Structural atrophy in the human brain predicts conversion from MCI to AD (for a review see (Ries et al., 2008
), (Mosconi et al., 2007
)), with atrophy rates accelerating as dementia progresses (Buckner et al., 2005
). Until recently our view of structural changes during the progression of AD has been fairly myopic – the majority of published MRI studies having focused on a limited number of structures known to degenerate in the disease, largely because the manual segmentation required to isolate regions for analysis is tedious and time-consuming. However, MRI data sets contain far more information than we currently extract. The development of automated segmentation and deformation-based mophometry methods would allow us to tap the full potential of MRI as a non-invasive, longitudinal imaging tool. Whole-brain automated analyses would provide an unbiased view of neurodegeneration and the possibility of identifying unrecognized temporal and spatial patterns of atrophy. Here we demonstrate the potential of two automated methods for the analysis of high resolution, active staining, contrast-enhanced MRI to identify wide-ranging volumetric changes in a mouse model of AD. We found that both methods detected volume loss prior to plaque formation in regions of the brain commonly measured in MRI studies of AD. More importantly, both methods also detected volume differences in brain structures well outside the areas thought to be affected in AD. In several cases, mid- and hindbrain structures identified by MR were found to develop amyloid pathology much later in the disease – the MR analyses predicted later histological abnormalities that would not have been recognized otherwise. High-resolution MR analysis has thus helped us locate degeneration and pathology in remote, but interconnected, areas through the use of automated, non-biased analyses. Our findings provide a rationale for expanding the use of automated structural analysis in human AD at very early stages, where we might find similarly unsuspected changes that broaden our view of the disease.
Given the interest in advancing the capability of MR for use in longitudinal studies of neurodegenerative disease, other groups have also used MR to view amyloid burden (Lee et al., 2004
) (Zhang et al., 2004
) (Wengenack et al., 2008
) and assess volumetric changes (Lau et al., 2008
) in the brains of APP transgenic mice. Several past studies have used manually segmented MRI to identify pre-amyloid decreases in dentate gyrus (Redwine et al., 2003
), cerebrum (Van Broeck et al., 2008
), and total brain volume (Delatour et al., 2006
), (Van Broeck et al., 2008
) of young transgenic mice. Because of the labor required to manually segment MR volumes, these studies examined only one or two (or at most five) structures, each of which had been previously implicated in AD. Our work extends this region-of-interest approach to an unbiased look at the entire brain, where we have discovered that cortical volume is significantly diminished early in disease, and with it, confirmed the decrease in total brain volume. We identified significant differences in cortical sheet thickness, cortical volume, and overall brain volume prior to the onset of amyloid plaques, suggesting that the pathogenesis of AD begins before histological lesions are apparent. These structural changes likely have functional consequences, as recent studies have shown that cognitive impairments in several mouse models can begin prior to the appearance of overt pathology (Moechars et al., 1999
), (Westerman et al., 2002
), (Kelly et al., 2003
Three recent studies have taken a more expansive look at the whole brain using automated techniques for MR analysis of APP transgenic mice. Similar to the work we present here, (Maheswaran et al., 2009
) compare automated atlas-based segmentation to deformation-based morphometry in an APP x PS1 model of AD, but start imaging after the onset of amyloid pathology. Suprisingly, they report that parts of the hippocampus, deep layers of cortex, and thalamus are all bigger, with some structures growing faster, in the double transgenic mice than in their WT controls. Our results are more consistent with those of (Lau et al., 2008
) who apply both a deformation-based approach and limited manual segmentation to longitudinal in vivo MR images. Lau et al. study a separate APP × PS1 line, with two time points prior to and two after the onset of amyloid. They examine a wide range of structures throughout the brain, and like us, find that most regions continue to grow in the DBL transgenic mice, albeit at a lower rate than in their WT siblings, with significant differences in the interaction between age and genotype detected for specific areas of the cortex, hippocampus, thalamus, striatum, and septum. In both of these studies, as well as in our own work, there is generally good correlation between the two techniques. The advantage of DBM over atlas-based segmentation is the ability to examine local changes that might be too small to impart a significant effect on the overall structure. In our study, for example, DBM identified significant changes in hippocampal volume of young DBL mice relative to controls that were not yet apparent in the automated segmentation. Similarly, DBM was able to detect changes in particular nuclei of the pons that were indistinguishable in the segmental analysis. We presented the uncorrected statistics, which are easier to interpret. An FDR correction at 0.001 level identified as significant a small number of sparse voxels, some localized in the dorsal hippocampal region, a region well-known to be affected in AD. The FDR correction proved to give an overly conservative view of the data, however we were able to use the uncorrected statistics to discover novel pathology in an independent cohort – pathology that wouldn’t have been discovered had we relied on corrected values. The generality of segmentation analysis is complemented by the local view of DBM, and conversely the complex interpretation of small changes in DBM benefits from the simple outcomes of segmentation.
