Recent imaging evidence in Alzheimer's disease suggests that neural involvement in early-stage disease is more complex than is encapsulated in the commonly held position of predominant mesial temporal lobe degeneration—there is also early posterior cingulate cortex and diencephalic damage. These findings suggest that early clinical Alzheimer's disease is underpinned by damage to an inter-connected network. If correct, this hypothesis would predict degeneration of the white matter pathways that connect this network. This prediction can be tested in vivo by diffusion magnetic resonance imaging. Most diffusion tensor imaging studies of white matter in neurodegenerative disorders such as Alzheimer's disease have concentrated on fractional anisotropy reductions and increased ‘apparent’ diffusivity; however, there is a lack of empirical biological evidence to assume that fractional anisotropy changes will necessarily capture the full extent of white matter changes in Alzheimer's disease. In this study, therefore, we undertook a comprehensive investigation of diffusion behaviour in Alzheimer's disease by analysing each of the component eigenvalues of the diffusion tensor in isolation to test the hypothesis that early Alzheimer's disease is associated with degeneration of a specific neural network. Using tract-based spatial statistics, we performed voxel-wise analyses of fractional anisotropy, axial, radial and mean diffusivities in 25 Alzheimer's disease patients compared with 13 elderly controls. We found that increased absolute (axial, radial and mean) diffusivities in Alzheimer's disease were concordant in a distribution consistent with the network hypothesis, highly statistically significant and far more sensitive than fractional anisotropy reductions. The former three measures identified confluent white matter abnormalities in parahippocampal gyrus and posterior cingulum, extending laterally into adjacent temporo-parietal regions as well as splenium and fornix. The caudal occipital lobe, temporal pole, genu and prefrontal white matter were relatively preserved. This distribution is highly consistent with expected predictions of tract degeneration from grey matter lesions identified by fluorodeoxyglucose positron emission tomography and structural magnetic resonance imaging. Concordant with results from these other imaging modalities, this pattern predominantly involves degeneration of the tracts connecting the circuit of Papez. These findings also highlight that early neuropathological processes are associated with changes of the diffusion ellipsoid that are predominantly proportional along all semi-principal axes.
Structural and metabolic imaging studies in early Alzheimer's disease have highlighted several abnormalities in cortical and subcortical grey matter. Hippocampal atrophy measured by volumetric magnetic resonance imaging (MRI) is an established finding and is present in incipient (mild cognitive impairment stage) disease (Jack et al., 1999; Pengas et al., in press); comparable atrophy is also evident in the adjacent entorhinal cortex (Du et al., 2004). Metabolic imaging with (18F)-2-fluoro-deoxy-d-glucose positron emission tomography (FDG-PET) also confirms mesial temporal lobe hypometabolism in incipient Alzheimer's disease (De Santi et al., 2001; Nestor et al., 2003), but suggests that posterior cingulate hypometabolism is the most severe metabolic lesion in early disease (Minoshima et al., 1997; Nestor et al., 2003). Although posterior cingulate hypometabolism was originally interpreted as being a remote consequence of mesial temporal lobe degeneration, recent volumetric MRI work has highlighted that this area is also atrophic in incipient Alzheimer's disease (Choo et al., in press; Pengas et al., in press). As posterior cingulate and mesial temporal lobe regions are connected via the circuit of Papez, it has been proposed that neuronal degeneration in early Alzheimer's disease may specifically involve this neuronal network (Nestor et al., 2003). Indeed sub-cortical elements of this circuit (mamillary bodies and anterior thalamus) are also hypometabolic in incipient disease (Nestor et al., 2003). The hypothesis that this is a true network degeneration would predict involvement of the white matter tracts connecting these structures (i.e. parahippocampal white matter, posterior cingulum bundle and fornix).
