40 (mean ± SD)	0.06 ± 0.02
pTau181, pg/mL (median [IQR])	16.9 [13.3-22.1]
CSF biomarker cutoff groups, n (% of sample)
A−/T−	166 (75.8%)
A−/T+	13 (5.9%)
A+/T−	18 (8.2%)
A+/T+	22 (10.0%)

3.2. Whole-Brain Voxel- and Vertex-Wise Analyses

Whole-brain analyses demonstrated a significant Aβ₄₂/Aβ₄₀ × pTau interaction on NDI that was distributed predominantly throughout temporal, parietal, and medial frontal cortical regions (Fig. 1A; Supplementary Fig. 1). This finding was most prominent in the right inferior parietal region (including angular and supramarginal gyrus), bilateral posterior middle temporal region, temporal pole, entorhinal region, and posterior cingulate. There were no subcortical gray matter regions with significant Aβ₄₂/Aβ₄₀ × pTau interaction on NDI. There was no significant Aβ₄₂/Aβ₄₀ × pTau interaction on ODI or cortical thickness, and there were no significant main effects of Aβ₄₂/Aβ₄₀ or pTau on NDI, ODI, or cortical thickness.

Figure 1. Gray matter-based spatial statistics (GBSS) results showing a significant Aβ₄₂/Aβ₄₀ × pTau interaction on cortical neurite density index (NDI).

(a) Overlay of significant (FWE-corrected P < 0.05) GBSS results on inflated surface projection. (b-c) Mean NDI values from significant GBSS voxels were extracted in native diffusion space for each participant and plotted as scatter plots. Points are sized and colored by the moderator variable, and fit lines represent conditional slopes from simple slopes analyses. (b) For Aβ × pTau status, there was a significant negative conditional slope of Aβ₄₂/Aβ₄₀ on NDI at the pTau− group and a significant positive conditional slope at the pTau+ group mean. (c) For pTau × Aβ status, there was a significant positive conditional slope of pTau on NDI at the Aβ− group and a significant negative conditional slope at the Aβ+ group mean. (d) NDI values were collapsed across biomarker groups and ANCOVA models showed that the A+/T− group had higher NDI than the A−/T− group, and the A+/T+ group had lower NDI than all other groups. *P_FDR < 0.05, **P_FDR < 0.01

NDI values were extracted from significant GBSS voxels in native diffusion space and used in linear regression models followed by simple slopes analysis in order to investigate the directionality of the continuous by continuous interaction effect of Aβ₄₂/Aβ₄₀ × pTau. As expected, there was a significant Aβ₄₂/Aβ₄₀ × pTau interaction on NDI values from significant GBSS voxels (β = 1.51, 95% CI 1.03 to 2.00, P < 0.0001). Using the mean values of pTau in the pTau− group (15.9 pg/mL) and pTau+ group (31.2 pg/mL), there was a significant negative conditional slope of Aβ₄₂/Aβ₄₀ on NDI at the pTau− group mean (β = −0.24, 95% CI −0.35 to −0.13, P < 0.0001; Fig. 1B) and a significant positive conditional slope at the pTau+ group mean (β = 0.22, 95% CI 0.11 to 0.33, P < 0.0001; Fig. 1B). Similarly, using the mean values of Aβ₄₂/Aβ₄₀ in the Aβ− group (0.068) and Aβ+ group (0.033), there was a significant positive conditional slope of pTau on NDI at the Aβ− group mean (β = 0.011, 95% CI 0.002 to 0.021, P = 0.023; Fig. 1C) and a significant negative conditional slope at the Aβ+ group mean (β = −0.041, 95% CI −0.056 to −0.027, P < 0.0001; Fig. 1C).

Finally, after collapsing participants into four biomarker groups based on dichotomous Aβ₄₂/Aβ₄₀ and pTau cutoffs, ANCOVA models using NDI values from significant GBSS voxels demonstrated significant differences in cortical NDI between biomarker groups (F₂₁₂ = 3.10, P = 0.0062; Fig. 1D). Specifically, post hoc comparisons indicated that while A−/T− and A−/T+ groups did not differ, the A+/T− group trended towards higher NDI than the A−/T− group (P_FDR = 0.074), and the A+/T+ group had lower NDI than all other groups (A+/T+ vs A−/T−: P_FDR = 0.032; A+/T+ vs A−/T+: P_FDR = 0.032; A+/T+ vs A+/T−: P_FDR = 0.0027).

