The Alzheimer's Disease Neuroimaging Initiative (ADNI) recently added diffusion tensor imaging (DTI), among several other new imaging modalities, in an effort to identify sensitive biomarkers of Alzheimer's disease (AD). While anatomical MRI is the main structural neuroimaging method used in most AD studies and clinical trials, DTI is sensitive to microscopic white matter (WM) changes not detectable with standard MRI, offering additional markers of neurodegeneration. Prior DTI studies of AD report lower fractional anisotropy (FA), and increased mean, axial, and radial diffusivity (MD, AxD, RD) throughout WM. Here we assessed which DTI measures may best identify differences among AD, mild cognitive impairment (MCI), and cognitively healthy elderly control (NC) groups, in region of interest (ROI) and voxel-based analyses of 155 ADNI participants (mean age: 73.5 ± 7.4; 90 M/65 F; 44 NC, 88 MCI, 23 AD). Both VBA and ROI analyses revealed widespread group differences in FA and all diffusivity measures. DTI maps were strongly correlated with widely-used clinical ratings (MMSE, CDR-sob, and ADAS-cog). When effect sizes were ranked, FA analyses were least sensitive for picking up group differences. Diffusivity measures could detect more subtle MCI differences, where FA could not. ROIs showing strongest group differentiation (lowest p-values) included tracts that pass through the temporal lobe, and posterior brain regions. The left hippocampal component of the cingulum showed consistently high effect sizes for distinguishing groups, across all diffusivity and anisotropy measures, and in correlations with cognitive scores.
The Alzheimer's Disease Neuroimaging Initiative (ADNI) recently added diffusion tensor imaging (DTI), among several other new imaging modalities, in an effort to identify sensitive biomarkers of Alzheimer's disease (AD). While anatomical MRI is the main structural neuroimaging method used in most AD studies and clinical trials, DTI is sensitive to microscopic white matter (WM) changes not detectable with standard MRI, offering additional markers of neurodegeneration. Prior DTI studies of AD report lower fractional anisotropy (FA), and increased mean, axial, and radial diffusivity (MD, AxD, RD) throughout WM. Here we assessed which DTI measures may best identify differences among AD, mild cognitive impairment (MCI), and cognitively healthy elderly control (NC) groups, in region of interest (ROI) and voxel-based analyses of 155 ADNI participants (mean age: 73.5 ± 7.4; 90 M/65 F; 44 NC, 88 MCI, 23 AD). Both VBA and ROI analyses revealed widespread group differences in FA and all diffusivity measures. DTI maps were strongly correlated with widely-used clinical ratings (MMSE, CDR-sob, and ADAS-cog). When effect sizes were ranked, FA analyses were least sensitive for picking up group differences. Diffusivity measures could detect more subtle MCI differences, where FA could not. ROIs showing strongest group differentiation (lowest p-values) included tracts that pass through the temporal lobe, and posterior brain regions. The left hippocampal component of the cingulum showed consistently high effect sizes for distinguishing groups, across all diffusivity and anisotropy measures, and in correlations with cognitive scores.
Alzheimer's disease (AD) is the most common type of dementia,
affecting 1 in 8 people over age 65 in the U.S. alone (Alzheimer's Disease Association,
2012). Its prevalence is predicted to more than double in the
next 40 years (Hebert
et al., 2003). It is important to identify individuals most
likely to develop AD, so that those at greater risk can be treated earlier. One
high-risk group consists of people with mild cognitive impairment (MCI) — a
transitional stage between normal aging and AD. People with MCI convert to AD at
a rate of about 10–15% per year (Petersen et al., 2001; Bruscoli and Lovestone,
2004). In addition to the more widely-accepted measures from
anatomical MRI, PET, and CSF measures of pathology, one major neuroimaging study
of AD – the Alzheimer's Disease Neuroimaging Initiative (ADNI) – recently
incorporated additional neuroimaging techniques including diffusion tensor
imaging (DTI) (Jack et
al., 2010; Jahanshad et al., 2010a; Zhan et al., in press).
DTI is a variant of MRI that measures the diffusion of water molecules in brain
tissue. Here we set out to assess which standard DTI measures may best identify
neuroanatomical differences between AD, MCI, and normal aging. In the end, DTI
offers a range of measures that might be sensitive to pathology, including
measures of brain connectivity (Daianu et al.,
2012; Nir et al., 2012; Prasad et al., 2013; Toga and Thompson, 2013; Daianu
et al., accepted for publication; Daianu et al., submitted for
publication). For these initial analyses, however, we aimed
to analyze more traditional measures and maps that are perhaps most likely to be
used in standardized multi-site DTI analyses, at least in the near future
(Jahanshad et al.,
2013).MRI-based image analysis methods have long been used to track
structural atrophy of the aging brain. MRI studies of AD reveal widespread
neuronal loss and atrophy in the brain's gray matter, especially in medial
temporal and hippocampal regions (Atiya et al., 2003; Chetelat and Baron, 2003; Thompson et al.,
2003; Anderson et al., 2005; Apostolova and Thompson, 2008; Bakkour et al.,
2009; Risacher et al., 2009; Apostolova et al., 2010; Desikan et al., 2010a;
Desikan et al., 2010b; Chiang et al., 2011; Weiner et al., 2012; Leung et
al., 2013). Beta-amyloid and tau proteins accumulate in the
brain, leading to inflammation, neuronal atrophy and cell death (Braak and Braak, 1991; Braak and Braak,
1995). As neurons are lost, white matter volume is also
reduced, due to both myelin degeneration and axon loss in neural fiber tracts
(Braak and
Braak, 1996; Bartzokis, 2011; Braskie et al., 2011; Hua et al.,
2013). Standard anatomical MRI is still the imaging
technique most often used in AD studies and clinical trials, but DTI is
sensitive to microscopic changes in white matter (WM) integrity not always
detectable with standard anatomical MRI (Xie et al., 2006; Canu et al., 2010). Although
this is debatable until more evidence is collected, some DTI changes may even
precede and predict volume loss (Hugenschmidt et al., 2008; Nir et al., 2012),
making it a potentially beneficial tool for capturing additional or
complementary markers of early neurodegeneration. Carriers of some AD risk genes
show differences on DTI as young adults, decades before the typical age of onset
of AD (Braskie et al.,
2011).Fractional anisotropy (FA) is perhaps the most widely accepted
DTI measure and reflects how directionally constrained the diffusion of water is
along axons. While higher FA values may imply more coherent or intact axons, or
a higher degree of myelination, lower FA may reflect loss of WM integrity and
injury. These physiological correlates of the DTI signal are widely accepted,
but the differences may have other interpretations, especially where fibers
cross (Leow et al., 2009; Zhan
et al., 2009). Mean diffusivity (MD) captures the average
rate of diffusion in all directions, and generally increases with WM injury,
especially if normal barriers to diffusion are damaged (such as myelin sheaths
on axons). Axial diffusivity (AxD) captures diffusion parallel to axonal fibers,
while radial diffusivity (RD) reflects perpendicular diffusion. These measures
are linked to axonal injury and demyelination, respectively (Song et al., 2003; Song et al.,
2005). To date, numerous DTI studies of AD and MCI find that
greater cognitive impairment, or poorer diagnosis, is associated with lower FA
in the corpus callosum, fornix, cingulum, superior longitudinal fasciculus, and
inferior longitudinal fasciculus (Ukmar et al., 2008; Stricker
et al., 2009; Mielke et al., 2009; Liu et al., 2011) and DTI
measures correlate with widely used clinical or cognitive ratings including the
mini-mental state exam (MMSE) (Bozzali
et al., 2002).Despite growing diffusion imaging evidence of AD-related WM
changes, it is not clear which regions and DTI measures are the most sensitive
for detecting diagnostic differences. In order to evaluate the power of drug
trial treatment to counteract degeneration, optimizing statistical power for
discerning differences and changes is crucial. We focused this current paper on
cross-sectional differences in patients and controls, as there are a number of
DTI measures, regions, and approaches that need to be compared and ranked in
terms of their effect sizes for picking up group differences. We set out to rank
the effect sizes for different DTI-based scalar measures in detecting
differences in both white matter voxel-based analyses (VBA) and within regions
of interest (ROIs). We first examined differences in DTI anisotropy and
diffusivity measures, between groups of cognitively healthy normal elderly
controls (NC), MCI, and ADpatients in both voxel-based and ROI analyses. We
also examined the association of anisotropy and diffusivity maps with widely
used clinical or cognitive ratings including the MMSE (Folstein et al., 1975), the
“sum-of-boxes” clinical dementia rating (CDR-sob) (Berg, 1988), and the Alzheimer's Disease
Assessment Scale-Cognitive (ADAS-cog) (Rosen et al., 1984). Finally, in a supplementary test, we
compared our ROI results to ROIs extracted along the skeleton from the widely
used tract-based spatial statistics (TBSS) method (Smith et al., 2006). Despite the popularity of
FA, we hypothesized that we would find the highest effect size and
discriminative power for MD measures, as recently suggested in a review of DTI
studies of AD by Clerx et al. (Clerx et
al., 2012). We also hypothesized that we would find the
greatest differences in temporal lobe WM and the corpus callosum (CC), as the
temporal lobe is usually the earliest region to be affected by amyloid and tau
pathology in AD and DTI studies are often better powered to find group
differences in regions such as the CC where fiber coherence is
highest.
Materials and methods
Clinical sample and
demographics
Baseline MRI, DTI, clinical, and
neuropsychological data were downloaded from the ADNI database
(http://adni.loni.ucla.edu). When the analysis was
performed (September 2012), data collection for the ADNI2 project was still
in progress. Here we performed an initial analysis of 155 participants from
14 data acquisition sites, of whom 44 were normal controls (NC), 88 amnestic
MCI subjects, and 23 ADpatients (see Inline Supplementary Table S1 for
distribution of subjects across sites). Unlike ADNI1, ADNI2 MCI participants
include the enrollment of a new early MCI cohort (e-MCI; n = 62), with milder episodic memory impairment than
the MCI group of ADNI1. The MCI group of ADNI1 is now referred to as late
MCI (l-MCI; n = 26) in ADNI2. Levels of
MCI (early or late) were determined using the Wechsler Memory Scale —
Logical Memory II (Wechsler,
1987). We evaluated the l-MCI and e-MCI groups both
separately and as one large MCI group. Detailed inclusion and exclusion
criteria are found in the ADNI2 protocol (http://adni-info.org/Scientists/Pdfs/ADNI2_Protocol_FINAL_20100917.pdf).Baseline MRI, DTI, clinical, and
neuropsychological data were downloaded from the ADNI database
(http://adni.loni.ucla.edu). When the analysis was
performed (September 2012), data collection for the ADNI2 project was still
in progress. Here we performed an initial analysis of 155 participants from
14 data acquisition sites, of whom 44 were normal controls (NC), 88 amnestic
MCI subjects, and 23 ADpatients (see Inline Supplementary Table S1 for
distribution of subjects across sites). Unlike ADNI1, ADNI2 MCI participants
include the enrollment of a new early MCI cohort (e-MCI; n = 62), with milder episodic memory impairment than
the MCI group of ADNI1. The MCI group of ADNI1 is now referred to as late
MCI (l-MCI; n = 26) in ADNI2. Levels of
MCI (early or late) were determined using the Wechsler Memory Scale —
Logical Memory II (Wechsler,
1987). We evaluated the l-MCI and e-MCI groups both
separately and as one large MCI group. Detailed inclusion and exclusion
criteria are found in the ADNI2 protocol (http://adni-info.org/Scientists/Pdfs/ADNI2_Protocol_FINAL_20100917.pdf).Distribution of study
participants by diagnostic group across each of
the 14 acquisition sites. We report the
p-values from a
χ2 test
between the actual and expected distribution of
subjects, which we estimated based on the total
number of subjects in each diagnostic group, as a
proportion of the total, across sites. Although
numbers for some sites are small, there is no
evidence that any site is enrolling a
statistically divergent proportion of subjects in
each category.Inline Supplementary Table S1 can be
found online at http://dx.doi.org/10.1016/j.nicl.2013.07.006.Each subject underwent cognitive evaluations. The
Mini-Mental State Examination (MMSE) was used to provide a global measure of
cognitive status, based on evaluating cognitive domains including
orientation to place, orientation to time, registration, attention and
concentration, recall, language, and visual construction (Folstein et al., 1975). The
total score ranges from 1 to 30, with lower scores indicating impairment.
