This study is representative of several current research efforts that use automated methods to measure hippocampal atrophy in AD, including large diffeomorphic metric mapping (Csernansky, et al. 2004
; Wang, et al. 2007
), volumetric analysis (Geuze, et al. 2005
), and fluid registration (van de Pol, et al. 2007
). Clinical measures of disease burden have also been correlated with regional hippocampal atrophy in several surface-based mapping studies of AD and MCI (Apostolova, et al. 2006a
; Becker, et al. 2006
; Frisoni, et al. 2006
), Lewy body dementia (Sabattoli, et al. 2007
), and vascular dementia (Scher, et al. 2007a
), and of conversion between MCI and AD (Apostolova, et al. 2006b
; Apostolova and Thompson 2007
For each surface model, a medial curve was defined as the line traced out by the centroid of the hippocampal boundary (Styner, et al. 2005
; Thompson, et al. 2004a
). The medial curve was defined separately in each individual, before averaging the surfaces. The operations of averaging surfaces and defining the medial curve from a surface are not commutative, because a medial curve derived from an average surface would not be the same as the average of the medial curves derived from each individual. Because we were interested in measuring radial atrophy in each individual, we computed these measures in each subject with reference to their own medial curve, but plotted the resulting statistics on the average surface for the groups being compared.
One advantage of using the radial distance measure - and computing it in each individual separately - is that it is invariant to overall 3D shifts or translations 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. This is because the nonlinear deformations that register structures across subjects in VBM are spatially regularized (smooth) and may not be precise enough to register the hippocampal boundaries exactly. By contrast, if AD is associated with some shifting of the mean stereotaxic position of the hippocampus, the radial distance measure will not be affected by it, and will only detect localized reductions in volume intrinsic to the hippocampus. This also makes the radial atrophy mapping technique somewhat robust to small rotational or translational errors in registering the images across subjects, as the radial distances are always measured with respect to a central line threading down the center of the structure. The radial distance is a reasonable proxy for volumetric loss, mapping its distribution in 3D, but will not be sensitive to some types of volumetric change. As radial atrophy is measured with respect to the medial curve, the distance measure reflects the thickness of the hippocampus in a given section, which may not reflect an overall 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 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.
Although an FDR correction for the number of elements in each p
-map is performed, no correction for the number of clinical markers is done. Ideally, we should lower our threshold according to Bonferroni principles. However, none of the permutation tests were close to the 0.05 level, so this is not necessary. In a recent study (regarding brain regions other than the hippocampus), we performed a 5-fold cross-validation to assess the predictive power of morphometric measures, for assessing conversion to AD from MCI over a 6 month period, using truly independent subsamples (Zhang, et al. 2008
Several prior studies used 3D surface-based maps to visualize the profile of hippocampal atrophy in AD and MCI, but they relied almost solely on manual tracing, which is extremely time-intensive. Becker et al. (Becker, et al. 2006
) found greater hippocampal atrophy in AD versus MCI (in a total of N
= 66 subjects) with greatest differences in the CA1 and subiculum regions. Frisoni et al. (Frisoni, et al. 2006
= 68), found greater atrophy in AD versus controls, with the main differences in the CA1 field and parts of the subiculum, with CA2 and CA3 regions relatively spared. Finally, Apostolova et al. (Apostolova, et al. 2006a
= 65) found MCI v. AD differences in CA1 for both hippocampi, and in CA2 and CA3 only on the right.
Another approach to automatically relate Alzheimer’s disease diagnosis and clinical scores with systematic differences in brain structure on MRI is voxel-based morphometry (VBM). VBM has shown promise in several previous studies, including Chetelat (Chetelat, et al. 2005
) which tracked gray matter loss in a longitudinal study of 18 MCI patients, Whitwell (Whitwell, et al. 2007
), who also showed gray matter loss over three years in 63 MCI subjects, and Good (Good, et al. 2002
) compared VBM to ROI analysis and showed that they compare favorably in detecting structural differences in Alzheimer’s disease.
In our maps, the right hippocampal head shows greater atrophy in MCI versus normal groups; the right hippocampal tail shows atrophy only in the AD-MCI and AD-control comparisons. The finding of right anterior hippocampal atrophy in MCI is consistent with findings by de Toledo-Morrell et al. (De Toledo-Morrell, et al. 2000
), who observed that the right entorhinal cortex had greatest atrophy in elderly patients. Some investigators have argued that both the EC and hippocampal formation degenerate before the onset of overt dementia, and that EC volume is a better predictor of conversion (Dickerson, et al. 2007
Correlations with symptoms have also been examined using hippocampal maps. Ballmaier et al. (Ballmaier, et al. 2007
) mapped atrophy of the hippocampal head in depressed versus non-depressed elderly controls. Statistical mapping results, confirmed by permutation testing, showed that regional surface contractions were greater in elderly subjects with late-versus early-onset depression in the anterior aspects of the subiculum, and lateral-posterior aspects of the CA1 subfield in the left hemisphere. Similar studies of patients with Lewy body dementia showed a more restricted pattern of hippocampal atrophy than AD patients at a comparable level of cognitive impairment (Sabattoli, et al. 2008
). Hippocampal maps may therefore complement cortical maps in revealing the selective atrophic patterns that characterize different types of dementia.