3.3. AD Composite Region Analyses

Within the AD composite region, there was a significant Aβ₄₂/Aβ₄₀ × pTau interaction on NDI (β = 0.84, 95% CI 0.36 to 1.32, P = 0.00071; Fig. 2A&B), but not cortical thickness (β = 2.64, 95% CI −5.1 to 10.4, P = 0.51; Fig. 2C&D). Specifically, examining Aβ × pTau status using simple slopes analysis indicated that there was a significant negative conditional slope of Aβ₄₂/Aβ₄₀ on NDI at the pTau− group mean (β = −0.14, 95% CI −0.25 to −0.03, P = 0.014; Fig. 2A) and a significant positive conditional slope at the pTau+ group mean (β = 0.12, 95% CI 0.005 to 0.23, P = 0.039; Fig. 2A). Examining pTau × Aβ status using simple slopes analysis indicated that while the conditional slope of pTau on NDI at the Aβ− group mean was not significant (β = 0.005, 95% CI −0.004 to 0.015, P = 0.28; Fig. 2B), there was a significant negative conditional slope of pTau on NDI at the Aβ+ group mean (β = −0.024, 95% CI −0.039 to −0.009, P = 0.0017; Fig. 2B). Similar results were observed in simpler interaction models where Aβ₄₂/Aβ₄₀ was kept continuous and pTau was dichotomized, and where pTau was kept continuous and Aβ₄₂/Aβ₄₀ was dichotomized (see Supplementary material). There were no main effects of Aβ₄₂/Aβ₄₀ or pTau on NDI (Aβ₄₂/Aβ₄₀: β = −0.010, 95% CI −0.095 to 0.075, P = 0.82; pTau: β = −0.003, 95% CI −0.011 to 0.006, P = 0.54) or cortical thickness (Aβ₄₂/Aβ₄₀: β = −0.06 , 95% CI −1.41 to 1.29, P = 0.93; pTau: β = 0.02, 95% CI −0.12 to 0.16, P = 0.78) within the AD composite region. Additionally, there were no differences in NDI, ODI, or cortical thickness between APOE ε4-positive and ε4-negative individuals.

Figure 2. Interaction of Aβ₄₂/Aβ₄₀ × pTau for neurite density index (NDI) and cortical thickness values extracted from AD composite region.

Mean NDI or cortical thickness were extracted from a bilateral AD composite region and plotted as scatter plots. Points are sized and colored by the moderator variable, and fit lines represent conditional slopes from simple slopes analyses. (a) For Aβ × pTau status, there was a significant negative conditional slope of Aβ₄₂/Aβ₄₀ on NDI at the pTau− group mean and a significant positive conditional slope at the pTau+ group mean. (b) For pTau × Aβ status, there was a significant negative conditional slope of pTau on NDI at the Aβ+ group mean, but no significant conditional slope at the Aβ− group mean. (c-d) There was no significant Aβ₄₂/Aβ₄₀ × pTau interaction on cortical thickness.

Overall, the GBSS and AD region results indicate an Aβ₄₂/Aβ₄₀ × pTau interaction effect on cortical microstructure in which lower levels of CSF Aβ₄₂/Aβ₄₀ (i.e. greater amyloid burden) in pTau− individuals were associated with higher cortical NDI. Additionally, lower levels of CSF Aβ₄₂/Aβ₄₀ in pTau+ individuals, as well as higher levels of CSF pTau in Aβ+ individuals, were both associated with lower cortical NDI.