The Clinical Dementia Rating (CDR) was also used as a global measure of
dementia severity (Berg,
1988). The “sum-of-boxes” CDR (CDR-sob) score is the sum
of 6 measures each assessing the degree of impairment in memory,
orientation, judgment and problem solving, community affairs, home and
hobbies, and personal care. The CDR-sob score ranges from 0 to 18 (no
dementia to severe dementia, respectively). Finally the Alzheimer's Disease
Assessment Scale-Cognitive (ADAS-cog), a global measure encompassing memory,
reasoning, language, orientation, ideational, praxis and constructional
praxis (Rosen et al.,
1984), was collected where scores range from 0 to 70 (no
dementia to severe dementia respectively). In
post-hoc analyses, we further homed in on
specific cognitive domains using the available ADNI composite scores for
executive function (ADNI-EF) (Gibbons et al., 2012) and memory (ADNI-MEM)
(Crane et al.,
2012) derived using data from the ADNI neuropsychological
battery. Detailed psychometric calculation protocols are available for
download at https://ida.loni.ucla.edu/. ADNI-EF was
calculated using a combination of WAIS-R Digit Symbol Substitution, Digit
Span Backwards, Trails A and B, Category Fluency, and Clock Drawing scores
(Gibbons et al.,
2012) and ADNI-MEM was calculated as a composite of the
Rey Auditory Verbal Learning Test (RAVLT), ADAS-Cog, and Logical Memory data
(Crane et al.,
2012).Demographics and diagnostic information for the participants
are shown in Table 1. Diagnostic
groups did not differ in age, however, education, an AD risk factor
(Sattler et al.,
2012), was marginally significant between controls and
AD. As would be expected, clinical measures that index cognitive decline
(MMSE, ADAS-cog, CDR-sob, ADNI-MEM, ADNI-EF) did show significant graded
differences between groups.
Table 1
Demographics and clinical scores for the
participants.
NC
MCI
e-MCI
l-MCI
AD
p-value for group difference
(n = 44)
(n = 88)
(n = 62)
(n = 26)
(n = 23)
NC vs MCI
MCI vs AD
e-MCI vs l-MCI
NC vs AD
Age
72.7 ± 5.9
73.3 ± 7.3
74.0 ± 7.9
71.8 ± 5.3
75.8 ± 10.0
0.59
0.27
0.15
0.17
Sex
22 M/22 F
53 M/35 F
38 M/24 F
15 M/11 F
15 M/8 F
–
–
–
–
Education
16.6 ± 2.7
15.9 ± 2.7
15.8 ± 2.8
16.2 ± 2.5
15.0 ± 3.0
0.18
0.22
0.50
0.05
MMSE
28.9 ± 1.3
27.8 ± 1.6
28.0 ± 1.5
27.5 ± 1.8
23.0 ± 1.9
1.01E− 4
4.90E− 12
0.24
1.57E− 14
ADAS-coga
5.3 ± 2.9
9.6 ± 4.2
8.4 ± 3.7
12.7 ± 4.0
20.4 ± 7.3
2.90E− 9
2.04E− 6
1.45E− 4
1.04E− 8
CDR-sob
0.03 ± 0.1
1.3 ± 0.7
1.2 ± 0.6
1.4 ± 0.9
4.96 ± 1.4
1.38E− 27
2.69E− 12
0.27
2.68E− 14
ADNI-MEMb
0.84 ± 0.52
0.22 ± 0.46
0.36 ± 0.42
− 0.12 ± 0.38
− 0.74 ± 0.68
1.01E− 8
4.14E− 6
5.97E− 6
3.64E− 10
ADNI-EFb
0.79 ± 0.74
0.16 ± 0.64
0.14 ± 0.61
0.19 ± 0.73
− 0.86 ± 0.87
1.65E− 5
5.16E− 5
0.75
2.51E− 8
Bold signifies significant (i.e. p < 0.05).
ADAS-cog data were available only for a subset of the
subjects, with the following numerical breakdown: NC n = 41, MCI n = 78
(e-MCI = 57, l-MCI = 21), AD n = 20.
ADNI-MEM and ADNI-EF composite scores were available only
for a subset of the subjects, with the following numerical breakdown: NC
n = 41, MCI n = 82 (e-MCI = 58, l-MCI = 24), AD n = 20.
We further assessed whether these measures revealed more
fine-grained differences between the l-MCI and e-MCI subgroups. We found
differences in ADAS-cog scores — a measure developed specifically for
Alzheimer's disease, and the most widely used primary outcome measure in
clinical trials that test AD drug treatment efficacy (Mohs et al., 1983). ADNI-MEM
was also significantly different as it is calculated using both ADAS-cog and
Logical Memory data, the measure used to initially classify the two MCI
subgroups.
MRI and DTI scanning
All subjects underwent whole-brain MRI scanning on 3 Tesla GE Medical Systems scanners at 14 acquisition sites
across North America. Anatomical T1-weighted SPGR (spoiled gradient echo)
sequences (256 × 256 matrix; voxel
size = 1.2 × 1.0 × 1.0 mm3; TI = 400 ms; TR = 6.98 ms; TE = 2.85 ms; flip angle = 11°), and diffusion-weighted images (DWI;
256 × 256 matrix; voxel size:
2.7 × 2.7 × 2.7 mm3; TR = 9000 ms; scan
time = 9 min; more
imaging details can be found at http://adni.loni.ucla.edu/wp-content/uploads/2010/05/ADNI2_GE_3T_22.0_T2.pdf)
were collected. 46 separate images were acquired for each DTI scan: 5
T2-weighted images with no diffusion sensitization (b0
images) and 41 diffusion-weighted images (b = 1000 s/mm2). This protocol was chosen after conducting a
detailed comparison of several different DTI protocols, to optimize the
signal-to-noise ratio in a fixed scan time (Jahanshad et al., 2010a; Zhan et al.,
in press). All T1-weighted MR and DWI images were
checked visually for quality assurance to exclude scans with excessive
motion and/or artifacts; all scans were included.
Image analysis
Preprocessing steps
For each subject, all raw DWI volumes were aligned to
the average b0 image (DTI volume with no diffusion
sensitization) using the FSL eddy_correct tool
(www.fmrib.ox.ac.uk/fsl) to correct for
head motion and eddy current distortions. All extra-cerebral tissue was
roughly removed from the T1-weighted anatomical scans using a number of
software packages, primarily ROBEX, a robust automated brain extraction
program trained on manually “skull-stripped” MRI data (Iglesias et al., 2011) and
FreeSurfer (Fischl et al.,
2004). Skull-stripped volumes were visually
inspected, and the best one selected and sometimes further manually
edited. Anatomical scans subsequently underwent intensity inhomogeneity
normalization using the MNI nu_correct tool
(www.bic.mni.mcgill.ca/software/).
Non-brain tissue was also removed from the diffusion-weighted images
using the Brain Extraction Tool (BET) from FSL (Smith, 2002). To align
data from different subjects into the same 3D coordinate space, each
T1-weighted anatomical image was linearly aligned to a standard brain
template (the downsampled Colin27 (Holmes et al., 1998): 110 × 110 × 110, with 2 mm isotropic voxels) using
FSL flirt (Jenkinson et al., 2002) with 6 degrees of
freedom (dof) to allow translations and rotations in 3D. To correct for
echo-planar imaging (EPI) induced susceptibility artifacts, which can
cause distortions at tissue–fluid interfaces, skull-stripped
b0 images were linearly aligned (FSL
flirt 9 dof) and then elastically registered
to their respective T1-weighted structural scans using an
inverse-consistent registration algorithm with a mutual information cost
function (Leow et al.,
2007) as described in (Jahanshad et al., 2010b). The resulting
3D deformation fields were then applied to the remaining 41 DWI volumes
prior to estimating diffusion parameters. To account for the linear
registration of the DWI images to the structural T1-weighted scan, a
corrected gradient table was calculated.
DTI maps
A single diffusion tensor (Basser et al., 1994), or ellipsoid, was
modeled at each voxel in the brain from the eddy- and EPI-corrected DWI
scans using FSL dtifit, and scalar anisotropy and
diffusivity maps were obtained from the resulting diffusion tensor
eigenvalues (λ1, λ2,
λ3) which capture the length of the longest,
middle, and shortest axes of the ellipsoid. The tensor model in DTI has
limitations, especially in regions where fibers cross, but we do not
investigate it further here; our other papers consider this in more
detail (Zhan et al., 2008; Leow et al., 2009; Zhan et al., 2009; Zhan et
al., 2010; Zhan et al., 2011; Zhan et al., 2012b; Zhan et al., in
press). Fractional anisotropy (FA), a measure of the
degree of diffusion anisotropy, was calculated from the standard formula:- where <λ> is the mean diffusivity (MD), or average
rate of diffusion in all directions. Axial diffusivity was defined as
the primary (largest) eigenvalue (AxD = λ1), and captures
the longitudinal diffusivity, or the diffusivity parallel to axonal
fibers (assuming of course that the principal eigenvector is indeed
following the dominant fiber direction, which may be unclear in regions
with extensive fiber crossing). Radial diffusivity (RD), which captures
the average diffusivity perpendicular to axonal fibers, was calculated
as the average of the two smaller eigenvalues:
White matter tract atlas ROI summary
measures
We linearly, then elastically registered (Leow et al., 2007) the FA
image from the Johns Hopkins University (JHU) DTI atlas (Mori et al., 2008) to each
subject's distortion corrected FA image. We then applied that
deformation to the stereotaxic JHU “Eve” atlas WM labels (http://cmrm.med.jhmi.edu/cmrm/atlas/human_data/file/AtlasExplanation2.htm),
using nearest neighbor interpolation to avoid intermixing of labels.
This is not the atlas that had the problem pointed out by Rohlfing (2013). We then
superimposed the atlas ROIs into the same coordinate space as our
results. We removed 10 ROIs from the analyses (including the left and
right middle cerebellar peduncle, pontine crossing tract, medial
lemniscus, inferior and superior peduncles) as they often fell partially
or completely out of the field-of-view (FOV) of the images. We also
excluded the body of the fornix as it is small and prone to
misregistration and partial voluming. AD researchers are specifically
interested in the fornix as it is the primary posterior pathway coming
out of the back of the hippocampus, a key target of pathology. While we
included the crus of the fornix/stria terminalis,
the body is just too small to be resolved well on DTI at this
resolution. In addition to the JHU labels, 4 more ROIs were evaluated:
bilateral genu, body, and splenium of the corpus callosum (as opposed to
just the lateralized measures), as well as the entire corpus callosum,
and a large “TOTAL” WM ROI made up of all the other ROIs, to obtain
total summary measures of these regions. We were then able to calculate
the average FA, MD, RD and AxD, within the boundaries of each of the 43
ROIs for each subject (Table 2).2
Table 2
Index of 43 ROIs from the WM tract atlas (Mori et al., 2008) followed by
their abbreviations.
