Abstract
Background: To facilitate image analysis, most recent 2-[18F]fluoro-2-deoxy-d-glucose PET (FDG-PET) studies of glucose metabolism (MRglc) have used automated voxel-based analysis (VBA) procedures but paradoxically none reports hippocampus MRglc reductions in mild cognitive impairment (MCI) or Alzheimer disease (AD). Only a few studies, those using regions of interest (ROIs), report hippocampal reductions. The authors created an automated and anatomically valid mask technique to sample the hippocampus on PET (HipMask).
Methods: Hippocampal ROIs drawn on the MRI of 48 subjects (20 healthy elderly [NL], 16 MCI, and 12 AD) were used to develop the HipMask. The HipMask technique was applied in an FDG-PET study of NL (n = 11), MCI (n = 13), and AD (n = 12), and compared to both MRI-guided ROIs and VBA methods.
Results: HipMask and ROI hippocampal sampling produced significant and equivalent MRglc reductions for contrasts between MCI and AD relative to NL. The VBA showed typical cortical effects but failed to show hippocampal MRglc reductions in either clinical group. Hippocampal MRglc was the only discriminator of NL vs MCI (78% accuracy) and added to the cortical MRglc in classifying NL vs AD and MCI vs AD.
Conclusions: The new HipMask technique provides accurate and rapid assessment of the hippocampus on PET without the use of regions of interest. Hippocampal glucose metabolism reductions are found in both mild cognitive impairment and Alzheimer disease and contribute to their diagnostic classification. These results suggest re-examination of prior voxel-based analysis 2-[18F]fluoro-2-deoxy-d-glucose PET studies that failed to report hippocampal effects.
Most metabolic neuroimaging studies paradoxically suggest absent hippocampal dysfunction in Alzheimer disease (AD) and mild cognitive impairment (MCI). Over 200 studies using PET with 2-[18F]fluoro-2-deoxy-d-glucose (FDG) have demonstrated abnormal reductions of glucose metabolism (MRglc) in large cortical regions,1–5 whereas only five reported evidence for hippocampal abnormalities in AD or MCI.6–10 The finding of hippocampal hypometabolism comes exclusively from FDG-PET studies that used a manual region-of-interest (ROI) sampling method guided by coregistered MRI.6–10 On the other hand, the failure to observe hippocampus effects is a consistent feature of those FDG-PET studies that used the automated voxel-based analysis (VBA). VBA is predominantly used for its convenience in image analysis and in statistical examination of group differences. VBA relies on fully automated image registration and size normalization protocols to standardize all brains to a common space, enabling assessment of the PET scans on a voxel-wise basis.11–13 We hypothesized that the negative VBA findings were due to a failed spatial alignment of relatively small structures like the hippocampus that are prone to high anatomic variability with aging and neurodegeneration.
We conducted two studies. Study 1 used MRI-ROIs to evaluate the VBA normalization accuracy for the hippocampus and to derive a precise and anatomically valid procedure for automatically sampling the hippocampus using a hippocampus mask (HipMask). In Study 2, we implemented the HipMask in an FDG-PET study of normal elderly (NL), patients with MCI, and patients with AD and compared the results with existing image analysis methods.
The results demonstrate the clinical utility of the HipMask procedure to determine hippocampal MRglc reductions in MCI and AD and its equivalence to the gold standard ROI. Hippocampal reductions were not detected by VBA.
Methods.
This section includes the general methods used as well as those specific to our two studies.
Subjects.
Subjects were drawn from New York University (NYU) School of Medicine AD Core Center. Written informed consent was obtained from all subjects (and when appropriate also from a caregiver). Subjects received an extensive screening and diagnostic battery that consisted of medical, neurologic, psychiatric, neuropsychological, and MR examinations.
Exclusion criteria.
Subjects were excluded if they had evidence of conditions affecting brain structure or function (e.g., stroke, clinically uncontrolled diabetes, major head trauma, active evidence of depression) or use of cognitively active medications.
Inclusion criteria.
Eighty-four subjects were included in this study. All were between 50 and 85 years of age and had a minimum of 12 years education. The elderly NL selected for study had Mini-Mental State Examination (MMSE)14 scores ≥ 28 and Global Deterioration Scale (GDS)15 scores of 1 or 2. The patients with MCI selected had MMSE scores > 24 and GDS scores = 3. The patients with mild to moderately severe AD received GDS scores of 4 or 5. The diagnosis of probable AD was consistent with the guidelines of the National Institute of Neurologic and Communicative Disorders and Stroke-AD and Related Disorders Association16 or the Diagnostic and Statistical Manual of Mental Disorders IV (DSM-IV)17 criteria.