Mapping changes in white matter tracts is presently a major research focus for understanding development, ageing and disease of the central nervous system. The potential of diffusion MRI as a non-invasive neuroimaging tool was first highlighted when Moseley et al. (1990) noted hyper-intensity in diffusion-weighted acquisitions at the site of ischaemic injury. This became known as the ‘light bulb’ effect and has found clinical application as an early marker of stroke. A challenge in dealing with this type of acquisition is the anisotropic nature of diffusion in white matter: water molecules diffuse faster along fibre tracts than perpendicular to them due to barriers such as axonal membrane and myelin sheath—this results in an image contrast that is dependent on the spatial characteristics of the applied diffusion weighting. To characterize this dependence better, Basser et al. (1994) introduced the concept of diffusion tensor imaging (DTI)—(for review see Assaf and Pasternak, 2008). In this formulation, the data at each point are modelled using a second-rank tensor, visualized by a diffusion ellipsoid that represents the three-dimensional (3D) character of water molecules’ diffusion. The tensor is rotationally invariant and is always aligned to the main diffusion direction at a given voxel; thus it can be decomposed into three mutually orthogonal eigenvectors and three real eigenvalues: in particular, the diffusion coefficient along the direction of maximal ‘apparent’ diffusion (λ1 or axial diffusivity), and the diffusion coefficients along two orthogonal directions embedded in the plane perpendicular to the main diffusion direction (λ2 and λ3). The second and third eigenvalues can be averaged and expressed as so-called radial diffusivity (RD).1994). In addition, the average diffusivity, known as mean diffusivity (MD) can be inferred from the overall dimensions of the diffusion ellipsoid.
DTI potentially enables mapping of white matter tract alterations in development, ageing and neurological disorders and has therefore become an important tool in the study neurodegeneration in Alzheimer's disease (Takahashi et al., 2002; Fellgiebel et al., 2004, 2008; Medina et al., 2006; Huang et al., 2007; Ringman et al., 2007; Zhang et al., 2007, 2009; Smith et al., in press; Damoiseaux et al., 2009; Mielke et al., 2009). The vast majority of studies to date have focused on FA changes (Chua et al., 2008), typically without consideration of the component eigenvalues. If changes in diffusion along the direction of the semi-major axis of the ellipsoid (λ1) were proportional to those of the semi-minor axes (RD), however, then FA—which is a function of this ratio—would remain relatively unchanged. We hypothesized that this might be the case in early Alzheimer's disease, in which—unlike stroke where FA changes were originally noted—there may be greater changes in the absolute dimensions of the diffusion ellipsoid rather than in its shape. This prediction would translate to concordant and more sensitive results for λ1, RD and MD when compared with FA.
In summary, the aim of this study was to use DTI measures of white matter tracts to investigate whether there was supporting evidence for the prediction that early Alzheimer's disease is associated with degeneration of structures in an inter-connected limbic–diencephalic neuronal network. Although, intuitively, FA may seem an appropriate metric to study white matter changes in neurodegenerative disease, there is neither any empirical evidence to indicate that this is the most sensitive measure nor is there evidence to suggest that other tensor behaviours could be superior in capturing the white matter tract changes in Alzheimer's disease. As such, a comprehensive analysis of diffusion measures was undertaken.
The hypothesis was explored in a cohort of 25 patients with early-stage probable Alzheimer's disease according to Dubois criteria (Dubois et al., 2007) compared with 13 age- and sex-matched controls (Table 1). All controls were screened to exclude neurological or major psychiatric illness and performed normally on a global cognitive measure [the Addenbrooke's cognitive examination—revised or ACE-R (Mioshi et al., 2006)]. Written informed consent was obtained from all the participants and the study was approved by the local regional ethics committee.
|Gender, M:F||15:10||5:8||NS [χ2(1) = 1.59]|
|Age at imaging, years||69.7 (6.3)||67.1 (5.5)||NS|
|Education, years||13.1 (2.6)||14.0 (2.7)||NS|
|MMSE at imaging/30||22.9 (4.1)||29.0 (1.2)||<0.0001|
|ACE-R at imaging/100||69.5 (12.5)||94.6 (3.8)||<0.0001|
|Gender, M:F||15:10||5:8||NS [χ2(1) = 1.59]|
|Age at imaging, years||69.7 (6.3)||67.1 (5.5)||NS|
|Education, years||13.1 (2.6)||14.0 (2.7)||NS|
|MMSE at imaging/30||22.9 (4.1)||29.0 (1.2)||<0.0001|
|ACE-R at imaging/100||69.5 (12.5)||94.6 (3.8)||<0.0001|
Values are given as mean (SD).