Genu of corpus callosum
GCC
Posterior thalamic radiation
PTR L,R
Body of corpus callosum
BCC
Sagittal stratum
SS L,R
Splenium of corpus callosum
SCC
External capsule
EC L,R
Full corpus callosum
CC
Cingulum (cingulate gyrus)
CGC L,R
Corticospinal tract
CST L,R
Cingulum (hippocampus)
CGH L,R
Cerebral peduncle
CP L,R
Fornix
(crus)/Stria
terminalis
FX/ST L,R
Anterior limb of internal capsule
ALIC L,R
Superior longitudinal fasciculus
SLF L,R
Posterior limb of internal
capsule
PLIC L,R
Superior fronto-occipital
fasciculus
SFO L,R
Retrolenticular part of internal
capsule
RLIC L,R
Inferior fronto-occipital
fasciculus
IFO L,R
Anterior corona
radiata
ACR L,R
Uncinate fasciculus
UNC L,R
Superior corona
radiata
SCR L,R
Tapetum
TAP L,R
Posterior corona
radiata
PCR L,R
All ROIs
TOTAL
TBSS tract atlas ROI summary
measures
Tract-based spatial statistics (TBSS) (Smith et al., 2006),
provided in the FSL software package (http://www.fmrib.ox.ac.uk/fsl/), was also performed
according to protocols outlined by the ENIGMA-DTI group: http://enigma.loni.ucla.edu/ongoing/dti-working-group/.
All subjects' corrected FA maps were linearly, then elastically
registered (Leow et al.,
2007) to the ENIGMA-DTI template in ICBM space. The
resulting 3D deformation fields were then applied to the three
diffusivity maps. All subjects' spatially normalized FA, MD, RD and AxD
data were projected onto the skeletonized ENIGMA-DTI template. Mean
anisotropy and diffusivity measures were calculated along the skeleton
in the same 43 ROIs (Table 2). This type of analysis has been used
previously in both genetic studies and studies of disease to home in on
associated WM tracts (Kochunov et al., 2011; Jahanshad et al.,
2013).
Template creation and spatial
normalization
A study-specific minimal deformation template (MDT)
(Gutman et al.,
2012) was created using 29 cognitively healthy
elderly control (NC) spatially aligned FA maps. An MDT deviates, on
average, the least (in some metric) from the anatomy of the subjects,
and can often improve registration accuracy and statistical power
(Gutman et al.,
2012; Lepore et al., 2007). The MDT was generated by
creating an initial affine mean template from all 29 subjects, then
registering all the aligned individual scans to that mean using a fluid
registration (Leow et al.,
2007) while regularizing the Jacobians (Yanovsky et al., 2007). A
new mean was created from the registered scans; this process was
iterated several times. Each subject's initial FA map was elastically
registered to the final MDT and the resulting deformation fields were
applied to the 3 diffusivity maps to align them to the same coordinate
space. To ensure white matter alignment across subjects, registered FA
maps were thresholded at FA > 0.2
to include only highly anisotropic anatomy and the thresholded maps were
elastically registered to the thresholded MDT (FA > 0.2). Again, the resulting deformation
fields were applied to all previously registered DTI maps. We also used
the tissue-specific, smoothing-compensated method
(T-SPOON) proposed by Lee et al. (2009) to
improve tissue specificity and reduce confounds caused by morphometric
differences that are not fully corrected by the elastic
registration.
Statistical analysis
We ran voxel-wise multiple linear regressions, covarying for
age and sex, to test for statistical effects of AD and MCI diagnosis
relative to the NC group, on measures of white matter integrity in FA, MD,
RD, and AxD maps. We also tested for associations between these DTI measures
and MMSE, CDR-sob, and ADAS-cog scores, controlling for age and sex, across
the entire study sample (i.e., including all AD, MCI and NC subjects), and
also within each subgroup. We also ran all of the regressions using a
random-effects regression model, grouping the data by acquisition site. To
limit statistical testing to white matter, where the power is greater to
detect differences, statistics were run only on voxels within the boundaries
of the MDT thesholded at FA > 0.2.
Prior studies have thresholded FA values between 0.2 and 0.3 to successfully
exclude gray matter or CSF (Wakana et al., 2005; Smith et al., 2006). As we were
studying a clinical population, we chose the more conservative (lower) limit
of the recommended FA threshold. We also only ran statistics on voxels of
the thresholded MDT present in all subject scans, as some scans had a
slightly cropped FOV. As such, we did not consider the inferior parts of the
cerebellum and brain stem.We ran random-effects regressions on the average anisotropy
and diffusivity measures within the 43 full ROIs, again covarying for age
and sex, testing for statistical effects of diagnosis (NC vs. MCI, or AD), 3
global clinical test scores, and post-hoc tests on
ADNI-MEM and ADNI-EF. We further tested and compared TBSS ROI
measures.Computing multiple association tests for each ROI, or
thousands of tests on a voxel-wise level can introduce a high false positive
error rate. To account for these errors, we used the standard false
discovery rate (FDR) method to control the false positive rate of each map
at q = 0.05
(Benjamini and Hochberg,
1995). All statistical maps shown in this paper
(Figs. 1–5) were
thresholded at the FDR critical p-value.To visualize and compare effect sizes in the anisotropy and
diffusivity maps, we computed and graphed the cumulative distribution
functions (CDF) of the p-values obtained from the
voxel-wise and ROI random-effects regressions. The ordered set of observed
p-values was plotted against the expected null
distribution. If there are no group differences (a null distribution), then
the plot would fall approximately along the y = x diagonal line.
However, if the CDF initially rises at a rate steeper than 20 times the null
CDF (y = 20x), then the corresponding maps are FDR
significant at q = 0.05. A greater slope reflects a larger effect size; the FDR critical
p-value (at q = 0.05) is the point at which the curve
intersects the y = 20x line.DTI measures calculated from a single-tensor diffusion model
have limitations in regions with extensive fiber crossing. For example, in
AD we might expect FA deficits throughout the brain, but FA may appear to be
artificially increased where crossing fibers
deteriorate (Douaud et al.,
2011). To make sure that these results were not false
positives we separated the full p-value map
(irrespective of the significance level of the voxels) into two maps: (1)
voxels that showed a negative association, and (2) positively associated
voxels. We then independently corrected each p-map
for multiple statistical comparisons using FDR (Benjamini and Hochberg, 1995).As an alternative way to visualize white matter integrity
differences between patients and controls, we also created maps showing the
“percent difference” in mean DTI measures between ADpatients and NC, within
the FDR significant regions.There is a mild bias in doing this, in that it selects
voxels that show effects, and then computes the magnitude of the effects.
Even so, it was used to simply describe how much difference was seen in the
regions shown, bearing in mind that the same difference would not be found
in the rest of the brain.
Results
Voxel-based analyses
White matter integrity differences between
diagnostic groups
AD vs controls
As measured by FA, ADpatients showed pervasive
deficits in fiber integrity (lower FA) compared to healthy controls,
throughout the WM across the entire brain, when using a general
linear regression model to adjust for age and sex (critical
p < 0.016; 31.9% of voxels survived the FDR threshold (Benjamini and Hochberg,
1995); minimum p-value,
1.65 × 10− 8). However, when we used a
random-effects regression model at every voxel to take into account
acquisition site differences, we found more significant and even
more widespread results (critical p < 0.020; 40.1% of voxels
survived the FDR threshold (Benjamini and Hochberg, 1995); minimum
p-value, 6.66 × 10− 9;
Fig. 1a).
We therefore proceeded with this model – fitting a site effect – for
the remainder of the analyses. A few significant voxels exhibited
associations in a direction opposite to the great majority of the
brain map, and contrary to what would traditionally be accepted as
showing impairment. For example, we found regions where the AD group
had a higher mean FA (Fig. 1e; blue regions).
These regions were largely found at the junction of the corpus
callosum commissural fibers and the corona
radiata. Such regions are notorious for fiber
crossings, which may artificially reduce FA, if FA is estimated
using the single-tensor diffusion model (Oishi et al., 2011).
This pattern has been reported in other AD studies, and may reflect
a selective sparing or selective degeneration of one of the pathways
in a region with crossing fibers (Douaud et al., 2011). When we
separated (and independently FDR-corrected) the map of voxels that
showed a negative association with AD and the map of the positively
associated voxels, the diffuse negative associations were still
significant, but the previously significant positive associations
were not.
Fig. 1
(Top panels) Statistical maps show − log10p-values within regions where AD patients exhibit
significantly higher (a) MD (FDR critical
p < 0.038),
(b) RD (FDR critical p < 0.041), and (c) AxD (FDR critical p < 0.027) than healthy controls (NC). The
cingulum, temporal lobe (including the hippocampal part of the cingulum) and
splenium of the corpus callosum show the “most significant” differences, i.e.
greatest effect sizes in the maps. (d) Significant but slightly less profuse FA
differences between groups were also found (FDR critical
p < 0.020).
(e) Beta-values (non-normalized slope of the regression)
within FDR significant regions largely revealed regions with
lower FA in AD, but there are small regions,
notorious for crossing fibers, where AD patients have higher mean FA than
controls (dark blue), contrary to what would
traditionally be accepted as showing impairment. (Bottom panel) Maps show the
percent difference in FA and diffusivity measures
between AD patients and NC within FDR significant regions. Again, the corpus
callosum, temporal lobe, cingulum (including the hippocampal part of the
cingulum), and the posterior thalamic radiations show the greatest degree of
difference (up to ~ 33%); we note that the significance values
and effect sizes in these regions depend on the mean group differences and its
standard error, which is more completely reflected in top panels. In both
figures, MD and RD differences are greater and more widespread. (For
interpretation of the references to color in this figure legend, the reader is
referred to the web version of this article.)
This pervasive pattern of significance was largely
replicated for the three other diffusivity measures. As predicted,
we found higher mean RD, AxD and MD values in ADpatients relative
to controls (Fig. 1 b–d) and the effects were larger and even
more diffuse than for FA. CDF plots confirm that FA was not the most
sensitive measure for differentiating groups. Comparisons of NC to
ADpatients (Fig. 2a) revealed that RD was the most sensitive
measure followed closely by MD.
Fig. 2
Effect sizes for FA, MD, RD and AxD measures in AD, MCI
and NC group comparisons and clinical cognitive test associations reveal that FA
is consistently the least sensitive measure. This type of plot is used to
compare effect sizes in statistical maps based on different diffusivity
measures. We used it to help decide which DTI measures best distinguish the
diagnostic groups. Here we show cumulative distribution function (CDF) plots of
the distribution of the p-values obtained from (a,d)
voxel-wise and (b–c) ROI linear regressions, which are subjected to multiple
comparisons correction using standard FDR (Benjamini and Hochberg, 1995). (a) VBA
comparisons of NC to AD patients revealed that RD was the most sensitive measure
(denoted by the higher critical p-values controlling the
FDR/the highest non-zero x-coordinate where the CDF
crosses the y = 20x line) followed closely by MD. (b) ROI
comparisons of NC vs AD confirm that FA was the least
sensitive measure. (c) ROI comparisons of NC vs MCI reveal that
AxD measures had the largest effect size, followed by MD (VBA comparisons were
not significant). (d) Voxel-wise and (e) ROI linear regressions on associations
between cognitive scores (MMSE, CDR-sob, and ADAS-cog) in the entire cohort and
anisotropy and diffusivity measures confirm that FA associations, while FDR
significant, show the smallest effect size, while MD and RD have the largest
effect sizes.
Early, late and all MCI vs
controls
No significant difference was detected between e-MCI
and l-MCI groups for any of the DTI measures, so we first assessed
the MCI group as a whole, followed by these subgroups. We found no
significant differences between NC and MCI as a whole (n = 88), or NC and the e-MCI
(n = 62) subgroup, for any
of the anisotropy or diffusivity measures. However, for NC compared
to l-MCI subjects (n = 26), we
did find significantly higher diffusivity measures in the left
hippocampal part of the cingulum (MD: FDR critical
p < 3.05 × 10− 5; RD: FDR critical
p < 2.72 × 10− 5; AxD: FDR critical
p < 2.46 × 10− 5; Fig. 3). While these measures are
significant on their own, the N was small. These contrasts will
benefit from further analysis, as the ADNI dataset grows.