This project used two study cohorts (table 1). In the development of the HipMask in Study 1 we used a training cohort of NL (n = 20), MCI (n = 16), and AD (n = 12), for a total of 48 subjects. In the implementation of the HipMask in Study 2 the testing cohort was comprised of NL (n = 11), MCI (n = 13), and AD (n = 12), for a total of 36 subjects. The various processing steps for Studies 1 and 2 are summarized in figures 1 and 2 and detailed in the Appendices.
Table 1 Subject characteristics
Figure 1. Flow chart showing the processing steps followed in study 1 (hippocampus mask development). This scheme represents the various steps followed to go from an initial sequence of MRI scans to the creation of the HipMask. (A) The pes hippocampus regions of interest (Hip ROIs, in red) were drawn on the T1-MRI scans of 48 subjects (20 healthy elderly, 16 mild cognitive impairment, and 12 Alzheimer disease) using a locally developed Multimodal Image Data Analysis System package (MIDAS, version 1.6). Afterwards, the MRI scans from the entire training cohort were averaged to create an elderly brain template image that approximates the Talairach and Tournoux space.23 The MRI scans and the individual pes hippocampus ROIs were then spatially normalized to the template using voxel-based analysis procedures,12 in order to place all the hippocampi in the same anatomic space. (B) The spatially normalized pes hippocampus ROIs (nROIs) were transferred back to MIDAS where they were superimposed slice by slice and the extent of overlap for the nROIs assessed in order to derive the optimal probabilistic overlap mask for the pes hippocampus (HipMask). (C) We created a probability image of the pes hippocampus in the stereotactic space with the number of subjects overlapping for each voxel ranging from 0 (blue, no overlap) to 48 (red, 100% overlap) (the probability image is here shown for one MRI slice only). (D) To test the normalization accuracy for the hippocampus, we determined the percentage spatial overlap between the template hippocampus (gray) and that hippocampal anatomy shared by all the 48 cases (red scale).
Figure 2. Flow chart showing the processing steps followed in study 2 (hippocampus mask implementation). This scheme represents the various steps followed to develop an automated procedure for accurate sampling the pes hippocampus on PET scans. The pes hippocampus regions of interest (Hip ROIs, in red) were drawn on the T1-MRI scans of 36 new subjects (11 healthy elderly, 13 mild cognitive impairment, and 12 Alzheimer disease) using MIDAS 1.6. The same subjects also received an FDG-PET scan. (A) First, we coregistered the MRI and PET scans in their native space and the pes hippocampus MRglc data were extracted with the manual ROIs (in white) from the MRI-coregistered PET scans. (B) The PET scans were then spatially normalized to the elderly brain template created in Study 1 using voxel-based analysis routines.12 Hippocampus MRglc data were automatically extracted with the HipMask (in white). (C) As per accepted methods the spatially normalized PET scans were smoothed with an isotropic Gaussian filter (12 mm FWHM) and examined for MRglc differences across groups using the voxel-based analysis as implemented in SPM99.12 Finally, we compared the MRglc data from the three methods: (A) individualized MRI-guided ROIs, (B) HipMask, and (C) voxel-based analysis.
MRI examinations.
All subjects received a standardized MRI scan protocol. The MRI scanning procedures have been previously described.8 Briefly, the MRI scans were acquired as coronal 1.3-mm-thick images obtained perpendicular to the long axis of the hippocampus using a 1.5 T General Electric Signa imager (General Electric, Milwaukee, WI). We used a T1-weighted 3D spoiled gradient recalled sequence (SPGR) with repetition time = 35 msec, echo time = 9 msec, and flip angle 60° (field of view = 18 cm, number of excitations = 1, matrix = 256 × 128) and reconstructed into 124 contiguous slices.
Regions of interest.
MRI scans were transferred to a Sun Sparc workstation (Sun Microsystems, Mountain View, CA). According to our previous PET studies demonstrating superior diagnostic accuracy for the larger anterior portion of the hippocampus,7–9 we focused on the pes hippocampus. Pes hippocampus ROIs were drawn in both hemispheres on threefold enlarged coronal MR images using a locally developed Multimodal Image Data Analysis System package (MIDAS, version 1.6).18 Each of two observers, blind to subject diagnosis, drew all pes hippocampal ROIs on half of the cases (randomly chosen). Each ROI was independently checked for accuracy by the other observer and any changes were made by joint agreement. Briefly, drawings were performed along the full anterior-posterior extent of the pes hippocampus and included a portion of the subiculum. The lateral hippocampal border was the temporal horn of the lateral ventricle and the medial border was the ambient cistern. The inferior border was the white matter (WM) of the parahippocampal gyrus (PHG). Most anterior, the hippocampus was distinguished from the amygdaloid body by fibers of WM interposed between these regions. The pes-hippocampus volumes were corrected for variations in head size by using the volume of the intradural supratentorial compartment.9
FDG-PET examinations.