M = male; F = female; NS = not significant. MMSE = Mini-mental state examination; ACE-R = Addenbrooke's cognitive examination–revised
All patients were scanned within an average of 0.9 months (SD = 1.5 months) from a cognitive assessment. MRI scans were performed on a Siemens Trio 3T system (Siemens Medical Systems, Erlangen, Germany), equipped with gradient coils capable of 45 mT/m and slew rate of 200 T/m/s. A 12 channel phased-array total imaging matrix head-coil (Siemens Medical Systems, Erlangen, Germany) was used with twice-refocused single-shot echo-planar imaging (Reese et al., 2003): repetition time (TR) / echo time (TE) / number of excitations (NEX) = 7800 ms/90 ms/1; matrix, 96 × 96; 63 contiguous axial slices; isotropic voxel resolution of 2 mm × 2 mm × 2 mm; bandwidth of 1628 Hz/pixel and echo spacing of 0.72 ms. The diffusion tensor was acquired with diffusion-sensitizing gradient orientations along 63 non-collinear directions (b = 1000 s/mm2) that were maximally spread by considering the minimal energy arrangement of point charges on a sphere (http://www.research.att.com/~njas/electrons/dim3), and with one scan without diffusion weighting (b = 0 s/mm2, b0). We allowed for parallel acquisition of independently reconstructed images using generalized autocalibrating partially parallel acquisitions [GRAPPA; (Griswold et al., 2002), acceleration factor of 2 and 39 reference lines], hence reducing the final scanning time to 8 min and 44 s.
In addition, true 3D, high-resolution, T2-weighted images were acquired to ensure that vascular pathology was not significant in either cohort. The turbo spin echo with variable flip-angle distribution (sampling perfection with application optimized contrasts using different flip angle evolution; SPACE) pulse sequence was used with the following scan parameters: TR/TE/NEX = 3200 ms/450 ms/2; matrix, 192 × 192; 144 slices; voxel resolution, isotropic 1.25 mm × 1.25 mm × 1.25 mm. GRAPPA acceleration with a factor of 2 was also implemented, allowing a scan time of 5 min and 9 s.
T1-weighted anatomic images were also acquired to evaluate regional differences in grey-matter density using voxel-based morphometry (Ashburner and Friston, 2000). The acquisition consisted of a 3D magnetization-prepared rapid gradient-echo pulse sequence: TR/TE/inversion time (TI)/flip angle = 2300 ms/2.86 ms/900 ms/9°, 144 slices, 192 × 192 matrix dimensions and 1.25 mm × 1.25 mm × 1.25 mm voxel size. Receiver bandwidth and echo spacing were 240 Hz/pixel and 6.7 ms, respectively; scan time was 7 min and 23 s.
The entire protocol was acquired during a single 25 min session, where the subject's brain was aligned in stereotactic space: the anterior commissure–posterior commissure line was aligned with the axial plane, and the inter-hemispheric fissure was aligned along the sagittal plane and at right angles to the coronal plane. The scanner isocentre was placed as near as possible to the centre of the thalamus in the mid-sagittal plane.
The Functional MRI of the Brain (FMRIB) software library (FSL v4.0.3) (Smith et al., 2004) was employed to process and analyse the raw DTI data. First, each diffusion-weighted volume was affine-aligned to its corresponding b0 image using the FMRIB's linear image registration tool (FLIRT v5.4.2) (Jenkinson and Smith, 2001); this pre-processing step corrects for possible motion artefacts and eddy-current distortions. Prior to fitting the tensor, brain masks of each b0 image were generated using the brain-extraction tool (BET v2.1) (Smith, 2002) with fractional threshold, f = 0.1, and vertical gradient, g = 0. FMRIB's diffusion toolbox (FDT v2.0) was then used to fit the tensor and compute the diagonal elements (λ1, λ2 and λ3), FA, RD and MD at each brain voxel.