Consistent with the general agreement between DBM and segmentation, both methods identified the most surprising results to come out of our automated, whole-brain volumetric analyses. While we had expected degeneration in the neocortex based on the spatial overlap with pathology, we had not anticipated significant volume differences in the pons, inferior colliculus, substantia nigra, cerebral peduncle, and spinal nuclei. The presence of atrophy in these structures may be due to the artificial pattern of Aβ production in the brains our mice; however, transgene expression under the CaMKIIα promoter is limited primarily to the forebrain and is absent from the mid- and hindbrain structures we’ve identified (Mayford et al., 1996
; Ochiishi et al., 1998
) (Kamata et al., 2006
). Although the transgene is not thought to be expressed in these regions, each connects to structures where the transgene is active. As a result of the MR findings, we revisited a timed series of amyloid stained-sections from this line to find diffuse amyloid appearing much later in disease in several (although not all) of the structures identified by MR as showing volumetric deficits 5–8 months earlier. This outcome highlights the strength of non-biased, automated MR volumetric analyses for their unique ability to assess wide-ranging changes and make predictions that inform our understanding of disease pathogenesis.
It is worth noting that not all areas of the brain were affected equally in our mice. Even though the transgene is widely expressed throughout the forebrain, the volume losses we detected were fairly restricted. Two structures from our segemental analysis that stand out for their relative preservation are the amygdala and olfactory bulbs. These areas develop amyloid pathology at approximately the same rate as the cortex and hippocampus, yet they show no significant difference in volume at either of the ages we examined. The DBM analysis suggests a somewhat more complex picture, in which local losses occurred but were not extensive enough to affect the size of the structure as a whole. Conversely, our segmental analysis identified volume loss in the cortex substantial enough to involve the entire structure, while morphometry mapped these changes to more limited sub-regions. The somatosensory cortex suffered the brunt of the damage, but other areas such as the motor, cingulate, and visual cortices were conspicuously spared. It is unclear why some regions were susceptible and others relatively resistant, but the specificity revealed through these MR analyses offers a comparison that future studies might exploit to understand how the human disease primarily affects areas involved in learning and memory while leaving other structures untouched.
Although our study has been performed in an artificial mouse model, several of our findings relate surprisingly well to what is known about the course of neurodegeneration in human AD. The losses we noted in cortex, hippocampus, and whole brain volumes are found consistently in both manually segmented and voxel-based morphological studies of human AD (reviewed in (Mosconi et al., 2007
), (Busatto et al., 2008
), (Thompson et al., 2007
), (Ries et al., 2008
)). More impressive than the overlap of atrophy in these regions commonly associated with AD is the overlap with regions not often considered in the disease. As in our AD mice, the pons can be dramatically affected in AD patients, where there are several reports of degeneration or cytoskeletal pathology in the locus ceruleus (Haglund et al., 2006
), (Strong et al., 1991
), (Busch et al., 1997
), (Lyness et al., 2003
), (Zarow et al., 2003
), (German et al., 2005
), (O’Neil et al., 2007
), pontine parabrachial nuclei, subpeduncular nucleus, tegmentopontine reticular nucleus, reticular formation, and pontine nucleus (Rub et al., 2001b
), (Rub et al., 2001a
), (Iseki et al., 1989
). Also reported, but poorly studied, is amyloid pathology in the inferior and superior colliculi of AD patients (Iseki et al., 1989
), (Leuba and Saini, 1995
), (Parvizi et al., 2001
). In contrast, involvement of the substantia nigra in AD has become so well recognized that its presence is subclassified as dementia with Lewy-bodies or Parkinson’s disease with dementia.
The relationship of pre-amyloid cortical atrophy in our mice to the human disease is somewhat more elusive: by the time cases of sporadic AD present with symptoms, many have already developed amyloid pathology (Kemppainen et al., 2007
), (Forsberg et al., 2008
). The pre-plaque stage at which our mice were imaged can only be studied in humans with autosomal dominant forms of inherited AD. Familial cases of AD are rare, but a handful of studies have examined these patients for signs of brain atrophy and loss of connectivity prior to the onset of symptoms. Longitudinal assessment of brain volume by MRI revealed significant atrophy in the entorhinal cortex and hippocampus several years before the appearance of memory impairments (Fox et al., 2001
; Fox et al., 1996
), (Schott et al., 2003
). The pre-amyloid cortical atrophy we describe in our transgenic mice is also seen in the rare cases of human AD where it can be examined with certainty of future prognosis. Thus, despite the artificial overexpression of a synthetic APP mutation, our transgenic mice reproduce many features of the human disease and serve as a rational model for examining pre- and post-amyloid volumetric changes.
The automated methods developed for analyzing mouse brain MRI data have allowed us to identify structural changes in parts of the brain we never would have considered a priori given the commonly established wisdom about degeneration in AD. These techniques show the true potential of MRI for structural studies: by extracting information from the entire data set, we’ve generated a temporal and spatial map of where atrophy begins and progresses during these early stages of disease. Future studies can now focus more closely on some of these unexpected sites of degeneration to explore what forebrain connections make them vulnerable, and whether the same regions are at risk in other mouse models of AD. More importantly, we can begin to understand what role these areas play in the progression of symptoms characteristic of the disease. For now, the tools available for automated analyses of longitudinal mouse brain MRI data provide a window into AD that the human disease cannot yet afford. But with major neuroimaging initiatives already underway for AD, we may soon have the opportunity to see the brain as broadly as we have here to confirm in human patients the novel sites of atrophy we’ve identified from the mouse models.