Background
By the time of symptom onset in Alzheimer’s disease (AD), characteristic neuroanatomical changes have already begun to manifest. Although cognitive impairments are just emerging, cortical atrophy and white matter degeneration are detectable using magnetic resonance imaging (MRI) [
1]. By the time these structural changes appear, the underlying neuropathology may render interventions to halt disease progression ineffective. Development of noninvasive tools to assess neural microstructure is critical to better characterize the earliest neurodegenerative events in AD, which in turn may permit the timely detection of incipient cognitive impairment and more effective intervention.
The earliest neuronal markers of AD include widespread amyloid-β (Aβ) deposits, and those that correspond most strongly with cognitive deficits appear in the medial temporal lobe of the brain, including neurofibrillary tangles, synapse loss, and neuronal death in the entorhinal cortex [
2,
3]. Synapse loss, tau pathology, and cell death, together with white matter damage, spread throughout the limbic system and eventually to more extensive neocortical regions [
4,
5]. Gray matter atrophy that mirrors progressive stages of this pathological cascade has demonstrated efficacy as an in-vivo marker of early AD. Although structural changes are widespread even at early disease stages [
6], those in the medial temporal lobe, including entorhinal cortex and hippocampal atrophy, have the strongest associations with clinical and cognitive metrics [
1,
7‐
11].
Diffusion MRI, of which the most widely used approach is diffusion tensor imaging (DTI), is based on the Brownian motion of water diffusion within biological tissue. DTI allows evaluation of neural microstructure that is complementary to standard morphometric MRI measures. Common DTI metrics include fractional anisotropy (FA; the magnitude of directed diffusion) and mean diffusivity (MD), which depend upon cellular barriers to water diffusion such as myelination, neuronal count, or number or density of neurites [
12]. DTI studies have identified white matter changes co-occurring with or preceding gray matter atrophy in mild cognitive impairment (MCI) and AD [
13]. White matter microstructure in the uncinate, superior longitudinal fasciculus, and fornix is altered in MCI and AD [
14,
15] and predicts memory decline and progression from MCI to AD [
16,
17]. Widespread white matter abnormalities in MCI and AD are at least partially independent of, and may precede, gray matter changes [
18‐
20]. Within gray matter, diffusion imaging-based microstructural measures may be more sensitive than morphometry to disease onset and early neuropathology [
21‐
23]. Diffusion neuroimaging may therefore have value for detecting preclinical neuropathological microstructural changes. However, studies have reported conflicting associations between FA and brain amyloid burden [
24,
25], which may stem from the diversity of biological factors contributing to the aggregate diffusion tensor.
Restriction spectrum imaging (RSI) is a diffusion MRI method that enables measurement of microstructural features undetectable by conventional diffusion imaging techniques, and thus may permit earlier, perhaps presymptomatic, detection of incipient disease. Traditional DTI metrics inform about voxel-level features, but are blind to underlying subvoxel complexities, including crossing or bending fibers [
26]. RSI resolves these properties by using multidirection, high b-value diffusion imaging to measure diffusion orientation and length scale [
27]. RSI can consequently account for within-voxel crossing fibers and separate volume fractions of restricted, hindered, and free water diffusion [
27], and may be less susceptible to the effects of edema or partial voluming than DTI. Histological examination in rodents has determined that the restricted volume fraction predominantly reflects intracellular diffusion within axons and dendrites, and is thus a valuable tool for probing gray and white matter neurite density (ND) [
27]. RSI is clinically useful for tumor detection [
28], and has characterized gray matter organization in autism [
29] and white matter pathology in epilepsy [
30].
This study examined the sensitivity of RSI metrics of ND and isotropic free water diffusion (IF) to memory impairment and disease status in MCI and AD. ND combines the volume fraction of mean restricted diffusion with the volume fraction of oriented diffusion that is highly restricted perpendicular to the direction of diffusion and is not attenuated by crossing fibers. ND therefore yields a combined measure of all restricted diffusion, which is likely dominated by neurites [
27]. Although ND is strongly related to FA, FA is unable to separate restricted and hindered compartments or to account for crossing fibers, and can be reduced by partial voluming if the aggregate diffusion is isotropic. IF measures the volume fraction of isotropic free water diffusion, reflecting contributions from cerebrospinal fluid (CSF) and excluding hindered and restricted diffusion components. In comparison, MD measures average diffusion from all compartments.