An additional problem with DTI in degenerative diseases, however, is that atrophy can lead to systematic misregistration to template for a disease group when compared with controls, such that changes in diffusion measures [either within a region of interest (ROI) or a whole-brain voxel-based analysis] may be artefactual (Snook et al., 2007]. In other words, due to anatomical differences across subjects or in the presence of atrophy, a given voxel can be contaminated by greater inclusion of adjacent white matter tracts, grey matter or cerebrospinal fluid; this will, in turn, give rise to apparent changes in diffusivity that may not necessarily be due to true diffusion alterations in the tract of interest. In order to circumvent this problem, the tract-based spatial statistics (TBSS) approach (Smith et al., 2006) suggests a method by which the nearest most relevant tract centre in each subject's spatially normalized FA image can be projected onto a skeleton containing the centre of all major tracts ‘common’ to all subjects; thereby allowing voxel-wise statistics to be carried out at the tract centres only, minimizing the effect of misregistration.
As opposed to its earlier version, where every FA image was co-registered to every other one, and re-warped to the subject with the minimum mean displacement prior to affine-alignment into MNI152 standard space (Montreal Neurological Institute, McGill, USA), TBSS v1.1 provided a 1 mm isotropic FA target image (FMRIB58_FA) in standard space, to which all FA images were non-linearly warped using the image registration toolkit (Rueckert et al., 1999). This resulted in warped versions of each subject's FA image, which were then averaged, turned into a white matter tract skeleton and thresholded at FA > 0.2 to further limit the effects of miswarping across subjects. The skeleton was created by (i) suppressing all non-maximum FA values in each voxel's local perpendicular direction and (ii) comparing each remaining non-zero voxel with its closest neighbours, thus searching for the centre of fibre bundles. Finally, all subjects’ normalized FA data were projected onto the mean FA skeleton and fed into voxel-wise statistics. Data for λ1, RD and MD were also generated by applying the FA non-linear registration to the additional parametric maps, and projecting them onto the skeleton using identical projection vectors to those inferred from the original FA data.
TBSS uses permutation-based non-parametric inference on unsmoothed statistical maps; this method delivers better spatial localization of abnormalities than assuming normal distribution of smoothed DTI parametric maps. Non-parametric, two-sample, unpaired t-tests of reduced and increased DTI indices in patients compared with controls were performed using ‘randomise v2.1’ (included in FSL v4.1); randomise computes permutation tests on the assumption that the null hypothesis implies complete exchangeability of the observations (Nichols and Holmes, 2002). In this study, we generated 5000 permutations of the data to test against, at threshold levels: (i) P < 0.01 uncorrected for multiple comparisons, (ii) false discovery rate (FDR), q < 0.05, which controls the proportion of false rejections of the null hypothesis (type I errors) among all significant voxels, and (iii) family-wise error (FWE) corrected P < 0.05, which controls the rate of type I errors for all voxels collectively by using the null distribution of the maximal voxel-wise test statistic. FWE does not allow for control over the proportion of false positives and is usually unnecessarily conservative, particularly when data smoothing is involved; however, permutation testing offers asymptotically exact control over the false positive rate, thereby adapting easily to control the FWE rate. In addition, cluster-like structures were enhanced using a recently proposed method known as threshold-free cluster enhancement (Smith and Nichols, 2009).
