Here we presented the largest hippocampal mapping study to date (using 980 scans from 490 subjects). There were four main findings. First, our results with an automated analysis technique agreed well with prior studies in smaller samples, performed using manual or other relatively labor intensive methods. We found that the mean hippocampal volume loss rates increased with worsening diagnosis (AD: 5.59%/year; MCI: 3.12%/year; normal: 0.66%/year). Our estimated hippocampal loss rates in AD are a little high compared to some other studies (Du et al., 2003
; Jack et al., 2000
), but are nonetheless close to the average rates reported in a recent meta-analysis. Barnes et al. (in press)
included nine studies from seven centers, pooling data from a total of 595 AD and 212 matched controls. They found that mean (95% CIs) annualized hippocampal atrophy rates were 4.66% (95% CI 3.92, 5.40) for AD subjects and 1.41% (0.52, 2.30) for controls. Factors that might affect these rates include differences in anatomic delineation criteria — for example, some protocols include the hippocampal tail and others do not. Also, if hippocampal loss rates truly accelerate as the disease progresses, it is plausible that any study of AD patients who are more severely affected may find a greater mean rate of atrophy. Even so, this is not a compelling argument in our case, as our AD patients are relatively mildly impaired and yet show atrophic rates slightly higher than the average rate in the Barnes et al. meta-analysis. Second, all of our change measures were correlated with MMSE, global and sum-of-boxes CDR, and with changes in these measures. All such correlations were in the anticipated directions. The automated measures satisfied a further criterion expected of a reliable measure of disease progression — reliable correlation with cognitive decline. Third, we found that ApoE4 carriers had faster rates of volume loss, both in the full sample of 490, and when only cognitively intact controls were examined. This finding may have a practical benefit for treatment trials, because samples pre-selected to over-represent E4 carriers may show greater interval changes on MRI and may therefore be better powered to detect a statistical reduction in loss rates (Jack et al., 2008b
). In addition, one might argue that those carrying the risk gene are better candidates for early intervention as an accelerated disease process is already detectable on MRI, at least at the group level. Even so, the case for using ApoE4 as an enrichment criterion in addition to using information on baseline cognitive deficits also depends on whether the knowledge of a person’s genotype provides added information for predicting atrophic rates relative to what can be predicted from cognition alone; this requires further study. Fourth, of the 245 MCI subjects examined, those who converted to AD during the 1-year evaluation period had greater loss rates than non-converters.
We presented several analyses using both hippocampal volumes and 3D significance maps, to determine whether maps provide more information than simple volumetric summaries. For the main effects, both maps and volumes gave similar results, for example regarding the changes over time in all 3 diagnostic groups, and the correlations with MMSE, CDR scores and interval changes in the scores. The map data also showed similarly localized regions of ongoing hippocampal loss in controls, MCI and AD groups ().
The cases where maps and volumes gave different answers were typically when detecting effects of subtle factors that were associated with loss rates with borderline significance. In the full sample, a higher educational level was associated with slower right hippocampal loss rates (p=0.0051), an effect detected in the maps but not in the volumetric analysis. This significance level is borderline when corrected for multiple comparisons (0.0051×10=0.051), so it requires independent replication.
ApoE4 gene carriers showed faster loss rates in the volumetric analysis, even in controls, but these effects did not reach the corrected significance threshold in the maps (after correction for multiple comparisons by permutation testing). Even so, local effects of genotype were consistently found in the same regions in controls, MCI and AD groups, providing a possible search region for future studies of gene effects.
A major surprisewas that the maps comparing atrophy rates between diagnostic groups did not differentiate MCI from controls. The volumetric analyses showed that atrophy rates increased in the order CTL, MCI, AD, and all groups showed progressive loss (except for normals on the right side). While both the volumetric and the 3D analyses showed that AD subjects have significantly greater atrophy rates vs. controls, only the volumetric data showed a difference between MCI and controls. While maps tend to outperform volumetric summaries when detecting effects that are relatively concentrated or localized, as the most affected subregions may show higher effect sizes than an overall volumetric measure, maps may not outperform overall numerical summary measures when the effects are relatively diffuse. Numeric summaries such as hippocampal volume often involve an implicit averaging that may counteract any highly localized boundary segmentation errors, which can deplete the power of maps in the same regions. In addition, the inability to distinguish MCI from controls based on the maps alone may reflect the clinical heterogeneity of the MCI group, as a combination of various conditions in a sample likely yields a regionally diffuse mean atrophy pattern for the group as a whole. This highlights the need for both volumetric measures and map-based statistical analysis.
To obtain the most powerful statistics from map-based statistical analysis, it may be beneficial in the future to use these thresholded maps as search regions of interest for factors that influence the rates of atrophy. By focusing on the regions that are changing the most, it may be possible to statistically define surface subregions of interest in training samples to develop more powerful or precise measures of hippocampal atrophy in future independent samples. In this sense, this is the same logic as using AdaBoost to label the hippocampus, by combining classifiers with the greatest capacity for error reduction.