We also aimed to identify regions of hippocampal degeneration that predicted either a change in diagnostic classification or a decline in standard clinical measures of functional decline. Although we correlated atrophy with several measures of subsequent decline, only the 1-year decline in sum-of-boxes CDR scores was close to being linked with baseline HP atrophy (p
= 0.056; a post hoc
test gave p
= 0.036, corrected
, but used a less stringent re-thresholding of the data to define suprathreshold statistics). This difficulty in predicting future decline may be due to the shortness of the follow-up interval; other studies also reported difficulties in predicting cognitive decline based on hippocampal volumes alone, especially when concomitant subcortical brain injuries, such as lacunar infarcts were present (Mungas, et al. 2002
). By contrast, Apostolova et al. (Apostolova, et al. 2006b
) recently used the same surface mapping approach (but based on manual HP segmentation) and predicted subsequent clinical decline in 20 MCI subjects. Over a 2-year follow-up period, 6 patients developed AD, 7 remained stable, and 7 improved. Smaller hippocampi and specifically CA1 and subicular involvement were associated with future conversion from MCI to AD, while MCI patients who improved and no longer met MCI criteria at follow-up tended to have larger hippocampal volumes and their subiculum and CA1 regions were relatively preserved.
In our study, right hippocampal atrophy was associated with depression severity, consistent with several prior studies of elderly depression. In elderly subjects with late-versus early-onset depression, Ballmaier et al. (Ballmaier, et al. 2007
) found greater HP atrophy in the anterior aspects of the subiculum, and lateral-posterior aspects of the CA1 subfield in the left hemisphere. In that study, hippocampal surface contractions correlated with memory measures in late-onset depressed patients. Our maps and those of Ballmeier et al. (Ballmaier, et al. 2007
) are consistent with most earlier studies (Bell-McGinty, et al. 2002
; Hickie, et al. 2005
; Lloyd, et al. 2004
; O’Brien, et al. 2004
; Steffens, et al. 2000
), showing smaller hippocampal volumes in elderly depressed patients compared to controls. We detected correlations for only the right HP in this study, consistent with several other reports that have shown differences as being more pronounced for the right than for the left hemisphere (Bell-McGinty, et al. 2002
; Steffens, et al. 2000
Next, we examined other hypothesized correlations between hippocampal morphology and educational level, blood pressure, homocysteine levels, and depression severity. Each of these factors has been associated with AD, but they have not yet been directly linked with hippocampal morphology. Long term high blood pressure has been associated with earlier AD onset (Skoog, et al. 1996
), elevated homocysteine levels have been linked with stroke, and to some extent with AD (Morris 2003
), a lower level of education has been shown to be associated with dementia (Stern, et al. 1994
), and depression commonly accompanies AD (Ballmaier, et al. 2007
; Ballmaier, et al. 2004
). We showed that a strong linkage is absent in all of these cases except for depression, which has been supported by other studies (Sheline, et al. 1996
). Neither ApoE4 nor ApoE2 was linked with morphology, either in those without AD or in the full patient sample, suggesting that it is not a powerful modulator of hippocampal morphology, despite its role as a risk gene for AD that was confirmed to be over expressed in our MCI and AD groups versus the normal controls. The lack of detectable hippocampal differences between ApoE4 groups is also consistent with other cross-sectional MRI studies that employed manual tracing of the hippocampus (Jak, et al. 2007
Finally, we computed empirically-based estimates of the minimal sample sizes necessary to detect the well-known correlations of hippocampal atrophy with diagnosis, and with clinical test scores. When conventional volumetric measures were used, Jack et al. (Jack, et al. 2003
) estimated that in each arm of a therapeutic trial, only 21 subjects would be required to detect a 50% reduction in the rate of decline if hippocampal volume were used as the outcome measure. This compared with 241 subjects if MMSE scores were used and 320 if the AD Assessment Scale Cognitive Subscale (ADAS-Cog) were used.
Here we found that the more subtle MCI state was difficult to distinguish from either AD or normal aging with fewer than 200-300 subjects overall, but that the four other associations (normal v. AD, and correlations between atrophy and MMSE scores, global CDR scores and sum of boxes CDR scores) only required 40 subjects to detect (24 was not sufficient). This finding was unexpected; prior studies, based on manual tracings, required far fewer subjects to differentiate MCI from AD and from controls (Apostolova, et al. 2006a
). The manual segmentations previously employed may have produced more accurate hippocampal models, but they were time consuming to create (often taking several weeks or months, making large-scale analyses difficult or prohibitive). In our current study, our automated approach greatly reduced our segmentation time for a large sample (requiring less than one minute of CPU time on a desktop computer). Whether or not this approach would be optimal for very small studies remains to be proven, as the improvement in automation versus the need to trace a small training set of around 20 images may become limiting in very small samples. Another factor that may have led to our higher sample size requirements for group differentiation is that ADNI collects scans at many acquisition sites. Even so, major efforts have been devoted to protocol design and calibration across scanners and sites.
In future, we plan to map the progression of hippocampal atrophy over time, once the follow-up (longitudinal) scans are available for the full ADNI sample. Morphometric correlates of disease progression may be easier to examine with longitudinal imaging data. A further avenue of work will apply this method to help distinguish neurodegenerative patterns between AD other types of dementia. In our recent studies of Lewy body dementia (Sabattoli, et al. 2007
) and vascular dementia (Scher, et al. 2007a
), we found a more restricted pattern of hippocampal atrophy than AD patients at a comparable level of cognitive impairment.