Grey matter analysis
Recently, Acosta-Cabronero et al. (2008) found that voxel-based morphometry using the statistical parametric mapping release, SPM5 (Ashburner and Friston, 2005), can be improved by skull-stripping and bias-correcting magnetic resonance images prior to being introduced in its generative model. In this study, all scans were pre-processed using the following automated pipeline: first, skull-stripping was performed using the hybrid watershed algorithm (Ségonne et al., 2004) in FreeSurfer v.4.05 (http://surfer.nmr.mgh.harvard.edu), which integrates an atlas-based term constraining the shape of the brain; stripped volumes were then bias-corrected using the non-parametric non-uniform intensity normalization algorithm or N3 v.1.10 (Sled et al., 1998) with default arguments; and finally, a fine brain extraction that excludes venous sinuses and cerebrospinal fluid was performed using BET v.2.1 with fixed arguments: f, set to 0.2 and g, set to 0. All volumes were then spatially normalized and segmented using the unified segmentation model in SPM5. The segments were also modulated to compensate for volumetric differences introduced into the warped images. Modulation has the effect of preserving the total amount of grey matter represented by multiplying by the relative volumes. Finally, grey matter segments were smoothed using an 8 mm full width at half maximum isotropic Gaussian kernel.
A group analysis was performed to compare regional differences in grey matter density between patients and control subjects. Sex was entered into the statistical model as a nuisance co-variate. In order to maintain consistency with the permutation-based statistical approach of TBSS, the smoothed grey matter segments were assessed using the standard, permutation-based non-parametric multiple comparisons procedure: statistical non-parametric mapping. One thousand permutations of the data were generated with variance smoothing of 8 mm full width at half maximum, masked by the relative threshold of 0.2 and statistical threshold of q < 0.05 (FDR).
TBSS results for increased FA and reduced diffusivities in 25 mild Alzheimer's disease patients compared with 13 elderly controls did not show any statistically significant difference at P < 0.01 uncorrected for multiple comparisons. The reduced FA contrast (Fig. 1), however, identified small discrete areas of significant change in the white matter of the right temporal lobe, right posterior cingulate region and right parieto-occipital region. FA reductions were also identified in the fornix (see Fig. 1: sagittal slice x = 0 mm; coronal slice y = –4 mm; axial slice z = 10 mm) as well as two small areas in the right cerebellar hemisphere and ponto-medullary junction. Results of the analyses are shown at uncorrected P < 0.01 overlaid onto the MNI152 template and the mean FA skeleton. The results for increased λ1, RD and MD (Fig. 1) were broadly concordant with each other and far more extensive than for FA. Axial and mean diffusivity clusters of significance were slightly more extensive than those for RD. Axial, radial and mean diffusivity contrasts identified similar significant changes to those seen with FA in fornix and cerebral white matter, but not in the brainstem or cerebellum. Unlike FA, however, changes in λ1, RD and MD were bilateral and confluent involving the white matter of caudal cingulum bundle, parahippocampal gyrus and spreading into adjacent lateral white matter areas of the caudal temporal and parietal regions. The caudal corpus callosum was also abnormal with all three measures, whereas there was relative sparing of caudal occipital lobe, temporal pole, prefrontal white matter and rostral corpus callosum. The rendered volumes and overlays shown in Fig. 2, using axial diffusivity, illustrate in detail this distribution of white matter abnormalities.
At a FDR of 0.05, the λ1, RD and MD FDR-corrected P thresholds (0.0088, 0.0082 and 0.0094, respectively) were almost equivalent to the ad hoc uncorrected significance level proposed above (Fig. 3A). The more conservative FWE rate of 0.05 (Fig. 3B) led to an expected decrease in sensitivity, but overall produced very similar clusters of significance to those for uncorrected and FDR-corrected critical values. In contrast, all FA suprathreshold voxels were expected to be false positives at these multiple comparison correction thresholds. Repeating the analyses after independently controlling for gender, age and education level difference yielded identical results.
The distribution of grey matter abnormalities in this cohort from the voxel-based morphometry analysis was in agreement with the above TBSS result (Fig. 4). Loss of grey matter density in the Alzheimer's disease group was most significant in the hippocampus, followed by entorhinal region. This extended to a less significant extent into temporal neocortex and insula. The posterior cingulate cortex was also significantly atrophic.
