Many investigators have used MRI and PET imaging to measure longitudinal progression of brain changes in normal aging, MCI and AD, with varying results. As drug candidates that might slow the progression of Alzheimer’s pathology began to be developed, the need to develop robust and sensitive imaging methods to quantify progression of Alzheimer’s disease has become increasingly important. To this end, the National Institute of Aging and pharmaceutical industry funded the Alzheimer’s Disease Neuroimaging Initiative, with the goal of developing improved methods based on imaging and other biomarkers, for AD treatment trials.

A variety of methods have been used to quantify the longitudinal changes in structural brain MRI including: region-of-interest measurements (especially of the hippocampus (Frisoni et al., 1999), the “boundary shift integral” technique which quantifies differences between two successive co-registered 3D MRIs (Fox et al., 2000), voxel-based morphometry (Good et al., 2001; Whitwell et al., 2007), and tensor-based morphometry (Studholme et al., 2004, 2006; Leow et al., 2007).

Tensor-based morphometry (TBM) is a relatively new image analysis technique that identifies regional structural differences in the brain, across groups or over time, from the gradients of the deformation fields that align, or ‘warp,’ images to a common anatomical template (reviewed in Ashburner and Friston, 2003). Highly automated methods such as TBM are being tested to examine their utility in large-scale clinical trials, and in studies to identify factors that influence disease onset and progression (Leow et al., 2005b; Cardenas et al., 2007), or normal development (Thompson et al., 2000a; Chung et al., 2001; Hua et al., 2007).

In this paper, TBM is applied to a longitudinal ADNI dataset by using a nonlinear registration algorithm to match 3D baseline structural MR images with follow-up images acquired 1 year later (for related approaches, see Leow et al. 2005a; Studholme et al., 2006; Studholme and Cardenas, 2007; van de Pol et al., 2007; Barnes et al., 2008; Ridha et al., 2008).

Color-coded Jacobian maps — which show the local expansion or compression factor at each point in the image — can be used to indicate local volume loss or gain relative to the baseline image (Freeborough and Fox, 1998; Chung et al., 2001; Fox et al., 2001; Ashburner and Friston, 2003; Riddle et al., 2004). Here we examined longitudinal brain changes, using these Jacobian maps, in groups of AD and MCI subjects relative to controls. We also investigated, at a voxelwise level, how ongoing brain atrophy correlated with clinical measures including MMSE scores, and the global Clinical Dementia Rating (CDR), as well as biomarkers of AD pathology including CSF levels of tau protein, 181-phosphorylated tau protein (p-tau), beta amyloid (ABeta 42), and tau/Abeta42 ratio (Andreasen et al. 2001; Itoh et al., 2001; Verbeek et al., 2003; Clark et al., 2003; Hampel et al., 2004; Lee and Trojanowski, 2006).

Prior studies of the disease trajectory in AD (e.g., Scahill et al., 2002; Thompson et al., 2003a,b), show a shift in the distribution of atrophy with advancing disease. In line with the trajectory of neurofibrillary pathology (Braak and Braak, 1991), the entorhinal and medial temporal lobes show the earliest signs of atrophy in MCI, with frontal atrophy typically occurring later, and primary sensory and motor cortices spared until late in the illness (see Thompson and Apostolova, 2007, for a review of this trajectory mapped with different imaging modalities). Consistent with this, in our MCI group, progressive atrophy was detected only in the temporal and parietal lobes, in line with evidence that ongoing changes are more anatomically restricted at this pre-dementia stage.