Fig. 3
Statistical maps show the − log10p-values within regions where l-MCI subjects exhibit
significantly higher AxD (FDR critical
p < 2.46 × 10− 5) than healthy controls (NC). The left hippocampal part of the
cingulum was also significant in MD and RD maps. This region is small, but it
passes the conventional FDR correction for multiple comparisons, and is in a
region implicated in early changes in AD (medial temporal regions). With a
greater sample size, this effect may be detectable in the other hemisphere as
well.
Correlation with MMSE, CDR-sob, and ADAS-cog
neuropsychological scores
Full cohort
We first assessed anisotropy and diffusivity map
associations with widely used clinical cognitive ratings in the
entire study population, including ADpatients, MCI, and NC subjects
(Fig. 4).
Fig. 4a
shows WM regions where anisotropy and diffusivity differences
correlated with the MMSE scores in the entire study population
(n = 155). MMSE was
significantly positively associated with FA (FDR critical
p < 0.018), and negatively associated with diffusivity (MD: FDR
critical p < 0.037; RD: FDR critical
p < 0.040; AxD: FDR critical p < 0.028). That is, lower FA and higher
diffusivity, which typically indicate greater WM deficits, were
associated with lower MMSE scores, indicative of greater impairment.
Fig. 4b
shows WM regions where FA and diffusivity differences correlated
with the CDR-sob scores (n = 155). As expected, FA was negatively associated with CDR-sob (FDR
critical p < 0.027), and diffusivity was positively associated
(MD: FDR critical p < 0.039; RD: FDR critical
p < 0.042; AxD: FDR critical p < 0.029). Fig. 4c shows WM regions where FA was
negatively associated with ADAS-cog (n = 139; FDR critical p < 0.016), and diffusivity
positively associated (MD: FDR critical
p < 0.040; RD: FDR critical p < 0.041; AxD: FDR critical
p < 0.029). Lower FA and higher diffusivity were associated with
higher (worse) ADAS-cog and CDR-sob scores.
CDF plots of MMSE, ADAS-cog, and CDR-sob voxel-wise associations
again reveal that FA associations, while FDR significant, show the
smallest effect size, while MD and RD have the largest effect sizes
(Fig. 2d).
Fig. 4
Statistical maps show that clinical scores on the (a)
MMSE, (b) CDR-sob, and (c) ADAS-cog are related to detectable differences in
fractional anisotropy, and mean, radial and axial diffusivity in the entire
cohort. These maps show the − log10p-values within regions that significantly correlate with
cognitive scores. RD and MD show the most widespread and robust associations.
The hippocampal part of the cingulum and surrounding temporal lobes show the
“most significant” associations (i.e., lowest p-values
and highest voxel-wise effect sizes) between cognitive performance and white
matter integrity.
As with diagnostic group differences, in all
clinical score analyses we found small significant regions with
associations in a direction opposite to what would traditionally be
accepted as showing impairment. Again, when we corrected for
multiple comparisons on voxel-wise p-maps of
each association direction separately, the associations in the
unexpected direction were no longer significant and only the
intuitive direction remained.
Early, late and all MCI
We further assessed associations between cognitive
scores and DTI measures in just the MCI group, as well as e-MCI and
l-MCI subgroups. We found a significant positive association between
all 3 diffusivity measures and CDR-sob within the MCI group (MD: FDR
critical p < 0.006; RD: FDR critical
p < 0.006; AxD: FDR critical p < 0.002; Fig. 5) and a small positive
association between ADAS-cog and MD in the middle or lateral
occipital WM in l-MCI (FDR critical p < 5.35 × 10− 6).
However, there was no detectable association with anisotropy or
diffusivity measures and MMSE scores.
Fig. 5
Statistical maps show where CDR-sob scores are
significantly positively associated with diffusivity measures in the MCI group,
considered on its own (n = 88). Higher
diffusivity, indicative of greater WM deficits, is associated with greater
CDR-sob scores indicative of greater impairment. We find the “most significant”
association (i.e., strongest voxel-wise effect) in the splenium of the corpus
callosum and the posterior cingulum.
ROI analyses
ROI differences between diagnostic
groups
We sorted the JHU atlas ROI results by
p-value to assess which ROIs showed the
greatest differences between groups. In doing so, we have to bear in
mind that this is not a ranking of the disease effects on the brain
by severity; results may also be affected by how accurately each
region can be measured, which in turn depends on the size and the
homogeneity of the region. With these caveats in mind,
Tables 3–4 list the ‘top 10’ most
significant ROIs for each regression analysis (here and in what
follows, we use the term ‘most significant’ to mean lowest
p-values in the tests of group
differences in mean values; we acknowledge that this term is not
always preferred by statisticians as the
p-values from different tests cannot
necessarily be compared). When comparing NC to ADpatients
(Table 3a), the splenium of the corpus callosum, left
fornix (crus)/stria
terminalis, left hippocampal part of the cingulum,
and total ROI map, were in the top 10 most significant ROIs, across
all anisotropy and diffusivity measures. While the superior
corona radiata and left sagittal stratum
(which includes inferior longitudinal fasciculus and inferior
fronto-occipital fasciculus) were in the top 10, across all
diffusivity measures.
Table 3
Atlas ROI average anisotropy and diffusivity. While many
regions were statistically significant (Benjamini and Hochberg, 1995), here we highlight regions
with the “top 10” significant p-values (greatest effect
sizes) when comparing mean values between diagnostic groups (in the left column
we show the diagnostic groups being compared). We also note the total number of
significant ROIs (after FDR correction) out of the 43 tested and the critical
FDR p-value.
FA
MD
RD
AxD
ROI
p-Value
ROI
p-Value
ROI
p-Value
ROI
p-Value
a)NCvsAD
Fx/ST L
1.44E− 7
SS L
2.09E− 11
SS L
6.19E− 11
SS L
6.29E− 11
SCC
1.40E− 6
CGH L
1.23E− 9
CGH L
1.69E− 09
CGH L
1.68E− 9
TAP L
2.46E− 6
TOTAL
4.63E− 9
TOTAL
6.89E− 09
TOTAL
7.47E− 9
PTR L
2.07E− 5
SCR L
1.11E− 8
SCR L
6.22E− 08
SCR L
1.74E− 8
CC
4.41E− 5
SCC
6.54E− 8
PTR L
7.64E− 08
IFO L
3.04E− 8
SFO R
5.86E− 5
Fx/ST L
1.07E− 7
Fx/ST L
7.70E− 08
SS R
9.02E− 8
TAP R
8.79E− 5
PTR L
1.10E− 7
SCC
8.05E− 08
Fx/ST L
2.54E− 7
ACR L
1.35E− 4
SS R
1.24E− 7
TAP L
1.25E− 07
EC L
3.21E− 7
TOTAL
1.62E− 4
TAP L
1.43E− 7
ACR L
1.40E− 07
SCC
3.34E− 7
CGH L
1.84E− 4
IFO L
2.77E− 7
SCR R
2.82E− 07
UNC L
3.46E− 7
# sig/43
24
(p < 2.32E− 2)
42
(p < 2.54E− 2)
42
(p < 2.72E− 2)
41
(p < 4.23E− 2)
b)NCvsMCI
–
–
CGH L
5.65E− 5
CGH L
8.47E− 5
CGH L
6.91E− 5
–
–
SCC
2.60E− 3
SCC
3.23E− 3
SS L
3.18E− 3
–
–
Fx/ST L
4.39E− 3
Fx/ST L
3.89E− 3
SCC
3.65E− 3
–
–
TAP L
5.14E− 3
TAP L
4.48E− 3
Fx/ST L
6.06E− 3
–
–
SS L
5.48E− 3
–
–
TOTAL
6.14E− 3
–
–
–
–
–
–
SS R
6.92E− 3
–
–
–
–
–
–
PTR L
7.20E− 3
–
–
–
–
–
–
SCR L
8.80E− 3
–
–
–
–
–
–
TAP L
8.86E− 3
–
–
–
–
–
–
EC R
1.01E− 2
# sig/43
0
5
(p < 5.47E− 3)
4
(p < 4.48E− 3)
12
(p < 1.23E− 2)
c)NCvse-MCI
–
–
CGH L
6.16E− 4
CGH L
7.33E− 4
CGH L
8.92E− 4
–
–
SCR L
3.18E− 3
–
–
–
–
–
–
SCC
3.35E− 3
–
–
–
–
# sig/43
0
3
(p < 3.35E− 3)
1
(p < 7.33E− 4)
1
(p < 8.92E− 4)
d)NCvsl-MCI
–
–
CGH L
4.71E− 5
CGH L
9.53E− 5
CGH L
3.01E− 5
–
–
–
–
–
–
SS L
1.88E− 4
–
–
–
–
–
–
SS R
1.47E− 3
–
–
–
–
–
–
Fx/ST L
5.38E− 3
–
–
–
–
–
–
PTR L
5.68E− 3
# sig/43
0
1
(p < 4.71E− 5)
1
(p < 9.53E− 5)
5
(p < 5.678E− 3)
As there were no significant differences between
l-MCI and e-MCI subjects across any of the measures, we compared the
entire MCI group to NC subjects (Table 3b). While no ROI showed
significant FA differences, the splenium of the corpus callosum,
left tapetum, left hippocampal part of the cingulum, and left fornix
(crus)/stria
terminalis were in the top 10, among all diffusivity
measures. We further assessed differences between e-MCI and NC and
l-MCI and NC. The left hippocampal part of the cingulum was
significant across all three diffusivity measures in both
comparisons (Table 3c–d). FA showed no statistically
significant differences.CDF plots of the ROI p-values
for NC vs AD (Fig. 2b) confirm VBA findings that FA was the
least sensitive measure, and that RD has a slightly higher effect
size than other measures. However, CDF plots of the ROI
p-values, for NC vs MCI, reveal that AxD
measures had the largest effect size, followed by MD (Fig. 2c). Table 3d also reveals
that AxD has a larger effect size when comparing l-MCI and
NC.
ROI correlations with clinical global
neuropsychological scores
The strongest associations with MMSE score
(Table 4a)
across all 4 measures were found in the left hippocampal part of the
cingulum, the left fornix (crus)/stria
terminalis, and the total ROI map. All three diffusivity
measures were additionally associated with MMSE in the left cingulum,
and the bilateral sagittal stratum (this includes the inferior
longitudinal fasciculus and inferior fronto-occipital
fasciculus).
Table 4
Atlas ROI average anisotropy and diffusivity. Here we
highlight regions with the 10 lowest FDR significant (Benjamini and Hochberg, 1995)
p-values (greatest effect sizes) when assessing
cognitive test score associations in the entire study population. We also note
the total number of significant ROIs (after FDR correction) out of the 43 tested
and the critical FDR p-value.