The λ1, RD and MD findings using the TBSS method offer what we propose is the most comprehensive view of the landscape of white matter tract degeneration in early Alzheimer's disease to date; and offers strong evidence that early Alzheimer's disease is associated with specific degeneration of a limbic–diencephalic neuronal network. Mesial temporal lobe atrophy and posterior temporo-parietal association cortex hypometabolism are established features of Alzheimer's disease (Nestor et al., 2004). Studies in recent years have identified focal posterior cingulate hypometabolism as the earliest physiological lesion (Minoshima et al., 1997); this area was subsequently found to be as significantly atrophic as the hippocampus in very early Alzheimer's disease (Pengas et al., in press). Evidence of lesions in mesial temporal and posterior cingulate regions was extended by the observation that hypometabolism in these regions as well as diencephalic structures (anterior thalamus and mamillary body) was specifically associated with very early stage Alzheimer's disease (Nestor et al., 2003). This led to the conjecture that these areas were degenerating as part of a functionally and anatomically integrated system (Nestor et al., 2003, 2006). These various observations would predict that the landscape of axonal tract degeneration in early Alzheimer's disease would include white matter of the parahippocampal gyrus extending to, and continuing along, the posterior cingulum bundle; that there would be extension of white matter degeneration from these areas into lateral (especially) posterior temporo-parietal areas; that preferential degeneration of these posterior associations cortices would lead to preferential degeneration of inter-hemispheric white matter connections passing through the caudal, as opposed to rostral, corpus callosum; that fibres connecting mesial temporal lobe and diencephalon via the fornix would degenerate; and that there would be relative preservation of white matter in prefrontal, temporal pole and caudal occipital areas. The findings for λ1, RD and MD, although not to such an extent for FA, identified precisely this distribution of white matter changes.
Consistent with this network hypothesis that links cortical degeneration and white matter pathways, the voxel-based morphometry results from the present cohort confirmed that the patient group had atrophy of the temporal lobes (especially hippocampus), and, to a lesser degree, the posterior cingulate cortex. Greater hippocampal involvement might suggest that this region is the primary driver of network degeneration—an interpretation consistent with early neurofibrillary tangle deposition in the mesial temporal lobe (Braak and Braak, 1991). It should be noted, however, that voxel-based morphometry may lack sensitivity in the posterior cingulate region; for instance, recently two independent manual volumetric studies of very early Alzheimer's disease (scanned at the mild cognitive impairment stage) both found significant posterior cingulate atrophy (Choo et al., in press; Pengas et al., in press) that was at least as severe as that of the hippocampus; pre-symptomatic Alzheimer's disease patients (autosomal dominant mutation carriers) have accelerated rates of atrophy using the serial fluid co-registration technique in both posterior cingulate and hippocampus but in no other regions (Scahill et al., 2002); and, finally, cortical thickness measurements in Alzheimer's disease suggest focal thinning of the posterior cingulate/precuneus region compared with the remainder of the mesial (inter-hemispheric) cortical surface (Du et al., 2007).
A recent study highlighted a remarkable homology between the atrophy profile in Alzheimer's disease (mesial temporal-posterior cingulate/precuneus-lateral temporo-parietal cortex) and an intrinsic functional connectivity network in healthy volunteers derived from resting state functional (f)MRI (Seeley et al., 2009). The present DTI findings provide very strong support for a direct relationship between these two observations, in that the anatomical connections between these structures were precisely those that were preferentially involved in the current study.