ND and IF are expected to be sensitive to microstructural neural changes in MCI and AD, including white matter damage due to axonal degeneration or demyelination, and gray matter changes associated with atrophy or expansion of the CSF space. We therefore hypothesized that ND would be reduced and IF increased with greater dementia severity, episodic memory impairment, and Aβ burden, and that these metrics would accurately discriminate cognitively impaired patients from healthy controls (HC). For validation, RSI measures were compared against conventional DTI metrics.
Methods
Participants
Participants were recruited from the University of California, San Diego (UCSD) Shiley-Marcos Alzheimer’s Disease Research Center (ADRC). Participants completed standardized clinical evaluation through the ADRC Clinical Core, reviewed by two senior neurologists to provide a consensus diagnosis. AD diagnosis was based on NINCDS-ADRDA clinical criteria [
31], and amnestic or multidomain MCI diagnosis was determined according to criteria outlined by Petersen et al. [
32]. Exclusion criteria included a Mini-Mental State Examination score <16 indicating severe dementia, safety contraindications for MRI, uncorrected vision or hearing loss, significant illness, substance abuse, or major psychiatric or neurologic illness. HC were recruited from the ADRC and community. In addition to the above criteria, HC were excluded if they were taking psychotropic or cognitive enhancing medications, or demonstrated impairment on the Mattis Dementia Rating Scale (DRS) or Clinical Dementia Rating tests. After excluding three participants for poor diffusion imaging data quality, the final sample (
n = 56; 30 women) included 31 HC, 12 participants with MCI, and 13 with AD, aged 63–93 years. CSF was obtained from a subset (68%; 24 HC, 7 MCI, 7 AD) of participants.
Study procedures were approved by the UCSD human subjects review board and participants provided informed, written consent prior to participation. Surrogate consent was provided for participants with advanced cognitive impairment.
Neuropsychological assessment
The neuropsychological test battery, described previously [
33], was administered by a trained examiner in a quiet room. Measures were selected for analysis based on their sensitivity to functional and memory impairments in AD. The DRS [
34] assesses the nature and severity of dementia. The Functional Activities Questionnaire (FAQ) [
35] assesses daily living activities. The WMS-R Logical Memory subtest [
36] requires participants to report details of a passage, immediately and after delay. The California Verbal Learning Test (CVLT) [
37] assesses the number of correctly recalled items from a list of 16 categorized words; immediate and delayed free recall were analyzed (CVLT-SFR and CVLT-LFR). The Consortium to Establish a Registry for Alzheimer’s Disease delayed recall (CERAD-DR) score [
38] is another measure of verbal memory that tests delayed recall of a 10-item word list. The American National Reading Test Verbal IQ (ANART-VIQ) [
39] was used as an estimate of premorbid ability.
Imaging data acquisition and processing
MRI data acquisition was performed at the UCSD Center for functional MRI on a 3.0 tesla Discovery 750 scanner (GE Healthcare, Milwaukee, WI, USA) with an eight-channel phased array head coil. The MRI protocol included a three-plane localizer, a sagittal 3D fast spoiled gradient echo T1-weighted volume optimized for maximum gray/white matter contrast (TE = 3.2 ms, TR = 8.1 ms, inversion time = 600 ms, flip angle = 8°, FOV = 24 cm, frequency = 256, phase = 192, voxel size = 1 × 1 × 1.2 mm, scan time 8:27), and an axial 2D single-shot pulsed-field gradient spin-echo echo-planar imaging sequence (45-directions, b-values = 0, 500, 1500, 4000 s/mm2 and 1, 6, 6, 15 unique gradient directions for each b-value, respectively; TE = 80.6 ms, TR = 8 s, frequency = 96, phase = 96, voxel size = 1.875 × 1.875 × 2.5 mm, scan time 6:34).