FA
MD
RD
AxD
ROI
p-Value
ROI
p-Value
ROI
p-Value
ROI
p-Value
a) MMSE
Fx/ST L
3.14E− 5
CGH L
8.01E− 15
CGH L
1.96E− 14
CGH L
1.96E− 14
PTR L
2.11E− 4
CGC L
2.05E− 8
SS L
5.86E− 8
SS R
4.47E− 8
ACR L
2.13E− 4
SS L
2.98E− 8
CGC L
9.93E− 8
SS L
6.95E− 8
ACR R
3.60E− 4
ACR L
1.57E− 7
ACR L
1.22E− 7
TOTAL
1.23E− 7
TOTAL
5.95E− 4
SS R
1.83E− 7
TOTAL
5.06E− 7
UNC L
1.67E− 7
SCC
6.18E− 4
TOTAL
2.46E− 7
ACR R
7.63E− 7
CGC L
3.61E− 7
CC
7.59E− 4
IFO L
3.99E− 7
SS R
9.68E− 7
EC L
4.56E− 7
CP R
1.27E− 3
Fx/ST L
9.60E− 7
IFO L
1.17E− 6
Fx/ST L
6.33E− 7
CGH L
1.42E− 3
ACR R
1.17E− 6
Fx/ST L
1.28E− 6
CGH R
1.42E− 6
BCC
1.46E− 3
CGH R
1.32E− 6
IFO R
1.91E− 6
RLIC L
2.04E− 6
# sig/43
24
(p < 2.78E− 2)
43
(p < 1.76E− 2)
43
(p < 3.66E− 2)
42
(p < 2.37E− 2)
b) CDR-sob
SCC
1.02E− 5
CGH L
2.21E− 12
CGH L
2.34E− 12
CGH L
1.47E− 11
CC
1.38E− 5
IFO R
1.06E− 9
IFO R
4.29E− 9
SS L
4.52E− 9
PTR L
4.49E− 5
SS L
2.96E− 9
SS L
9.54E− 9
SS R
9.72E− 9
GCC
8.80E− 5
IFO L
6.57E− 9
IFO L
1.98E− 8
IFO R
1.36E− 8
TOTAL
9.75E− 5
CGC L
9.75E− 9
TOTAL
6.20E− 8
TOTAL
3.48E− 8
BCC
1.09E− 4
SS R
3.75E− 8
CGC L
6.69E− 8
IFO L
5.56E− 8
Fx/ST L
1.28E− 4
TOTAL
3.80E− 8
ACR L
1.10E− 7
CGC L
9.34E− 8
TAP L
2.17E− 4
ACR L
1.44E− 7
SS R
2.16E− 7
CC
2.57E− 7
CGH L
2.75E− 4
CC
1.48E− 7
CC
2.68E− 7
PCR L
2.79E− 7
ACR L
4.61E− 4
SCR L
2.23E− 7
SCR L
4.76E− 7
UNC L
4.09E− 7
# sig/43
25
(p < 2.71E− 2)
42
(p < 3.59E− 2)
42
(p < 2.83E− 2)
40
(p < 4.22E− 2)
c) ADAS-cog
CGH L
1.68E− 4
CGH L
3.26E− 12
CGH L
2.14E− 12
CGH L
6.56E− 11
PTR L
4.28E− 4
SS L
2.09E− 10
SS L
5.39E− 10
SS L
4.55E− 10
SCC
6.74E− 4
SS R
3.16E− 9
SS R
2.43E− 8
SS R
6.36E− 10
Fx/ST L
8.57E− 4
IFO R
1.91E− 8
IFO R
1.09E− 7
EC L
3.63E− 8
GCC
1.32E− 3
IFO L
8.13E− 8
IFO L
6.38E− 7
IFO R
1.19E− 7
CC
1.71E− 3
CGC L
1.24E− 7
CGC L
7.75E− 7
TOTAL
2.01E− 7
SFO R
1.74E− 3
TOTAL
5.02E− 7
TOTAL
1.30E− 6
IFO L
2.99E− 7
ACR R
1.75E− 3
EC L
2.49E− 6
ACR R
3.45E− 6
CGC L
6.08E− 7
PTR R
3.29E− 3
CGC R
3.61E− 6
PTR L
1.06E− 5
UNC L
8.48E− 7
TAP L
3.35E− 3
ACR R
5.54E− 6
ACR L
1.36E− 5
SLF L
9.20E− 7
# sig/43
17
(p < 1.8E− 2)
41
(p < 3.62E− 2)
41
(p < 4.70E− 2)
39
(p < 3.80E− 2)
The strongest associations with CDR-sob score
(Table 4b)
across all 4 measures were found in the entire corpus callosum, the left
hippocampal part of the cingulum, and the total ROI map. All three
diffusivity measures additionally associated with CDR-sob in the left
cingulum, and bilaterally in the sagittal stratum and inferior
fronto-occipital fasciculus.The strongest associations with the ADAS score
(Table 4c)
across all 4 measures were found in the left hippocampal part of the
cingulum, and in all 3 diffusivity measures in the left cingulum,
bilaterally in the sagittal stratum and inferior fronto-occipital
fasciculus, and in the total ROI map.As with voxel-wise measures, we further assessed the
association between cognitive scores and ROI DTI measures in just the
MCI group as a whole and also in the e-MCI and l-MCI groups separately
(Table 5). As would
be predicted, we found significant negative MMSE associations with
diffusivity in both the MCI group and e-MCI subgroup in the left
hippocampal part of the cingulum. We found significant positive
associations between MD and AxD and CDR-sob scores in the MCI group and
no associations with ADAS-cog.
Table 5
Atlas ROI average anisotropy and diffusivity. Here we
highlight regions with the “top 10” FDR significant (Benjamini and Hochberg, 1995)
p-values (greatest effect sizes) when assessing
cognitive test score associations within diagnostic subgroups. We also note the
total number of significant ROIs (after FDR correction) out of the 43 tested and
the critical FDR p-value.
FA
MD
RD
AxD
ROI
p-Value
ROI
p-Value
ROI
p-Value
ROI
p-Value
MMSEMCI
–
–
CGH L
2.65E− 6
CGH L
5.02E− 6
CGH L
3.50E− 6
# sig/43
0
1
(p < 2.65E− 6)
1
(p < 5.02E− 6)
1
(p < 3.50E− 6)
MMSEe-MCI
–
–
CGH L
9.37E− 5
CGH L
1.52E− 4
CGH L
1.03E− 4
# sig/43
0
1
(p < 9.37E− 5)
1
(p < 1.52E− 4)
1
(p < 1.03E− 4)
CDR-sobMCI
–
–
SLF R
4.15E− 4
–
–
SLF R
6.94E− 5
–
–
–
–
–
–
SLF L
2.98E− 4
–
–
–
–
–
–
PCR R
1.03E− 3
–
–
–
–
–
–
SCC
1.64E− 3
–
–
–
–
–
–
CGC L
3.46E− 3
# sig/43
0
1
(p < 4.15E− 4)
0
5
(p < 3.46E− 3)
Again, CDF plots of the ROI
p-values for cognitive score associations
(Fig. 2b)
confirm that FA was the least sensitive measure.
ROI analyses summary
Full ROIs that were consistently most significant across
all 3 diffusivity measures in more than one analysis (Fig. 6) included the splenium of the CC, the left cingulum,
particularly the hippocampal part, the left fornix
(crus)/stria
terminalis which projects to the dorsal region of the
hippocampus, bilateral temporal lobe sagittal stratum, and the average
measure across all ROIs. They also included the bilateral inferior
fronto-occipital fasciculus in clinical score analyses. Of these ROIs,
the left fornix (crus)/stria
terminalis and hippocampal cingulum, and total ROI map
were most significant across all anisotropy and diffusivity measures in
at least two analyses. However, the left hippocampal part of the
cingulum was in the top 10 ROIs in by far the most analyses.
Fig. 6
Of the 43 full ROIs, these 9 ROIs were consistently more
sensitive to detecting differences across diffusivity and anisotropy measures in
at least two cognitive and/or diagnostic analyses. These graphs show the average
FA, MD, RD and AxD in the ROIs in each diagnostic group. “TOTAL” refers to the
ROI generated by combining all the atlas ROIs. Overall, we see a decrease in
anisotropy and increase in diffusivity with each stage of the
disease.
For complete tables of TBSS ROI
diagnostic group and cognitive score analyses results, please see
Inline
Supplementary Tables S2–S4. To summarize, TBSS ROIs
that were consistently most significant (i.e., giving greatest effect
sizes) across all four anisotropy and diffusivity measures in more than
one analysis included the splenium of the CC, bilateral hippocampal part
of the cingulum, the left sagittal stratum and uncinate, and the average
measure across all ROIs (Inline
Supplementary Fig. S1). While full ROI analyses
additionally revealed bilateral inferior fronto-occipital fasciculus,
left fornix (crus)/stria
terminalis, left cingulum, and right sagittal stratum,
as top ROIs and TBSS did not, TBSS ROIs additionally revealed the left
uncinate and right hippocampal part of the cingulum. For a comparison of
TBSS and full ROI analyses p-value CDF plots see
Inline Supplementary Fig.
S2.For complete tables of TBSS ROI
diagnostic group and cognitive score analyses results, please see Inline
Supplementary Tables S2–S4. To summarize, TBSS ROIs that were
consistently most significant (i.e., giving greatest effect sizes)
across all four anisotropy and diffusivity measures in more than one
analysis included the splenium of the CC, bilateral hippocampal part of
the cingulum, the left sagittal stratum and uncinate, and the average
measure across all ROIs (Inline Supplementary Fig. S1). While full ROI
analyses additionally revealed bilateral inferior fronto-occipital
fasciculus, left fornix (crus)/stria
terminalis, left cingulum, and right sagittal stratum,
as top ROIs and TBSS did not, TBSS ROIs additionally revealed the left
uncinate and right hippocampal part of the cingulum. For a comparison of
TBSS and full ROI analyses p-value CDF plots see
Inline Supplementary Fig. S2.Atlas TBSS ROI average
anisotropy and diffusivity. While many regions
were statistically significant (Benjamini and Hochberg,
1995), here we highlight regions
with the “top 10” significant
p-values (greatest effect
sizes) when comparing mean values between
diagnostic groups. We also note the total number
of significant ROIs (after FDR correction) out of
the 43 tested and the critical FDR
p-value.Atlas TBSS ROI average
anisotropy and diffusivity. Here we highlight
regions with the 10 lowest FDR significant
(Benjamini and Hochberg, 1995)
p-values (greatest effect
sizes) when assessing cognitive test score
associations in the entire study population. We
also note the total number of significant ROIs
(after FDR correction) out of the 43 tested and
the critical FDR
p-value.Atlas TBSS ROI average
anisotropy and diffusivity. Here we highlight
regions with the 10 lowest FDR significant
(Benjamini and Hochberg, 1995)
p-values (greatest effect
sizes) when assessing cognitive test score
associations in MCI subjects. We also note the
total number of significant ROIs (after FDR
correction) out of the 43 tested and the critical
FDR
p-value.Of the 43 TBSS ROIs,
these 8 ROIs were consistently more sensitive to
detecting differences across diffusivity and
anisotropy measures in at least two cognitive
and/or diagnostic analyses. These graphs show the
average FA, MD, RD and AxD in the TBSS ROIs in
each diagnostic group. “TOTAL” refers to the ROI
generated by combining all the atlas ROIs.
Overall, we see a decrease in anisotropy and
increase in diffusivity with each stage of the
disease.Cumulative
distribution function (CDF) plots of the
distribution of the
p-values obtained from TBSS
and Full ROI linear regression analyses, subjected
to multiple comparisons correction using standard
FDR (Benjamini and Hochberg, 1995). If
the CDF initially rises at a rate steeper than 20
times the null CDF (y = 20x), then the corresponding maps are
declared to be FDR significant at q = 0.05. The larger
the deviation is from this line, the greater the
effect sizes. (a) NC and AD group comparisons, (b)
NC and MCI group comparisons, and (c) clinical
cognitive test associations in both full and TBSS
ROIs reveal that FA is consistently the least
sensitive measure.Inline Supplementary Tables
S2–S4, Figs. S1 and S2 can be found online at http://dx.doi.org/10.1016/j.nicl.2013.07.006.
Post-hoc ROI correlations with memory and
executive function scores
The strongest associations between full ROI measures and
ADNI-EF (Table 6a) in the
entire study cohort (n = 143) across
all four DTI measures were found in the bilateral sagittal stratum and
the total ROI map. All three diffusivity measures were additionally
associated with ADNI-EF in the bilateral cingulum, left hippocampal
cingulum, and left superior longitudinal fasciculus.
Table 6
Cognitive correlates of the average anisotropy and
diffusivity in different atlas regions of interest. Here we highlight regions
with the “top 10” FDR significant (Benjamini and Hochberg, 1995)
p-values (greatest effect sizes) when assessing executive
function (ADNI-EF) and memory (ADNI-MEM) composite neuropsychological score
associations in the entire study population. We also note the total number of
significant ROIs (after FDR correction) out of the 43 tested and the critical
FDR p-value.