3.4. AD Composite Subregion Analyses

We performed simple slopes analyses on the AD composite subregions in order to investigate the regional patterns of Aβ₄₂/Aβ₄₀ × pTau interaction effects on NDI. For Aβ × pTau status, all subregions had negative conditional slopes of Aβ₄₂/Aβ₄₀ on NDI at the pTau− group mean and positive conditional slopes of Aβ₄₂/Aβ₄₀ on NDI at the pTau+ group mean (Fig. 3A). At the pTau− group mean, conditional slopes were significant (P < 0.05, FDR-corrected) in right inferior parietal, bilateral inferior temporal, and bilateral fusiform, indicating that greater amyloid burden was associated with higher NDI in these subregions for individuals who were pTau−. At the pTau+ group mean, there were no subregions with significant conditional slopes that survived FDR correction.

Figure 3. Regional patterns of Aβ₄₂/Aβ₄₀ × pTau interaction effects on NDI in the AD composite subregions.

Estimates of conditional slopes (with 95% confidence intervals) for each subregion and hemisphere. • P < 0.05 (uncorrected), *P < 0.05 (FDR corrected). (a) For Aβ × pTau status, subregions showed negative conditional slopes of Aβ₄₂/Aβ₄₀ on NDI at the pTau− group mean and positive conditional slopes of Aβ₄₂/Aβ₄₀ on NDI at the pTau+ group mean, indicating that greater amyloid burden was associated with higher NDI in pTau− individuals and lower NDI in pTau+ individuals. (b) For pTau × Aβ status, all subregions (except left inferior parietal and bilateral entorhinal) showed negative conditional slopes of pTau on NDI at the Aβ+ group mean, indicating that higher CSF pTau levels were associated with lower NDI in Aβ+ individuals.

For pTau × Aβ status, conditional slopes of pTau on NDI at the Aβ− group mean were close to zero and not significant for all subregions (Fig. 3B). However, at the Aβ+ group mean, nearly all subregions had significant negative conditional slopes of pTau on NDI (with the exception of left inferior parietal and the bilateral entorhinal subregion), indicating that higher levels of CSF pTau were associated with lower NDI in nearly every AD composite subregion for Aβ+ individuals.

4. DISCUSSION

Hypothetical models describing the temporal ordering of biomarker abnormalities along the continuum of AD propose sequential detectable changes in amyloid and tau pathophysiology, followed by neurodegeneration [1]. While conventional T1-weighted MRI-derived macrostructural measures (e.g. cortical thickness or gray matter volume) are often used as markers of neurodegenerative changes, recent work has proposed that microstructural features of neurodegeneration should also be considered on this continuum and may be detectable prior to macrostructural changes [2]. Here, we examined the extent to which AD-associated neurodegeneration is detectable in the preclinical asymptomatic phase of the disease using multi-shell DWI, and found that indeed, microstructural alterations, especially altered neurite density, were present among cognitively unimpaired individuals with amyloid and tau pathology. Moreover, there were no changes observed when using the T1-weighted derived measure of cortical thickness, suggesting that DWI microstructural metrics are more sensitive than macrostructural features for detecting early cortical neurodegenerative changes in preclinical AD.

Previous studies have reported alterations in cortical NODDI microstructure for individuals with both clinical AD dementia and mild cognitive impairment [5,6]. Here, we find that altered cortical neurite density is also present in the preclinical, asymptomatic phase of the disease. Specifically, we observed that greater amyloid burden (indexed by lower CSF Aβ₄₂/Aβ₄₀ levels) was associated with higher cortical NDI, but only when CSF pTau was low, suggesting an earlier disease stage. Secondly, we observed that greater amyloid burden accompanied by higher tau pathology was associated with lower NDI, suggesting loss of cortical neurites with worsening AD pathology. These findings were observed in both whole-brain voxel-wise GBSS analyses and in follow-up targeted analyses using NDI values extracted in native diffusion space from AD-specific cortical regions. Together, these findings suggest a non-monotonic relationship between cortical NDI and progression of AD pathology, where early amyloid deposition is associated with greater restricted intracellular diffusion, and subsequent accumulation of pathology is associated with less restricted intracellular diffusion, suggestive of neurite loss.