The current results are also of interest to contrast with the default mode network hypothesis in Alzheimer's disease. It has been noted that there is a striking homology between regions of high resting state fMRI activity in healthy volunteers (default mode) and the distribution of amyloid (using Pittsburgh Compound-B-PET) in Alzheimer's disease patients: in each the highlighted regions are posterior cingulate/precuneus/lateral temporo-parietal cortex and ventral frontal cortex (Buckner et al., 2005). Synaptic dysfunction in Alzheimer's disease, as measured with FDG-PET, is only partially consistent with this distribution, in that the posterior components are abnormal early whereas the frontal component is not a prominent early feature. As already discussed, the present DTI results sided with FDG-PET in demonstrating preferential degeneration of the limbic–diencephalic network—the nodes of which had been shown previously to be specifically hypometabolic in very early Alzheimer's disease (Nestor et al., 2003); the present results did not, however, identify significant ventral prefrontal white matter changes. Bringing these streams together suggests that there is not simply a stepwise progression from local resting state activity in health, to local amyloid deposition and to local neuronal degeneration. Assuming (i) that default mode activity and amyloid distribution are not simply coincidental and (ii) that their overlap is an important early trigger to a cascade of events that causes dementia in Alzheimer's disease, one then has to reconcile why the neural network that degenerates first in Alzheimer's disease is only partially overlapping. One possibility could be that although amyloid deposits in both places, there are additional factors that make the posterior regions selectively vulnerable to neurodegeneration and that this, in turn, involves the whole limbic–diencephalic network. If this speculation were correct then identifying those additional factors would be of considerable importance. The anatomical relationship of posterior cingulate to mesial temporal lobe suggests that one potential candidate is that this is the pathway by which, as Price and Morris (1999) speculated, the presence of amyloid turns a ubiquitous and relatively benign age-associated tauopathy towards an accelerated phase of neurofibrillary tangle formation and neuronal loss.
Turning to the methodological aspects of this study, although DTI is an established method to visualize diffusional properties of white matter, the biological underpinning of tensor changes remains incompletely understood. Using FA changes as a marker of axonal integrity imposes the assumption that degeneration will give rise to changes in the shape of the diffusion ellipsoid (Fig. 5A and B). It is widely believed, for instance, that phenomena such as demyelination and axonal loss may cause a scenario in which FA reductions are driven by enhanced radial diffusion and constant axial diffusivity coefficient (Fig. 5A) (Beaulieu, 2002; Song et al., 2002). In contrast, simultaneous axial diffusivity reduction and radial diffusivity increase is thought to be associated to fibre re-organization (Fig. 5B) (Dubois et al., 2008). It is conceivable, however, that contributions of other neurobiological processes involved in white matter tract degeneration such as destruction of neurofibrils and glial alterations can give rise to tensor behaviours that are not fully captured by anisotropy changes. FA may lack sensitivity if water diffusion changes proportionally along the direction of all three eigenvectors (Fig. 5C). Furthermore, it does neither follow that these biological processes will be the same across different pathologies nor even within a single pathology at different stages of illness: differential effects on glia, mechanisms of axonal degeneration, inflammatory responses and even the rate of degeneration could give rise to different tensor behaviours in different pathological states. In summary, exploration of the full tensor behaviour would be advantageous in studying all degenerative diseases rather than assuming that a single metric (such as FA) is sufficiently sensitive to different pathological states.
In order to explore comprehensively the behaviour of the diffusion tensor in neurodegeneration, we performed TBSS analyses, not only of FA maps, but also of axial, radial and total mean diffusities (λ1, RD and MD, respectively) in Alzheimer's disease patients compared with elderly controls. The concordance of observations for the absolute diffusivity measures and the gross lack of sensitivity of FA, caused by smaller group mean differences in the latter, are consistent with the hypothesis that early neurodegeneration in Alzheimer's disease is associated with predominantly proportional tensor variations in all three dimensions (Fig. 5C). This conclusion was underlined by the observation that changes in λ1, RD and MD, but not FA, survived correction for multiple comparisons using either FDR or, the most stringent, FWE methods. These findings suggest that FA is not the appropriate metric for studying white matter tract degeneration in Alzheimer's disease, a finding that could prove true of other neurodegenerative diseases though this would need to be demonstrated in these different diseases rather than assuming that it is the case.