FA
MD
RD
AxD
ROI
p-Value
ROI
p-Value
ROI
p-Value
ROI
p-Value
a) ADNI-EF
SS L
5.77E− 5
SS L
1.85E− 9
SS L
9.15E− 10
SS L
4.21E− 8
TOTAL
1.38E− 4
SS R
8.12E− 9
SS R
8.81E− 9
SS R
4.78E− 8
GCC
2.46E− 4
CGC L
4.68E− 8
CGC L
1.39E− 7
CGC L
1.55E− 6
SS R
3.42E− 4
CGC R
3.10E− 7
CGH L
3.29E− 7
CGC R
1.75E− 6
PTR R
4.33E− 4
CGH L
6.04E− 7
CGC R
1.62E− 6
RLIC L
2.17E− 6
ACR L
4.66E− 4
RLIC L
1.46E− 6
TOTAL
1.63E− 6
TOTAL
4.02E− 6
CC
4.83E− 4
TOTAL
1.61E− 6
IFO R
7.14E− 6
CGH L
5.08E− 6
Fx/ST L
7.55E− 4
IFO R
3.31E− 6
ACR L
9.49E− 6
EC L
5.85E− 6
PTR L
7.90E− 4
SLF L
4.27E− 6
ACR R
1.08E− 5
SCC
7.35E− 6
SCC
8.84E− 4
EC L
8.76E− 6
SLF L
1.85E− 5
SLF L
1.21E− 5
# sig/43
28
(p < 2.39E− 2)
43
(p < 3.24E− 2)
43
(p < 3.27E− 2)
41
(p < 3.49E− 2)
b) ADNI-MEM
–
–
CGH L
1.38E− 9
CGH L
1.68E− 9
CGH L
5.28E− 9
–
–
SS L
4.70E− 8
SS L
1.26E− 7
SS L
7.59E− 8
–
–
CGC L
4.14E− 7
IFO R
2.41E− 6
CGC L
2.88E− 7
–
–
IFO R
5.05E− 7
CGC L
4.24E− 6
SS R
4.83E− 7
–
–
IFO L
1.03E− 6
IFO L
8.30E− 6
EC L
8.94E− 7
–
–
SS R
2.21E− 6
SS R
9.67E− 6
IFO R
1.40E− 6
–
–
CGC R
4.87E− 6
CGC R
3.68E− 5
IFO L
1.44E− 6
–
–
EC L
2.14E− 5
EC L
2.22E− 4
CGC R
3.15E− 6
–
–
CGH R
1.33E− 4
CGH R
3.36E− 4
UNC L
4.44E− 5
–
–
UNC L
1.65E− 4
UNC L
3.69E− 4
CGH R
5.00E− 5
# sig/43
0
35
(p < 3.06E− 2)
33
(p < 3.35E− 2)
35
(p < 3.51E− 2)
There were no associations between FA and ADNI-MEM
(Table 6b) in
the entire cohort, but all three diffusivity measures were associated
with ADNI-MEM bilaterally in the cingulum and hippocampal cingulum,
bilateral inferior fronto-occipital fasciculus and sagittal stratum and
left uncinate and external capsule.These measures were not linked to DTI measures in the
MCI group as a whole, nor in the e-MCI and l-MCI groups
separately.
Discussion
In this ADNI study, we found that DTI indicators of white matter
impairment have the potential to emerge as useful clinical tools for
differentiating diagnostic groups in studies of AD. We had three main findings:
(1) in ROI and VBA analyses, we found widespread anisotropy and diffusivity
disruptions, often in tracts that pass through the temporal lobe and posterior
brain regions (especially the left hippocampal cingulum), in elderly AD and MCI
patients; (2) these disruptions were also associated with neuropsychological and
cognitive deficits; and (3) diffusivity (MD, RD, AxD) measures were more
sensitive for detecting differences than FA measures, and could detect more
subtle MCI differences, where FA could not.Alzheimer's disease (AD) is characterized by neuronal loss and
widespread gray matter atrophy, but it is also marked by a disturbance in the
brain's WM pathways, perhaps secondary to the effect of cortical neuronal loss.
Changes in white matter neuropathology include partial loss of axons, myelin
sheaths, and oligodendroglial cells (Brun and Englund, 1986; Sjobeck et al., 2005).
ADpatients have been shown to have significantly more WM hyperintensities (WMH)
than controls. WMH are significantly related to cortical atrophy (Capizzano et al., 2004) entorhinal
cortex (Guzman et al., in
press), and hippocampal atrophy (de Leeuw et al., 2004) in ADpatients.
Significant WM atrophy has also been reported (Hua et al., 2008; Hua et al., 2010;
Migliaccio et al., 2012).There is growing diffusion imaging evidence of AD related WM
changes as well. Most studies report lower FA and higher MD in all lobes of the
brain in both MCI and AD when compared to cognitively healthy controls, with
emphasis on medial temporal lobe structures (Kantarci et al., 2001; Bozzali et al., 2002; Takahashi et al.,
2002; Fellgiebel et al., 2004; Medina et al., 2006; Rose et al., 2006; Xie
et al., 2006; Zhang et al., 2007; Kavcic et al., 2008; Stebbins and Murphy,
2009), consistent with regions showing earliest pathological
changes (Braak and
Braak, 1991; Braak and Braak, 1995; Thompson et al., 2007)
and our results. Many DTI methods have been used to assess WM differences in AD
including TBSS (Smith et al.,
2006) and tract or connectivity analyses (Daianu et al., 2012; Daianu et al.,
accepted for publication; Daianu et al., submitted for
publication). Here we use the most common methods, VBA and
ROI analyses, with ROIs based on a stereotaxic WM ROI atlas (Mori et al., 2008), which has been
shown to accurately parcellate anatomical regions in ADpatients with severe
atrophy (Oishi et al.,
2009). Using predetermined, template-based ROIs can limit
findings and registration across subjects may vary due to morphological
differences. However, using a common atlas helps to make quantification
efficient and easier to standardize across sites (Jahanshad et al., 2013). Given that VBA findings
are diffuse, ROIs are arguably helpful to begin to rank the most sensitive
regions for detecting group differences.When comparing ADpatients to NC, the CC
splenium, the left fornix (crus)/stria
terminalis, and average measure across all ROIs where in the top
10 most significant (lowest p-values) ROIs across all
anisotropy and diffusivity measures. However, we found that the most significant
ROIs were not necessarily the most coherent tracts. Large clinical DTI studies
with limited spatial and directional resolution are often better powered to find
deficits in regions where the FA and fiber coherence is highest, such as the
splenium of the CC. In a supplementary test, we assessed whether the average of
the FA ROI values in the NC group was correlated with the effect size of each
ROI (which is related to the p-values obtained) in the NC
vs AD comparison (see supplementary text). Inline Supplementary Fig. S3 shows that the
average FA ROI values in the NC group was not correlated with the effect size
(defined by the Z-score corresponding to the p-values) of
each ROI. VBA revealed similar diffuse patterns, with the largest differences
found in the corpus callosum, temporal lobe, cingulum and hippocampal cingulum,
and regions near the posterior thalamic radiations. Only diffusivity measures
were sensitive enough to detect more subtle differences between MCI groups and
NC. The splenium of the corpus callosum, left tapetum, left hippocampal part of
the cingulum and left fornix (crus)/stria
terminalis consistently differed in all diffusivity measures but
not FA. Only the left hippocampal part of the cingulum was significant across
all three diffusivity measures in both comparisons of e-MCI and l-MCI to
controls. Similarly, only the left hippocampal part of the cingulum displayed
significant increased diffusivity in l-MCI in the VBA analyses.When comparing ADpatients to NC, the
largest VBA differences were found in the corpus callosum, temporal lobe,
cingulum and hippocampal cingulum, and regions near the posterior thalamic
radiations. ROI analyses revealed similar diffuse patterns with the CC splenium,
the left fornix (crus)/stria
terminalis, and average measure across all ROIs in the top 10
most significant (lowest p-values) ROIs across all anisotropy and diffusivity
measures. However, we found that the most significant ROIs were not necessarily
the most coherent tracts. Large clinical DTI studies with limited spatial and
directional resolution are often better powered to find deficits in regions
where the FA and fiber coherence is highest, such as the splenium of the CC. In
a supplementary test, we assessed whether the ROI average FA values in the NC
group was correlated with the effect size of each ROI (which is related to the
p-values obtained) in the NC vs AD comparison (see
supplementary text). Inline Supplementary Fig. S3 shows that the ROI average FA
values in the NC group were not correlated with the effect size (defined by the
Z-score corresponding to the p-values) of each
ROI.Only diffusivity measures were sensitive enough to detect more
subtle differences between NC and MCI groups. ROIs including the splenium of the
corpus callosum, left tapetum, left hippocampal part of the cingulum and left
fornix (crus)/stria terminalis
consistently differed in all diffusivity measures but not FA. Only the left
hippocampal part of the cingulum was significant across all three diffusivity
measures in both comparisons of e-MCI and l-MCI to controls. Similarly, only the
left hippocampal part of the cingulum displayed significant increased
diffusivity in l-MCI in the VBA analyses.Average FA for each
ROI across all normal control (NC) participants.
ROIs are ordered from those showing the most
significant differences (lowest
p-value) in average FA
between ADpatients and NC to least significant
differences. ROIs that show a significant
difference (critical
p < 0.023; shown
in green) are not necessarily those
with the highest average FA, compared to the
non-significant ROIs
(gray). In other words, the
effect sizes we find are not simply driven by
higher FA values or fiber
coherence.Inline Supplementary Fig. S3 can be
found online at http://dx.doi.org/10.1016/j.nicl.2013.07.006.The lack of larger differences between the MCI and normal
control groups is surprising in light of other studies showing that these groups
often differ on other neuroimaging measures. This may be due the small sample
size and also heterogeneity in the MCI cohort. The MCI group referred to in
papers using ADNI1 data is now called late MCI (l-MCI) in ADNI2. In ADNI2, the
majority of the MCI group includes the enrollment of a new cohort called early
MCI (e-MCI), with milder episodic memory impairment than the l-MCI group. Among
the MCI subjects with DTI data available, a smaller percent will be l-MCI than
in phase one of ADNI, perhaps contributing to the apparent discrepancy with
other ADNI studies/results. This difference in the ADNI2 versus ADNI1 and other
MCI cohorts should be considered in expectations of what the effect sizes should
be.Strongest ROI associations with cognitive scores in the full
cohort were consistently found in the left hippocampal part of the cingulum
across all anisotropy and diffusivity measures. The total ROI map was also one
of the top 10 lowest p-values in almost all FA and
diffusivity analyses, while left cingulum and bilateral sagittal stratum were in
the lowest 10 in all diffusivity analyses. The bilateral inferior
fronto-occipital fasciculus was in the lowest 10 in almost all full cohort
diffusivity analyses. When it came to picking up subtle differences within the
MCI group, CDR-sob scores were most widely correlated with diffusivity in both
VBA and ROI analyses. Prior studies also suggest that CDR-sob is a more
sensitive clinical assessment than ADAS-cog and MMSE, and it may better relate
to measures of atrophy on anatomical MRI (Hua et al., 2009).Post-hoc analyses were conducted to assess if there was any
association between DTI summary measures and executive function (ADNI-EF) and
memory (ADNI-MEM) scores. These analyses revealed some of the strongest
associations in temporal lobe tracts that subserve a variety of functions known
to deteriorate during aging, including the bilateral sagittal stratum, that
carries parts of the inferior fronto-occipital fasciculus and inferior
longitudinal fasciculus, the hippocampal cingulum as well as the cingulum.
ADNI-EF associations were additionally found in the total ROI and superior
longitudinal fasciculus, while ADNI-MEM was associated with DTI measures in the
external capsule, inferior fronto-occipital fasciculus and uncinate fasciculus.