Data were processed using an automated FreeSurfer-based processing stream (
http://surfer.nmr.mgh.harvard.edu) with additional tools developed at the UCSD Multimodal Imaging Laboratory. Images were visually inspected for quality, and data containing motion or other artifacts were excluded from analysis. RSI data were corrected for motion and eddy current distortions [
40], spatial and intensity distortions [
41], and distortions caused by gradient nonlinearities [
42]. Images were automatically registered and rigidly resampled into standard orientation, based on registration to T
1-weighted structural images [
43]. White matter tracts were labeled using a probabilistic atlas (AtlasTrack) [
44] that combines information about fiber tract location and orientation to estimate the a posteriori probability that a voxel belongs to a tract of interest. To minimize partial volume effects, voxels containing primarily gray matter or CSF were excluded from the analysis of white matter tracts [
45]. To correct for cortical surface partial volume effects, each voxel was assigned a volume fraction from 0–1 according to its proportion of gray or white matter. A weighting factor for each voxel was computed using Tukey’s bisquare weight function [
46], setting volume fractions less than 0.5 to 0 and those above 0.5 to a weight between 0–1, to generate gray and white matter volume fraction maps. Gray matter, white matter, and CSF boundaries were delineated and cortical regions of interest were defined according to the Desikan-Killiany atlas [
47]. DTI measures of FA and MD, and RSI measures of ND and IF were calculated within fibers and regions of interest. DTI measures were computed from all shells of the RSI acquisition using a nonlinear fitting procedure. Analysis of an independent dataset showed better correspondence to DTI measures derived from standard DTI data when using this method than a log transform followed by a linear fit.
CSF collection and measurement
Lumbar puncture was performed by a board-certified neurologist, using a Sprotte atraumatic 24-gauge needle, between 8 am and 11 am after the participant had fasted overnight. Two milliliters of CSF were sent to a local laboratory for measurement of cell count, total protein, and glucose. The remainder (typically 15–20 ml) of CSF was gently mixed, centrifuged in a polypropylene conical tube at 1500 rpm for 10 min, then aliquotted into Sarstedt 0.5-ml cryotubes, snap-frozen immediately, and stored at –80°C until assayed. Levels of Aβ40 and Aβ42 were measured using mass spectrometry (Quest Diagnostics). CSF samples with gross blood contamination or with red blood cell counts >10/ml were not used.
Data analysis
To minimize the number of fibers examined, analyses focused on tracts with connectivity to the temporal lobe that have previously demonstrated altered diffusion signal in MCI or AD [
14,
15,
48]. FA and ND were measured in selected tracts, including fornix, parahippocampal cingulum, uncinate fasciculus, inferior longitudinal fasciculus (ILF), inferior fronto-occipital fasciculus (IFOF), and arcuate fasciculus. MD and IF of hippocampus and entorhinal cortex were assessed because of the critical role of these regions in memory and their vulnerability to degenerative changes in early AD.
Because there were no significant interactions between hemisphere and participant group for any RSI measure (all
p > 0.01), values were averaged across hemispheres. Associations between diffusion imaging metrics and memory were assessed with partial correlations. Group differences in cognitive test scores, neuroimaging measures, and Aβ were evaluated using univariate general linear modeling (GLM). Post-hoc pair-wise group comparisons were adjusted for multiple comparisons using Bonferroni correction. Linear regression was conducted to identify neuroimaging metrics that predict cognitive function. To minimize the number of candidate variables and to allow comparison of RSI and DTI models, regression analyses were first performed separately for ND and IF, and for FA and MD. Significant predictors from these models were input as candidate variables into the final combined regression model for each neuropsychological measure. Linear discriminant analysis (LDA) was performed to distinguish HC from MCI/AD, and cross-validated classification accuracies were computed. LDAs were run separately for RSI metrics and for DTI metrics. Measures selected from these preliminary models were input into the combined LDA. Not all data met assumptions of normality and equal group covariances; however, LDA has been shown to be robust to data distribution and covariance violations [
49]. Area under the receiver operating characteristic curve (AUC) was computed for each classifier, and AUCs were statistically compared [
50]. RSI regression and LDA models were repeated with the inclusion of Aβ42 levels. Pearson’s correlations were calculated between memory and discriminant scores, and between Aβ levels and diffusion imaging metrics or memory scores. Significant differences between correlations were tested using Fisher r-to-z transformations.