One notable aspect of the topography of brain matter loss shown in Figs. 1 and ​and22 is that the greatest proportion of brain matter loss appears to lie in the white matter rather than at the voxels on the cortical surface. There are two reasons for this, both technical: (1) the deformations are spatially smooth and partial volume averaging effects occur and diminish the signal somewhat at tissue boundaries, such as the cortex/CSF interface, and (2) the registration accuracy of TBM is poorer at the cortical surface, at least relative to some approaches that explicitly model the cortical surface. To clarify this, note that Figs. 1 and ​and22 visualize group differences by averaging rates of volumetric changes (i.e., Jacobian maps), after nonlinearly aligning individual maps of change to the minimal deformation template (MDT). This deformation field is spatially smooth, so some partial volume effects between cortex and CSF are inevitable and are more pronounced along the MDT boundary. As a result, some signal spillover from outside of the brain tissue may be present, explaining the reduction in the atrophy signal along the boundary. This effect has been noted in our prior work (Hua et al., 2008), where the disease-related expansion in the ventricles spills over into the subcortical white matter by about 1–2 mm, in the average maps. Similarly, the cortical atrophy signal is partially canceled by the signal in the CSF outside the brain, which may not show the same level of atrophy, and if anything, may show slight expansion over time. Second, and perhaps more importantly, the deformation fields are based on automated matching of intensities in the images, and the spatial smoothness of the fields makes it difficult to register the entire cortical mantle within subjects from one time-point to the next, as would be required to gauge the atrophy of cortical gray matter. Alternative approaches may be used that compute thickness at each point, but these are typically more time-consuming as they generally require extraction of explicit models of the cortical surface as geometric meshes, prior to computing the cortical thickness either directly from the meshes (Lerch and Evans, 2005), or by tissue classification of the images and voxel coding (Thompson et al., 2004; Aganj et al., 2008).

There are at least two possible solutions to better sensitizing our TBM approach for detecting cortical gray matter loss. The first is to use a method termed voxel-based morphometry (VBM; Ashburner and Friston, 2000) or a related approach termed RAVENS (Davatzikos et al., 2001). In VBM, the deformation-based compression signal at each point is multiplied by maps of gray matter density, which are based on smoothing maps of gray matter voxels derived from an explicit tissue classification into gray and white matter and CSF. An additional modulation step is also included that preserves information on the volume of gray matter in the baseline images after warping. When gray matter density and deformation signals are multiplied together in this way, VBM maps in AD do typically show progressive cortical gray matter atrophy in a temporal-to-frontal pattern that matches the spread of neurofibrillary tangle pathology (Baron et al., 2001). A second approach to identifying cortical gray matter atrophy with TBM was developed by (Studholme et al., 2003), in which deformation-based compression signals at each point are smoothed adaptively depending on the amount of gray matter lying under the filter kernel. This is a way to avoid some of the signal depletion that occurs when atrophying gray matter is partial-volumed with CSF. A third solution is to run the deformation maps at a very high spatial resolution and with less spatial regularization, or with a regularization term that enforces continuity but not smoothness. Because of the complexity of differentiating cortical gray matter changes from underlying white matter changes, we seek to assign signals to the cortex without surface-based modeling, which can be time consuming for larger analyses (Thompson et al., 2004), Thus, in this paper, we decided to combine both gray and white matter for each region of interest (ROI) in our analyses, instead of separating them.

Although much of the literature has suggested that gray matter loss is the primary change in AD that is observable on MRI (see Thompson and Apostolova, 2007, for a review), there has been substantial theoretical and empirical evidence supporting white matter pathology in AD (Bartzokis et al., 2003). For example, in Rose et al. (2000), the authors summarized recent DTI studies in AD, and many have reported reduced FA (fractional anisotropy) in temporal, frontal, and parietal lobes, especially in the internal capsule and limbic association fibers, the corticothalamic pathway, superior longitudinal fasciculus, and posterior cingulate bundle (Yoshiura et al., 2002). White matter degeneration in AD has also been detected with MR relaxometry (Bartzokis et al., 2003) and myelin and oligodendrocyte reductions have been detected in neuropathological studies of AD. Future MRI studies, using state-of-the-art techniques such as diffusion-weighted MRI (Rose et al., 2000; Choi et al., 2005; Medina et al., 2006) or High Angular Resolution Diffusion Imaging (HARDI), are likely to further elucidate white matter pathology in AD.

A further notable feature, which requires some explanation, is that a related technique (the voxel-compression method) has shown unequivocally temporal gray matter loss and ventricular expansion, but no change in the white matter (Fox et al., 2001) in serial MRI studies of AD. As already noted, we did indeed detect subtle and diffuse changes in the subcortical white matter, but the ability to detect them depends to some degree on the level of regularization used in TBM. In TBM, there is a smoothness term, which causes correlations in the deformation signals at neighboring voxels. In general, for simplicity and practicality, an elastic (Leow et al., 2005a) or fluid (Fox et al., 2001) model of registration is used, in which the Green’s function of the governing operator is spatially uniform and fixed. If the correlations are assumed to be long-range (i.e., the deformations are spatially quite smooth), there is more signal enhancement in large homogeneous regions such as the white matter, whereas if the correlations are assumed to be short-range (i.e., the deformations are spatially quite rough, as in the fluid registration model of Fox et al., 2001), there is typically more sensitivity to finer-scale differences (as found in the gray matter in the Fox et al. study), while sacrificing some power to detect broader-scale differences (e.g., the failure to detect white matter atrophy in Fox et al., 2001). In future, the differential sensitivity of both approaches could be combined by estimating these spatially varying correlations empirically from anatomical landmarks using 6-dimensional covariance tensors (Fillard et al., 2008) and incorporating them into a statistically-based adaptive registration model as we have begun to do (Brun et al., 2007 and 2008).