These regional differences could shed light on the network of brain regions,
connected via white matter fiber bundles, involved in each task. Prior DTI
studies have linked similar regions to performance in executive function and
memory (Kantarci et al., 2011;
Sasson et al., 2013).As patterns of differences were often diffuse across the VBA WM
map, the total WM ROI was sometimes one of the 10 lowest
p-values, but it was never the absolute best.
Therefore assessing individual ROIs may offer slightly more power to detect
changes than a global summary measure (Jahanshad et al., 2013). In general, the most significant
ROIs were found mainly in the left hemisphere (for example, the left hippocampal
part of the cingulum was in the top 10 ROIs in by far the most analyses), which
has also been found in some prior MRI and DTI studies (Fox et al., 1996;
Thompson et al., 2001; Scahill et al., 2002; Thompson et al., 2003; Muller
et al., 2005). A supplementary analysis revealed no evidence
of greater variability in ROI measures in the right hemisphere that may be
causing the better performance of left hemisphere ROIs (see supplementary text).
TBSS ROI results did not differ dramatically from the full average ROI results
and left hippocampal cingulum was again ‘top 10’ in the most analyses; however,
it was closely followed by the right hippocampal cingulum.Aside from the total ROI and bilateral IFO, the ROIs that were
consistently significant in at least two analyses (Fig. 6) corroborate a pattern of degeneration in
the temporal lobe and posterior temporo-parietal circuitry found in many other
DTI studies of MCI and AD (Head et al., 2004; Stahl et al., 2007; Chua et al., 2008;
Stebbins and Murphy, 2009); these include limbic tracts in
the parahippocampal white matter, posterior cingulum, fornix, and splenium of
the CC, and have been linked to lower cognitive scores (Rose et al., 2000; Bozzali et al., 2002; Takahashi et al., 2002;
Yoshiura et al., 2002; Fellgiebel et al., 2004; Fellgiebel et al., 2005;
Duan et al., 2006; Medina et al., 2006; Rose et al., 2006; Zhang et al.,
2007; Medina and Gaviria, 2008; Mielke et al., 2009). DTI
studies have found lower anisotropy in the white matter pathway of the cingulate
gyrus, particularly the posterior cingulum, which connects to the entorhinal
cortex and plays a role in the cholinergic system, known to be impaired in AD
(Perry, 1980; Selden et al., 1998; Takahashi et al., 2002; Medina et al.,
2006; Zhang et al., 2007). In fact it has been highly
implicated in studies comparing MCI to controls, especially on the left as in
our study (Fellgiebel et al., 2005; Medina et al., 2006; Rose et al., 2006; Zhang et
al., 2007; Chua et al., 2008). Cognitive scores like MMSE
have also been linked to DTI diffusivity values in the posterior cingulate gyrus
in AD (Yoshiura et al., 2002;
Fellgiebel et al., 2005). Most AD studies also find WM
deficits in the corpus callosum, but it is not clear whether the genu is more
affected (Xie et al.,
2006), in line with the hypothesis that later maturing
regions are the first affected, or the splenium, corroborating a pattern of
degeneration in the posterior temporal parietal circuitry (Takahashi et al., 2002; Medina et
al., 2006; Duan et al., 2006). A review by Chua et al. (2008) reports that in
healthy aging, DTI abnormalities occur in the frontal regions, specifically the
frontal white matter, anterior cingulum and the genu of the corpus callosum,
while in AD DTI abnormalities are concentrated in the posterior regions. Some
studies have dissociated the regional effects of age and dementia, with age
effects greater in the anterior corpus callosum and frontal white matter,
suggesting an anterior-to-posterior gradient, while individuals with early-stage
dementia exhibit minimal additional frontal deficits, but rather show greater
white matter deterioration in posterior lobar regions (Head et al., 2004).As in prior studies of aging and AD
(Sullivan et al., 2010;
Acosta-Cabronero et al., 2010), FA was the
least sensitive measure when comparing diagnostic
groups and cognitive score associations in both ROI and VBA analyses
(Fig. 2). A recent
study found that FA measured using TBSS was more strongly associated with
spectroscopic measurements than FA measured using voxel-wise averaging which is
susceptible to partial-averaging artifacts (Wijtenburg et al., 2012). However, we found that
TBSS ROI results were not substantially different from data computed from
averaging diffusion indices in ROIs (Inline Supplementary Fig. S2). A recent AD DTI review –
based on 55 different studies – noted that MD values have more discriminative
power than FA values and higher effect sizes for case–control differences in the
frontal, parietal, occipital and temporal lobes (Clerx et al., 2012). To explain age related
increases in MD without significant FA changes, Zhang et al. (2011) suggested that brain
degeneration in aging may be caused, in part, by tissue damage due to processes
such as focal ischemia. This may result in lower tissue density, increasing
water diffusivity but maintaining underlying directional structure. FA and MD
are summary measures based on the ratio and mean of the eigenvalues
respectively, but AxD may reflect axonal injury, and RD may reflect
demyelination (Song et al., 2003; Song et al., 2005; Sun et al., 2006; Hofling et al.,
2009). Here, when comparing NC to ADpatients in both ROI
and VBA analyses, RD was the most strongly associated with WM deficits
(Fig. 2a–b), followed
by MD. Increased RD in AD relative to controls, therefore, may reflect
demyelination in AD. As in prior studies that reported AxD increases in normal
aging and Alzheimer's disease (Fellgiebel et al., 2004;
Sullivan et al., 2010; Acosta-Cabronero et al., 2010; Agosta et al.,
2011), we also found higher AxD in MCI, AD, and associated
with clinical impairment. In fact, we found slightly larger effect sizes for AxD
when comparing CN and l-MCI subjects in the VBA analyses and CN and MCI subjects
in the full ROI analyses (Table 3b). Recent longitudinal and cross-sectional studies
suggest that AxD might be more sensitive to detecting early changes, while RD
becomes progressively better as disease progresses (O'Dwyer et al., 2011; Acosta-Cabronero et
al., 2012). In a study by Bendlin et al. (2012),
cerebrospinal fluid biomarkers of AD, T-Tau and Aβ42, predicted
axial and radial diffusivity, suggesting these CSF biomarkers are potentially
related to both neuronal axon integrity and health of oligodendrocyte
synthesized myelin. The relationship may have a common cause, if cerebral
amyloid and tau burden is related to CSF measures and to axonal loss and poorer
myelination. Ultimately, the positive relationships between AD and both AxD and
RD may explain why FA, a function of the ratio of these measures, reveals the
least differences (Acosta-Cabronero et
al., 2010), and could not pick up more subtle deficits in MCI
(Fig. 2b,c;
Table 3b–d,
Table 4).As in prior studies of aging and AD
(Sullivan et al., 2010;
Acosta-Cabronero et al., 2010), FA was the
least sensitive measure when comparing diagnostic
groups and cognitive score associations in both ROI and VBA analyses
(Fig. 2). A recent
study found that FA measured using TBSS was more strongly associated with
spectroscopic measurements than FA measured using voxel-wise averaging which is
susceptible to partial-averaging artifacts (Wijtenburg et al., 2012). However, we found that
TBSS ROI results were not substantially different from data computed from
averaging diffusion indices in ROIs (Inline Supplementary Fig. S2). A recent AD
DTI review – based on 55 different studies – noted that MD values have more
discriminative power than FA values and higher effect sizes for case–control
differences in the frontal, parietal, occipital and temporal lobes
(Clerx et al.,
2012). To explain age related increases in MD without significant
FA changes, Zhang et al.
(2011) suggested that brain degeneration in aging may be
caused, in part, by tissue damage due to processes such as focal ischemia. This
may result in lower tissue density, increasing water diffusivity but maintaining
underlying directional structure.FA and MD are summary measures based on the ratio and mean of
the eigenvalues respectively, but AxD may reflect axonal injury, and RD may
reflect demyelination (Song et al., 2003; Song et al., 2005; Sun et al., 2006; Hofling
et al., 2009). Here, when comparing NC to ADpatients in
both ROI and VBA analyses, RD was the most strongly associated with WM deficits
(Fig. 2a–b), followed
by MD. Increased RD in AD relative to controls, therefore, may reflect
demyelination in AD. As in prior studies that reported AxD increases in normal
aging and Alzheimer's disease (Fellgiebel et al., 2004;
Sullivan et al., 2010; Acosta-Cabronero et al., 2010; Agosta et al.,
2011), we also found higher AxD in MCI, AD, and associated
with clinical impairment. In fact, we found slightly larger effect sizes for AxD
when comparing CN and l-MCI subjects in the VBA analyses, and CN and MCI
subjects in the full ROI analyses (Table 3b). Recent longitudinal and cross-sectional studies
suggest that AxD might be more sensitive to detecting early changes, while RD
becomes progressively better as disease progresses (O'Dwyer et al., 2011; Acosta-Cabronero et
al., 2012). Ultimately, the positive relationships between
AD and both AxD and RD may explain why FA, a function of the ratio of these
measures, reveals the least differences (Acosta-Cabronero et al., 2010), and could not
pick up more subtle deficits in MCI (Fig. 2b,c; Table 3b–d, Table 4).In addition, a summary measure such as FA, derived from a simple
tensor model, may not capture the complexity of white matter architecture. DTI
has some limitations in gauging fiber integrity in regions with extensive fiber
crossing and mixing. High angular diffusion imaging (HARDI) can better resolve
white matter multi-fiber distribution than DTI single-tensor models
(Leow et al., 2009; Zhan et
al., 2009). More sophisticated measures, such as a modified
version of FA calculated from the HARDI data “tensor distribution function”
(TDF), can better characterize the anisotropy in regions of fiber crossings
(Zhan et al., 2009),
and may better reveal FA differences that are due to loss in fiber coherence
rather than simply partial-volume effects.Further limitations of this study include the
small sample size from each of the 14 sites, and unequal distribution of cases
across sites (see Inline Supplementary
Table S1). Despite accounting for site affects using a random
effects regression model and grouping the data by acquisition site, some
cross-site differences may be unaccounted for in the study design. Additionally
we did not adjust for WMH burden, or explore the extent the DTI metrics
corresponded with WMH. Finally, the ADNI data set is constantly re-evaluating
subject diagnosis, cognitive, and basic demographic data. The data evaluated
here was what was available at the time of download and may be changed by ADNI
at any time.Further limitations of this study
include the small sample size from each of the 14 sites, and unequal
distribution of cases across sites (see Inline Supplementary Table S1). Despite
accounting for site effects using a random-effects regression model and grouping
the data by acquisition site, some cross-site differences may be unaccounted for
in the study design. Additionally we did not adjust for WMH burden, or explore
the extent the DTI metrics corresponded with WMH. Finally, the ADNI data set is
constantly re-evaluating subject diagnosis, cognitive, and basic demographic
data. The data evaluated here was what was available at the time of download and
may be modified by ADNI.
Conclusion
DTI offers an extensive set of biomarkers for disease detection
and monitoring of cognitive decline. We found anisotropy and even more
widespread diffusivity disruptions in projection, association, and commissural
tracts, particularly in tracts that pass through the temporal lobe and posterior
brain regions (especially the left hippocampal cingulum), in elderly AD and MCI
patients. These disruptions were associated with neuropsychological and
cognitive deficits. As ADNI2 progresses, new subjects are scanned and new
measures of WMH are being added to the database. These results should be
verified with larger sample sizes and the relationship between DTI based WM
disturbances and WMH can be further resolved. Future tract-based connectivity
studies may also shed light on how integrity in the tracts affects regional gray
matter. We may also be able to create a statistical region-of-interest that may
outperform atlas based ROIs (Hua et al., 2013; Gutman et al., 2013). ADNI2 is a
longitudinal study that will eventually allow us to investigate which of these
subjects develop AD, and if these early WM aberrations help predict future
deficits and conversion to AD. Future studies combining machine learning methods
with different modalities – including CSF and proteomic markers – may ultimately
determine the best way to distinguish diagnostic groups.