For evaluation of whole brain group differences, voxel-based analysis was performed using the Advanced Normalization Tools (ANTS)-Groupwise processing pipeline, a modified Tract-Based Spatial Statistics (TBSS) processing pipeline (
http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/TBSS) which has improved algorithm accuracy and superior registration compared to standard TBSS [
51,
52]. T
1-weighted images were iteratively registered to form a groupwise map using the ANTS-Symmetric Normalization ver-1.9.4 algorithm [
53]. Two-sample
t tests (5000 general linear model permutations per contrast;
http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/Randomise/UserGuide) were performed on voxel-wise white matter ND and gray matter IF, contrasting MCI versus HC and AD versus HC. Voxels showing significant group differences (
p < 0.01, threshold-free cluster enhancement with family-wise error (FWE) correction) were overlaid on the groupwise structural map from all participants in each contrast.
Partial correlations, GLMs, regressions, LDAs, and voxel-wise contrasts were adjusted for age, sex, and education. P < 0.01 was considered statistically significant. Data were analyzed using SPSS version 23.0 (IBM Corp, Armonk, NY, USA).
Discussion
This study evaluated RSI metrics for sensitivity to disease status and cognitive deficits in MCI and AD. RSI-based measures of gray and white matter microstructure correlated with disease severity, functional ability, memory, and Aβ load.
White matter integrity of several tracts with temporal lobe projections significantly correlated with memory and distinguished impaired from healthy participants. These findings are consistent with prior reports that reduced FA in these pathways distinguishes MCI or AD from normal aging [
14,
48], and tracks [
15] or predicts [
17,
54] disease progression. Effects were strongest for the IFOF, uncinate, and arcuate, long-range association fibers that connect temporal with frontal, occipital, and parietal cortex and may support integrative processing critical to memory or other cognitive functions that decline with disease progression.
Entorhinal cortex IF and MD were the strongest correlates of memory impairment and disease status, in line with the origins of AD neuropathology in the entorhinal area [
2,
3]. Thus, this notable sensitivity of entorhinal cortex microstructural changes may reflect more advanced pathology in this focal region with relative sparing of other regions at the mild disease stages examined here. Entorhinal cortex but not hippocampal IF correlated with Aβ42, and increased free water diffusion has also been observed in the hippocampus but not entorhinal cortex in normal aging, highlighting the possible specificity of entorhinal microstructural change to prodromal AD [
55].
RSI and DTI metrics are highly correlated, as demonstrated by their comparable associations with memory. However, we found preliminary evidence that RSI may offer additional information on microstructural changes to neural tissue and their association to early AD neuropathology. Intriguingly, lower ND and higher IF were respectively associated with greater burden (lower CSF levels) of Aβ42, the principal component of amyloid plaques and considered a crucial peptide for AD pathogenesis. In contrast, neither FA nor MD in any region correlated with Aβ levels, suggesting that RSI tracks neuropathological burden more closely than DTI. Prior studies report conflicting associations of Aβ with increased [
25] or decreased [
24] FA. These inconsistencies in the literature, and the lack of association of Aβ with FA or MD in our data, may derive from limitations of DTI to resolve microarchitectural complexities such as crossing fibers. Thus, RSI may provide the advantage over conventional diffusion imaging techniques of more complete characterization of tissue microstructure which may better inform about the relationship between tissue disorganization and its underlying pathophysiology.