In this paper, the correlations between atrophy and CSF biomarkers are also of significant interest. Prior literature has indicated that there is lower ABeta42, but higher tau and p-tau protein, in the CSF of AD patients versus those with other dementia subtypes or normal subjects (Andreasen et al. 2001; Itoh et al., 2001; Verbeek et al., 2003; Clark et al., 2003; Hampel et al., 2004). More recently, researchers have also investigated the utility of using these markers for predicting conversion from MCI to AD (Fagan et al., 2007; Li et al., 2007). In our results, progressive temporal lobe atrophy was highly correlated with baseline p-tau, tau/ABeta42 ratio (for both AD groups and all subjects pooled), ABeta42, and tau (the latter two only for data pooled across all diagnostic groups). This suggests that p-tau and tau/Abeta42 may be more clinically useful than Abeta42 or tau in predicting ongoing atrophy (in Fig. 4, CDF curves rise more rapidly for the correlations with p-tau and with the tau/ABeta42 ratio).

We could not demonstrate significant correlations between biomarkers and ongoing temporal lobe atrophy in the MCI group. This is perhaps not surprising due to the heterogeneous nature of MCI, and the relatively small sample of 40 subjects.

Clinical measures correlated more strongly with atrophy rates in MCI than in AD, supporting the use of serial neuropsychiatric testing in monitoring disease progression in MCI. In AD, atrophy rates exceeded those in MCI, but did not correlate so strongly with interval changes in neuropsychiatric test scores. This may suggest that (1) decline in cognition is more tightly linked with atrophy rates early in the illness, or (2) in late AD, atrophy rates may eventually plateau or slow down, which may disrupt any correlation between the absolute rate of tissue loss and further changes in cognition, or (3) correlations may only be detectable in samples that are larger and/or have a broader range of disease severity. Our AD sample was only half the size of our MCI sample, and was somewhat restricted in disease severity to reflect relatively mild AD; by contrast, in recent study of 52 subjects with mild-to-moderate AD (Ridha et al., 2008), there was a strong association between brain atrophy rate and MMSE decline (r=0.59, p<0.0001). In addition, there is some evidence that atrophy rates do not slow down as AD progresses; Chan et al. (2003) found that in 12 patients with mild dementia (MMSE=23), scanned from a presymptomatic stage through to moderately severe dementia, mean yearly loss of brain volume was 2.8% (95% CI: 2.3–3.3), but rose by 0.32% per year (0.15–0.50). In 39 healthy control subjects, Scahill et al. (2003) also found rates of atrophy accelerated nonlinearly with increasing age, with the most marked changes occurring after the age of 70.

To summarize, our results further support the value of serial MRI scanning, combined with quantitative nonlinear registration, for tracking disease progression in Alzheimer’s and MCI. Our detailed 3D Jacobian maps, reflecting regional brain atrophy, correlated well with disease progression and conversion to AD, as well as with various biomarkers and clinical measures. Moreover, groups of MCI converters and non-converters were differentiated by measures of temporal lobe atrophy over time.

Lastly, instead of separating hippocampus in our analysis, it was included as part of the temporal lobe. Any TBM study is limited by the accuracy with which deformable registration can match anatomical boundaries between individual brains and corresponding regions on the template. Our mean deformation template (MDT) was created after rigorous nonlinear registration, and geometric centering. Most anatomical features and boundaries are well-preserved in the MDT, and the hippocampus is sufficiently discernible to be labeled by hand. Even so, it may not always be possible to achieve accurate regional measurements of atrophy in small regions such as the hippocampus, since that would require a locally highly accurate registration. Some research groups have successfully computed hippocampal atrophy rates from fluid registration methods (e.g., Crum et al., 2001), and found that they can be superior to manual delineations in separating AD from controls (p<0.0001; Barnes et al., 2007a,b) and more reliable (van de Pol et al., 2007).