It is usually assumed that the main diffusion direction is aligned to white matter tracts, and that radial diffusivities, which average diffusion perpendicular to fibre bundles, are modulated by the extracellular distance between membranes, axon diameter and degree of myelination. Consequently, it is thought that loss of fibre tracts and disruption of myelin sheath during neurodegeneration should not affect longitudinal diffusivities, and could uniquely be explained by enhanced transverse diffusion and therefore, apparent FA reductions (Beaulieu, 2002; Song et al., 2002). This simplistic approach, however, does not include the effect of other plausible biophysical phenomena in neurodegenerative disorders such as fibre re-organization, increase in membrane permeability, destruction of intracellular compartments and glial alterations (Beaulieu, 2002). The magnitude of the impact of these processes on the diffusion tensor is unknown, but it is plausible that they may induce water molecules to diffuse faster in unanticipated directions. This may be the case in this study, where we observed regions of significant increase of λ1 that were concordant to those of RD and MD, resulting in insubstantial anisotropy changes. These results highlight the importance of exploring the full power of the diffusion tensor, and indicate that changes in the size of the diffusion ellipsoid (i.e. changes in diffusion eigenvalues that may be the product of complex interactions between several neurodegenerative processes) are more sensitive and more biologically plausible than changes in the anisotropy of the ellipsoid.
It was clear, however, that λ1 increase was slightly more emphatic than the change in transverse diffusion (Fig. 3); this should lead to a slight change in FA, even though FA is inherently less sensitive because RD and λ1 changed in the same direction. Nevertheless, the slightly disproportionate effect on λ1 might be the reason why significant changes in FA have been reported in previous ROI studies of Alzheimer's disease (e.g. Mielke et al., 2009; Zhang et al., 2009). In other words, subtle changes in FA can be detected within ROIs that do not survive more stringent whole-brain statistical analysis. Notably, however, one ROI study that looked at FA and MD found that the latter measure was more sensitive within ROIs (Fellgiebel et al., 2004), consistent with the present results. Another important consideration, and the reason that ROI analyses were avoided in the present study, is that it also possible that previously reported FA changes from ROI analyses may be, to a significant degree, artefactual. Even a subtle amount of atrophy in a tract of interest could lead to a systematic mis-sampling of that tract between patient and control groups. For instance, when an ROI is applied, subtle distortion of the anatomy (secondary to atrophy) means that the ROI in the disease group could include a greater amount of an adjacent tract which would, in turn, translate to a significant FA difference that was not actually a consequence of degeneration in the tract of interest (i.e. control's ROI = tract X; patient's ROI = tract X + Y). Colour-coding of tracts prior to ROI placement has been proposed to minimize this problem (Fellgiebel et al., 2005); however, because this colour coding only highlights the three principal orthogonal planes, it is unlikely that such an approach can completely eliminate the effect of contamination from adjacent tracts which might run in any direction rather than only being orthogonal to a tract of interest.
In summary, using TBSS to minimize preprocessing and analytic inaccuracies, this study investigated changes in both the anisotropic properties of diffusion (FA) and non-fractional measures of diffusivity (λ1, RD and MD). The results indicated that the latter are more sensitive than FA in Alzheimer's disease, seriously challenging the notion that changes in axonal integrity can be simply captured by FA reductions. It is still premature to conclude that DTI can fully describe the physical phenomena that characterize white matter degeneration in Alzheimer's disease, but the results are very encouraging in that the diffusion tensor is highly sensitive to changes in mild Alzheimer's disease patients, and that these changes are highly concordant with predictions of expected white matter change derived from prior knowledge of other imaging modalities. Beyond Alzheimer's disease, these measures may equally inform the nature of neural network changes in other neurodegenerative diseases and are also of potential relevance to normal maturation and ageing. With respect to Alzheimer's disease, the findings offer compelling evidence that the limbic–diencephalic network is selectively vulnerable to the degenerative process. Understanding what makes the neurons in this network preferentially degenerate is a key question for future research. Identification of molecular differences in this neuronal network compared with others less susceptible to degeneration in Alzheimer's disease might offer a bridgehead to a deeper mechanistic understanding of the pathogenesis of this disease.
Medical Research Council, UK (to P.J.N.); Alzheimer's Research Trust (to G.P.).
We are also grateful for support from the National Institute for Health Research (Cambridge Biomedical Research Centre).
diffusion tensor imaging
(18F)-2-fluoro-deoxy-d-glucose positron emission tomography
false discovery rate
magnetic resonance imaging
region of interest
tract-based spatial statistics