Several studies indicate that amyloid alone is often insufficient to induce clinical manifestation of AD [29,30], and recent studies have demonstrated that both amyloid and tau are necessary for longitudinal cognitive decline in cognitively unimpaired individuals [31-33]. Investigating the extent to which amyloid and tau pathology interact to impact alterations in brain structure in asymptomatic individuals is critical to improving the timing of clinical interventions, as well as improving the understanding of preclinical AD progression. Due to its sensitivity to disruption in tissue barriers, DWI has been proposed as a more sensitive marker of cortical changes in AD than conventional T1-weighted imaging [2]. The most commonly used DWI technique is diffusion tensor imaging (DTI), which models the diffusion signal as a simple ellipsoid tensor [34]. A limited number of previous studies have used DTI (and specifically the metric of mean diffusivity [MD]) to investigate cortical microstructure in preclinical familial [35] and preclinical sporadic AD [36]. These studies were instrumental in demonstrating a biphasic trajectory of cortical microstructure in asymptomatic individuals in which cortical MD was lower in A+/T− individuals and higher in A+/T+ individuals relative to A−/T− individuals. In the current study, we used the multi-compartment NODDI model and observed consistent results, whereby intracellular diffusion was highly restricted in the presence of amyloid, followed by less restricted intracellular diffusion among individuals with greater disease severity.

In expanding on previous research on cortical microstructure in preclinical AD, a major advantage of the present study is the use of NODDI as a more complex and biologically relevant multi-compartment model of the diffusion signal. NODDI metrics may capture microstructural properties and quantify neuronal cytoarchitecture better than DTI metrics, which fail to capture several aspects of complex tissue microstructure [3]. Furthermore, by directly modeling isotropic diffusion (i.e. free water), NODDI helps account for partial volume effects, which are of particular concern in neurodegenerative diseases where cortical atrophy can cause contamination of the diffusion signal from adjacent CSF and result in overestimation of MD changes [2,37]. While previous studies have shown that cortical thickness and MD are highly correlated across almost the entire cerebral cortex [36,38], in the current study there was no significant association between cortical thickness and NDI in the AD composite region (Pearson’s r = 0.10, 95% CI −0.03 to 0.23, P = 0.13). While there may be regional variation in the relationship between cortical microstructure and thickness, our findings suggests that cortical NDI quantifies meaningful microstructural changes associated with AD pathology that are independent of alterations in cortical thickness and less influenced by partial voluming.

Microstructural alterations underlying higher or lower NDI may reflect multiple pathological processes occurring in AD. Previous studies in animal models of AD indicate that both Aβ deposition and neuroinflammation are associated with higher NDI [39] and other multi-shell DWI markers of highly restricted diffusion [40,41]. Additionally, higher NDI may reflect neuronal hypertrophy in response to amyloid deposition [42-44]. On the other end of the spectrum, lower NDI likely reflects loss of cellular barriers, reduction in myelinated axon density [3,45,46], and loss of synapses and dendrites [47,48], all of which occur with accumulation of AD pathology. Thus, in context of our current findings, changes in NDI likely reflect different stages of pathological processes that occur during the asymptomatic, preclinical phase of the disease. Subregion analyses demonstrated that differential changes in NDI were most pronounced in bilateral fusiform and inferior temporal subregions, which suggests that microstructural alterations in these regions may be more sensitive to accumulation of pathology. Histopathological correlation is required to further characterize these associations.

In the current study, while AD pathology was associated with altered cortical NDI, we did not observe significant main or interactive effects of CSF Aβ₄₂/Aβ₄₀ and pTau on cortical thickness or ODI. While the question of whether amyloid alone has an impact on brain structure has been controversial [30], several previous studies have shown a synergistic effect of amyloid and tau on decreased cortical thickness in preclinical AD [36,49-52]. It is worth noting that the individuals included in our current study were largely younger (66.8±7.5 years at DWI) compared to other preclinical AD cohort studies (typically early to mid 70s), which may explain the lack of observed effects on cortical thickness. Therefore, detecting alterations in NDI but not cortical thickness in our younger cohort highlights the utility of NODDI metrics as sensitive markers of early cortical changes associated with AD pathology. Additionally, previous studies have reported lower cortical ODI in individuals with clinical AD dementia, but not individuals with MCI [5,6]. The current study provides further evidence that reductions in cortical ODI, which may reflect complexity of dendritic arborization [3,46], are observed in the later stages of disease severity when overt clinical dementia is present. Finally, we found no effect of APOE ε4 on cortical microstructure in cognitively unimpaired APOE ε4 carriers vs non-carriers, consistent with previous studies suggesting APOE ε4 status influences amyloid accumulation rather than gray matter structure and metabolism [53].