Some comparison is warranted between this approach and other methods for hippocampal mapping. Other hippocampal mapping studies by our group have contrasted the atrophy patterns in Lewy Body dementia, vascular dementia (Scher et al., in press
), amnestic MCI (Becker et al., 2006
), fronto-temporal dementia (Frisoni et al., 2006
), and in those at genetic risk for AD (Boccardi et al., 2004
) revealing morphometric signatures characteristic of each condition. These prior reports used the same surface parameterization methods and radial distance measures as were used in this report (statistics of radial atrophy), but the studies relied on manual tracing rather than automated segmentation, which greatly accelerates the rate at which scans can be analyzed.
Our work is related to that of Styner, Gerig, and Yushkevich on M
-reps (medial representations) (Styner et al., 2003
; Yushkevich et al., 2005
), in which a medial curve is defined through a structure, and distances of boundary points to the centerline are examined. The radial atrophy mapping technique is robust to small rotational or translational errors in registering the images across time, as the radial distances are always measured with respect to a central line threading down the center of the structure. Other methods, such as voxel-based morphometry, for example, may incorrectly pick up global shifts of the hippocampus as compressions or expansions inside the hippocampus, as the nonlinear deformations that register structures are spatially regularized (smooth) and may not be precise enough to register the hippocampal boundaries exactly. As radial atrophy is measured with respect to the medial curve, the distance measure reflects the thickness of the hippocampus in a given section, and it is not exactly the same as a volume difference. For example, any anterior-to-posterior shortening of the hippocampus would be detected by a volume measure, but the radial distance measure is not sensitive to this type of change — it only measures the radial thickness of the structure relative to a centerline. This may in general be considered a benefit rather than a limitation, as there are some variations across normal subjects in the anterior–posterior extent of the hippocampus, and these variations will be discounted by the radial mapping approach and will not be a source of confounding variance. Even so, discounting information on the anterior–posterior extent of the hippocampus may not be beneficial when measuring within-subject rates of change, unless the benefit of the information is outweighed by variations in segmentation accuracy across time that tend to cause errors in reproducibly defining the posterior limit of the hippocampus.
A related approach using large-deformation diffeomorphic metric mapping (Joshi and Miller, 2000
) has been used to deform labeled anatomical templates of the hippocampus onto new images, using a combination of manual landmarking of points on the hippocampus and 3D fluid image registration (Csernansky et al., 2000
; Haller et al., 1996
; Wang et al., 2007
). The surface of the hippocampus was parcellated a priori
on a neuroanatomical template into three zones approximating the locations of underlying subfields, and Large Diffeomorphic Metric Mapping (LDDMM) was used to generate the hippocampal surfaces of all subjects and to register the surface zones across subjects. In a cross-sectional study, Wang et al. (2006)
found that inward deformities of the hippocampal surface in proximity to the CA1 subfield and subiculum may be used to distinguish subjects with questionable AD from nondemented subjects. In a longitudinal study similar to ours, Wang et al. (2003)
used LDDMM to analyze hippocampal change over time in 18 subjects with questionable AD (CDR 0.5) and 26 age-matched nondemented controls (CDR 0) scanned 2 years apart. In CDR 0 subjects, they observed shape changes between baseline and follow-up largely confined to the head of the hippocampus and subiculum, while in the CDR 0.5 subjects, shape changes involved the lateral body of the hippocampus as well as the head region and subiculum. In a subanalysis of 9 subjects from the same sample, who converted from the nondemented (CDR 0) to the questionable AD (CDR 0.5) state, Qiu et al. (2008)
found that compared to the non-converters, the lateral aspect of the left hippocampal tail showed inward surface deformation in the converters. With a similar method, Csernansky et al. (2005)
found that inward deformation of the left hippocampal surface in a zone corresponding to the CA1 subfield is an early predictor of the onset of DAT in nondemented elderly subjects. This is consistent with our finding of more rapid hippocampal loss rates in MCI converters than non-converters. Surface-based maps may be used in the future to define hippocampal subregions where changes predict imminent cognitive decline or the onset of dementia. Using the radial distance method we have likewise demonstrated that CA1 and subicular atrophy associates with future conversion from MCI to AD (Apostolova et al., 2006b
) and from normal cognition to MCI (Apostolova et al., in press
Related shape modeling studies have involved modeling the hippocampal surface using spherical harmonic functions (SPHARM) (Styner et al., 2004
; Thompson and Toga, 1996
), and using the coefficients of the harmonic expansion to infer shape differences between dementia patients and controls. In Gutman et al. (2008)
, we used a support vector machine classifier based on spherical harmonics to classify 49 AD patients and 63 controls with 75.5% sensitivity and 87.3% specificity, with 82.1% correct overall. This approaches the 89–96% classification accuracy of the best diagnostic predictors (Vemuri et al., 2008
). Given the proliferation of new MRI-based measures of hippocampal degeneration based on automated surface matching methods (Wang et al., 2005
), automated partitions of surfaces (Shi et al., 2007
), random field theory on surfaces (Bansal et al., 2007
) and machine learning methods (Li et al., 2007
), future studies will likely focus on defining which MR-based measures provide optimal statistical power for detecting factors that slow the progression of AD, to optimize the power of future interventional trials.