Table S1
Distribution of study
participants by diagnostic group across each of
the 14 acquisition sites. We report the
p-values from a
χ2 test
between the actual and expected distribution of
subjects, which we estimated based on the total
number of subjects in each diagnostic group, as a
proportion of the total, across sites. Although
numbers for some sites are small, there is no
evidence that any site is enrolling a
statistically divergent proportion of subjects in
each category.
Site
Total
NC
MCI
AD
χ2p-value
1
16
9
3
4
0.74
2
4
0
3
1
0.81
3
12
5
6
1
0.95
4
14
3
7
4
0.93
5
14
1
12
1
0.84
6
8
0
6
2
0.81
7
11
6
4
1
0.85
8
3
0
3
0
0.68
9
14
5
7
2
0.99
10
12
5
5
2
0.95
11
12
1
10
1
0.86
12
11
0
9
2
0.82
13
16
5
9
2
1.00
14
8
4
4
0
0.85
Table S2
Atlas TBSS ROI average
anisotropy and diffusivity. While many regions
were statistically significant (Benjamini and Hochberg,
1995), here we highlight regions
with the “top 10” significant
p-values (greatest effect
sizes) when comparing mean values between
diagnostic groups. We also note the total number
of significant ROIs (after FDR correction) out of
the 43 tested and the critical FDR
p-value.
FA
MD
RD
A × D
ROI
p-Value
ROI
p-Value
ROI
p-Value
ROI
p-Value
a)NCvsAD
SCC
2.61E− 6
CGH L
4.68E− 12
CGH L
4.41E− 12
CGH L
7.44E− 11
CGH L
3.89E− 6
SS L
3.22E− 8
SS L
1.07E− 7
TAP L
3.69E− 7
Fx/ST L
4.76E− 5
CGH R
1.59E− 7
UNC L
1.10E− 7
CGH R
6.30E− 7
ACR L
1.95E− 4
UNC L
2.17E− 7
CGH R
1.79E− 7
IFO L
7.56E− 7
SFO R
3.70E− 4
SCC
2.63E− 7
SCC
3.32E− 7
SS L
9.69E− 7
Fx/ST R
3.81E− 4
Fx/ST R
4.00E− 7
Fx/ST L
3.86E− 7
SCR L
1.15E− 6
CC
3.90E− 4
SFO L
8.70E− 7
Fx/ST R
5.40E− 7
UNC L
1.27E− 6
SS L
4.74E− 4
TOTAL
1.30E− 6
SFO L
6.21E− 7
TOTAL
1.36E− 6
SFO L
5.27E− 4
IFO L
1.44E− 6
TOTAL
3.23E− 6
SCC
2.28E− 6
TOTAL
6.34E− 4
Fx/ST L
1.99E− 6
SFO R
9.86E− 6
SFO L
2.44E− 6
#
sig/43
22 (p < 1.99E− 2)
41 (p < 4.68E− 2)
36 (p < 3.15E− 2)
35 (p < 3.94E− 2)
b)NCvsMCI
–
–
CGH L
9.87E− 5
CGH L
1.12E− 4
CGH L
3.64E− 4
–
–
CGH R
8.16E− 4
CGH R
9.99E− 4
CGH R
9.53E− 4
#
sig/43
0
2
(p < 8.16E− 4)
2
(p < 9.99E− 4)
2
(p < 9.53E− 4)
c)NCvse-MCI
–
–
CGH L
8.54E− 4
CGH L
8.00E− 4
RLIC L
7.63E− 4
–
–
CGH R
9.37E− 4
CGH R
1.46E− 3
CGH R
1.08E− 3
–
–
–
–
–
–
CGH L
3.32E− 3
–
–
–
–
–
–
CGC R
3.91E− 3
#
sig/43
0
2
(p < 9.37E− 4)
2
(p < 1.46E− 3)
4
(p < 3.91E− 3)
d)NCvsl-MCI
–
–
CGH L
6.54E− 6
CGH L
1.66E− 5
CGH L
1.41E− 5
#
sig/43
0
1
(p < 6.54E− 6)
1
(p < 1.66E− 5)
5(p < 1.41E− 5)
Table S3
Atlas TBSS ROI average
anisotropy and diffusivity. Here we highlight
regions with the 10 lowest FDR significant
(Benjamini and Hochberg, 1995)
p-values (greatest effect
sizes) when assessing cognitive test score
associations in the entire study population. We
also note the total number of significant ROIs
(after FDR correction) out of the 43 tested and
the critical FDR
p-value.
FA
MD
RD
A × D
ROI
p-Value
ROI
p-Value
ROI
p-Value
ROI
p-Value
a)
MMSE
CGH L
2.87E− 5
CGH L
3.12E− 12
CGH L
1.65E− 12
CGH L
5.28E− 10
CGH R
7.55E− 5
UNC L
8.18E− 9
UNC L
3.10E− 9
UNC L
8.36E− 8
ACR L
1.56E− 4
CGH R
5.68E− 8
CGH R
3.37E− 8
SS L
4.47E− 7
ACR R
2.46E− 4
SS L
7.43E− 8
SS L
9.68E− 7
TOTAL
8.91E− 7
UNC L
6.83E− 4
TOTAL
2.18E− 6
ACR L
2.96E− 6
CGH R
1.17E− 6
CGC R
8.43E− 4
CGC L
2.60E− 6
Fx/ST R
3.51E− 6
SS R
2.69E− 6
TOTAL
8.48E− 4
IFO L
3.14E− 6
ACR R
3.76E− 6
SCR L
5.88E− 6
Fx/ST R
9.37E− 4
Fx/ST R
3.38E− 6
CGC L
6.53E− 6
TAP L
6.45E− 6
CP R
1.01E− 3
ACR L
5.58E− 6
CGC R
6.87E− 6
IFO L
1.90E− 5
CC
1.24E− 3
ACR R
6.06E− 6
Fx/ST L
2.85E− 5
CC
1.98E− 5
#
sig/43
22 (p < 2.43E− 2)
38 (p < 3.72E− 2)
34 (p < 3.22E− 2)
35 (p < 3.44E− 2)
b)
CDR-sob
CGH L
3.28E− 6
CGH L
2.86E− 11
CGH L
6.23E− 12
CGH R
2.05E− 9
CGH R
8.87E− 6
CGH R
3.66E− 11
CGH R
2.19E− 11
CGH L
1.20E− 8
SCC
2.01E− 5
SS L
1.43E− 7
UNC L
1.83E− 7
TOTAL
7.95E− 7
CC
1.23E− 4
UNC L
5.52E− 7
SS L
5.05E− 7
SCC
8.78E− 7
UNC L
1.29E− 4
SCC
6.97E− 7
TOTAL
2.39E− 6
SCR R
2.05E− 6
TOTAL
2.25E− 4
TOTAL
8.68E− 7
SCC
2.58E− 6
SS R
3.65E− 6
Fx/ST R
4.14E− 4
CC
3.33E− 6
Fx/ST R
2.93E− 6
SS L
3.68E− 6
CP L
4.25E− 4
SS R
3.89E− 6
ACR L
1.13E− 5
TAP L
4.12E− 6
SS L
5.32E− 4
Fx/ST R
4.84E− 6
CC
1.22E− 5
SCR L
4.80E− 6
GCC
6.98E− 4
CGC L
5.22E− 6
CGC R
1.71E− 5
IFO L
4.92E− 6
#
sig/43
24 (p < 2.36E− 2)
38 (p < 2.89E− 2)
37 (p < 4.18E− 2)
34 (p < 1.48E− 2)
c)
ADAS-cog
CGH L
3.39E− 6
CGH L
4.98E− 13
CGH L
1.85E− 13
CGH L
2.01E− 10
CGH R
4.72E− 4
CGH R
1.38E− 8
CGH R
1.47E− 8
UNC L
8.50E− 8
SS L
6.63E− 4
UNC L
3.98E− 8
UNC L
3.55E− 8
CGH R
1.68E− 7
SCC
1.17E− 3
SS L
8.17E− 8
SS L
4.98E− 7
SS L
5.17E− 7
ACR R
2.21E− 3
IFO L
1.21E− 6
Fx/ST R
2.33E− 6
TOTAL
1.95E− 6
Fx/ST R
2.79E− 3
CGC R
1.26E− 6
CGC R
5.10E− 6
SS R
3.10E− 6
Fx/ST L
2.81E− 3
Fx/ST R
1.70E− 6
CGC L
1.72E− 5
IFO L
3.85E− 6
CC
3.14E− 3
SS R
2.15E− 6
TOTAL
2.59E− 5
SLF L
8.29E− 6
TOTAL
3.19E− 3
CGC L
2.47E− 6
Fx/ST L
4.92E− 5
SCC
8.48E− 6
GCC
3.62E− 3
TOTAL
7.14E− 6
SS R
5.48E− 5
IFO R
9.62E− 6
#
sig/43
16 (p < 1.65E− 2)
39 (p < 3.76E− 2)
35 (p < 3.210E− 2)
35 (p < 1.48E− 2)
Table S4
Atlas TBSS ROI average
anisotropy and diffusivity. Here we highlight
regions with the 10 lowest FDR significant
(Benjamini and Hochberg, 1995)
p-values (greatest effect
sizes) when assessing cognitive test score
associations in MCI subjects. We also note the
total number of significant ROIs (after FDR
correction) out of the 43 tested and the critical
FDR
p-value.
Authors: Gloria C Chiang; Philip S Insel; Duygu Tosun; Norbert Schuff; Diana Truran-Sacrey; Sky Raptentsetsang; Clifford R Jack; Michael W Weiner Journal: Radiology Date: 2011-04-05 Impact factor: 11.105
Authors: K Kantarci; M L Senjem; R Avula; B Zhang; A R Samikoglu; S D Weigand; S A Przybelski; H A Edmonson; P Vemuri; D S Knopman; B F Boeve; R J Ivnik; G E Smith; R C Petersen; C R Jack Journal: Neurology Date: 2011-05-18 Impact factor: 9.910
Authors: Rachael I Scahill; Jonathan M Schott; John M Stevens; Martin N Rossor; Nick C Fox Journal: Proc Natl Acad Sci U S A Date: 2002-04-02 Impact factor: 11.205
Authors: Talia M Nir; Neda Jahanshad; Julio E Villalon-Reina; Dmitry Isaev; Artemis Zavaliangos-Petropulu; Liang Zhan; Alex D Leow; Clifford R Jack; Michael W Weiner; Paul M Thompson Journal: Magn Reson Med Date: 2017-03-07 Impact factor: 4.668
Authors: Ali Ezzati; Mindy J Katz; Michael L Lipton; Molly E Zimmerman; Richard B Lipton Journal: Brain Imaging Behav Date: 2016-09 Impact factor: 3.978
Authors: Pauline Maillard; Evan Fletcher; Baljeet Singh; Oliver Martinez; David K Johnson; John M Olichney; Sarah T Farias; Charles DeCarli Journal: Neurology Date: 2019-04-05 Impact factor: 9.910
Authors: Elizabeth E Moore; Timothy J Hohman; Faizan S Badami; Kimberly R Pechman; Katie E Osborn; Lealani Mae Y Acosta; Susan P Bell; Michelle A Babicz; Katherine A Gifford; Adam W Anderson; Lee E Goldstein; Kaj Blennow; Henrik Zetterberg; Angela L Jefferson Journal: Neurobiol Aging Date: 2018-06-28 Impact factor: 4.673
Authors: Farshid Sepehrband; Ryan P Cabeen; Jeiran Choupan; Giuseppe Barisano; Meng Law; Arthur W Toga Journal: Neuroimage Date: 2019-04-30 Impact factor: 6.556
Fig. 1
Fig. 2
Fig. 3
Fig. 4
Fig. 5
Fig. 6