Several limitations of the current study should be noted. First, we acknowledge that an inherent limitation of using DWI to study cortical microstructure is the voxel resolution of the DWI scans (2 × 2 × 2 mm³) relative to the thickness of the cortex (1.5–5 mm) [2]. Thus, there is the risk of partial volume effects in thinner areas of the cortex where the diffusion signal may be contaminated from underlying white matter or from adjacent CSF. While the processing steps in GBSS are designed to remove voxels that do not include sufficient gray matter (including in thinner areas of cortex), we recognize that these steps do not completely eliminate the risk of partial volume effects. Additionally, while the multi-compartment NODDI model specifically models free water diffusion (making it particularly advantageous for reducing CSF partial volume effects), cortical NODDI metrics in thinner areas of the cortex may still be partially influenced by imperfect suppression of CSF signal. Recent advances in DWI acquisition, including accelerated multi-slice imaging and 7T MRI sequences have allowed for high-quality acquisition of DWI scans with voxel resolution approaching 1 × 1 × 1 mm³ [54,55]. Additionally, more sophisticated DWI processing techniques such as multi-tissue constrained spherical deconvolution [56] may be able to provide better CSF, WM, and GM tissue parcellation directly from DWI data without requiring registration to anatomical T1 or T2 images. Future studies employing these techniques may reduce the risk of partial volume effects and are needed in order to continue advancing our understanding of cortical microstructure. A second limitation is that while the study included a large sample of well-characterized cognitively unimpaired participants, the analyses were cross-sectional. Longitudinal studies are needed to determine how AD pathology impacts changes in cortical NODDI microstructure over time, both in preclinical and clinical populations. Additionally, future work will focus on the relationship between cortical microstructural changes and development of cognitive impairment, especially in individuals who are CSF AD biomarker positive yet cognitively unimpaired. Third, while CSF Aβ₄₂/Aβ₄₀ and pTau provide good sensitivity and dynamic range, especially in asymptomatic individuals, they inherently do not provide regional information regarding accumulation of pathology. In particular, CSF pTau may not primarily reflect accumulation of tangle pathology in our preclinical population, but more likely captures increased tau phosphorylation and secretion as a neuronal response to amyloid accumulation [57,58]. Future studies combining PET imaging biomarkers of amyloid and tau with NODDI are needed to better characterize the regional relationships between amyloid deposition, tau accumulation, and cortical NODDI microstructure.

Conclusion

In combination with previous studies of NODDI-derived measures of microstructure in symptomatic AD, the current results demonstrate that NODDI is a powerful tool for characterizing cortical microstructural alterations along the continuum of AD. Specifically, in preclinical AD, initial amyloid accumulation is associated with higher NDI, while subsequent further accumulation of amyloid and tau pathology is associated with lower NDI. In the symptomatic phase of the disease, several key early AD cortical regions show lower NDI in individuals with MCI, while both NDI and ODI are lower across widespread cortical areas in individuals with clinical AD dementia [5,6]. Notably, changes in NDI seem to occur prior to reliably detectable changes in cortical thickness. In the context of the recently proposed AT(N) Alzheimer’s research framework [59], these findings suggest that cortical NODDI metrics provide unique insights into the underlying cytoarchitectural changes occurring throughout disease progression, and may be particularly informative biomarkers of neurodegenerative changes (or “N”). Future work investigating how NODDI metrics are related to cognitive decline or predict conversion to MCI or clinical dementia will be vital for further characterizing cortical microstructural alterations in AD.