Furthermore, there was a trend for stronger group differences for ND than FA in all fiber tracts, and for entorhinal cortex IF than MD, and discriminant analysis selected RSI over DTI metrics to distinguish impaired from healthy participants. When both RSI and DTI metrics were submitted to regression models, only RSI measures predicted dementia severity and CVLT scores, whereas no cognitive measure was predicted by DTI measures only. RSI and DTI measures differentially related to delayed recall, with a stronger association of arcuate ND to CVLT scores and of IFOF and fornix FA to CERAD scores (entorhinal cortex IF predicted both). The CVLT may be more sensitive to subtle memory impairments than the CERAD [
56] and, here, delayed recall deficits for impaired individuals were also more severe for the CVLT than the CERAD. Microstructural changes in entorhinal cortex and arcuate, detectable with RSI, may therefore be powerful indices of mild memory impairment.
Although other advanced diffusion MRI techniques may also improve sensitivity to neuropathological tissue microarchitecture compared to DTI, each approach is distinct in its implementation efficiency and characterization of complex fiber geometry. RSI is a multishell, multicompartment extension of traditional high-angular resolution diffusion imaging (HARDI). While HARDI can resolve complex fiber orientations [
57] it is blind to length scale information and thus cannot distinguish hindered and restricted diffusion pools. Diffusion kurtosis imaging (DKI) [
58], which also employs a multishell, multidirection acquisition, only indirectly estimates the structural complexity of tissue from measures of diffusional kurtosis. Neurite orientation dispersion and density imaging (NODDI) [
59], which like RSI integrates a multishell HARDI acquisition with a multicompartment model, can also separate tissue compartments and assess complex tissue microstructure. However, whereas NODDI characterizes the degree of fiber dispersion, RSI further identifies the geometric pattern of dispersion (e.g., crossing fibers) within a more efficient acquisition time (6.5 versus 30 min [
59]). Our findings add to a small but growing literature demonstrating sensitivity of advanced diffusion imaging techniques such as DKI and NODDI to neural microstructural changes in AD. RSI may provide metrics of tissue architecture complementary to these approaches, to offer a more complete characterization of pathological brain microstructure in AD.
While further study is needed to identify the neurobiological substrates underlying reduced ND and increased IF in MCI and AD, these changes broadly reflect fewer barriers to diffusion within brain tissue. Prior histological validation indicates that ND correlates highly with neurite integrity [
27], and reduced ND may arise from various factors including reduced axon or dendrite count or density, demyelination, or synapse loss [
12,
60]. Greater isotropic free water diffusion in neurodegenerative disease could reflect expansion of the extracellular space related to neuronal loss, cell shrinkage, or tissue disorganization. Because RSI can isolate isotropic free water diffusion from restricted and hindered diffusion compartments, IF provides a more specific measure than average voxel diffusion, which may explain the more limited spatial distribution of AD-related changes observed for IF than MD in whole-brain analyses. Additional histological comparison and integration of RSI with multimodal imaging and computational modeling [
61] may better clarify how microstructural brain changes relate to underlying cell pathology and brain network reorganization.
Although RSI overcomes many obstacles posed by conventional diffusion imaging methods, some remaining limitations warrant consideration. Partial volume effects may artificially deflate estimates of both FA and ND, although we attempted to account for partial voluming in the cortical surface and when identifying fiber tracts. These artifacts would be most problematic for fine tracts adjacent to CSF, such as the fornix [
62]; here, the strongest predictors of cognitive function and disease were from thicker long-range association fibers that should be more robust to partial volume effects [
63]. Clinical diagnoses were made according to standard criteria; nevertheless, MCI is a heterogeneous condition and even a diagnosis of AD is not definitive without postmortem validation. The stepwise increases in dementia severity and functional ratings from HC to MCI to AD suggest that these groups represent a spectrum from healthy to mildly impaired. However, because we enrolled only mildly impaired AD participants, and MCI and AD participants did not differ on measures of brain microstructure or memory, clinical overlap may exist between these groups. Because neural microstructure may not change linearly with disease progression, nonlinear associations of RSI metrics with severity of cognitive impairment will be a topic for follow-up investigation. Finally, diffusion imaging cannot directly inform about the cellular pathology mediating neuroanatomical differences. Future studies are needed to further assess the biological underpinnings of RSI biomarkers and evaluate their sensitivity to preclinical changes indicative of subsequent cognitive and functional decline.