Within 3 months of the MRI, the 36 subjects included in the testing cohort received an FDG-PET scan at Brookhaven National Laboratories (BNL, Long Island, NY) using a Siemens CTI-931 scanner (Knoxville, TN). The scanner generated 15 axial tomographic slices covering 108 mm. The in-plane axial resolution was 6.2 mm FWHM (direct slices) and for the cross-slices was 6.7 mm. The interslice distance was 6.75 mm. Images were reconstructed using the Hanning filter with a frequency cutoff of 0.5 cycles/pixel, yielding 128 × 128 matrix with a pixel size of 1.56 mm. Each subject's head was positioned using two orthogonal laser beams and imaged with the scanner tilted 25° negative to the canthomeatal plane.8 This plane runs approximately parallel to the long axis of the hippocampus. To reduce head movement during scanning, a molded plastic head holder was custom-made for each subject. Attenuation correction was obtained using 68Ga/68Ge transmission scans. One hour prior to the FDG-PET scan, a radial artery catheter and contralateral antecubital venous lines were positioned. Subjects received 5 to 6 mCi of FDG IV while laying supine in a dimly lit room. Arterial blood samples were obtained at standard intervals throughout the study to monitor glucose and 18F levels in the blood. PET images were acquired 35 minutes after isotope injection and lasted for 20 minutes. The absolute glucose consumption rate was calculated for each pixel using the Sokoloff equation19 with standard kinetic constants.20
PET/MRI coregistration.
Using MIDAS, each PET scan was coregistered with the corresponding MRI by using a three-dimensional method based on minimizing the variance of the signal ratios.21 Our implementation calls for a preliminary spatial alignment, using intrinsic anatomic landmarks. The final version of coregistered PET/MRI data consisted of coronal sections perpendicular to the long axis of the hippocampus, with 230 mm FOV and 256 × 256 matrix.
Study 1.
Assessment of the VBA normalization accuracy for the hippocampus.
VBA involves spatial normalization of individual scans to a standardized brain-volume template in order to reduce the interindividual variability in brain anatomy. This protocol enables different populations to be compared on a voxel-by-voxel basis across the whole brain.11,12 Statistical Parametric Mapping (SPM99, Wellcome Department of Cognitive Neurology, London) was chosen for the automated normalization of the MRI and PET scans11,12,22 because of its wide utilization in PET studies. An anatomic MRI elderly brain template was created from the entire training cohort (n = 48) in order to provide a template image appropriate to the population sample.22 Making the elderly template involved spatially normalizing each MRI scan to the SPM T1-MRI template image from the Montreal Neurologic Institute (MNI), which approximates the Talairach space.23 The spatial normalization involves estimating the optimum (least squares) 12-parameter affine transformation, followed by an iterative estimate of local alignment based on a family of 7 × 8 × 7 discrete cosine functions.12 The spatially normalized MR images were then resampled on a 105 × 126 × 91 matrix with a voxel size of 1.5 × 1.5 × 1.5 mm and the origin set at x = 53, y = 76, z = 34 mm. The individually normalized scans were smoothed with an 8-mm full-width at half-maximum (FWHM) isotropic gaussian kernel, and averaged to form the elderly brain template.
The MRI scans from the training cohort (in native space without the embedded ROIs) were then registered to the elderly brain template12 using the same linear and non-linear transformation protocol described for the template creation. Finally, the calculated transformation parameters were applied to spatially normalize the ROIs. This produced a data set with all subject scans and hippocampus ROIs sharing the same stereotactic space.
A detailed description of the procedures utilized to assess the effects of the normalization accuracy of the hippocampus can be found in appendix E-1 on the Neurology Web site at www.neurology.org. Briefly, as shown in figure 1, the hippocampal MRI-ROIs were spatially normalized according to the VBA protocol and superimposed to determine the percentage of hippocampal volume correctly shared by all subjects. Use of the spatial normalization technique assumes that the normalized brain regions occupy the same anatomic position and have the same shape. To test this assumption, we determined the overlap across subjects for each voxel. This measure can range between 0% (no overlap) and 100% (overlap across all subjects).
Creation of the optimal hippocampus mask.
After estimating the normalization errors for the hippocampus using standard VBA procedures, we created a sample HipMask that, after intersubject averaging, only included those portions of the hippocampus where the voxel-wise overlap was maximized across subjects. The optimal HipMask was determined using bootstrapping procedures and positive likelihood ratio calculations24,25 to identify a hippocampal mask that was independent of sample size and yielded the highest sampling precision. Appendix E-1 details the statistical and technical procedures utilized.
Study 2. Implementation of the ROI, HipMask, and VBA methods in an FDG-PET study of NL, MCI, and AD.
Hippocampal ROIs.
After PET coregistration with the corresponding MRI scans, the FDG-PET images of the testing cohort (n = 36) were sampled by using the MRI hippocampus ROIs (figure 2A).
HipMask.
The MRI-coregistered PET images were spatially normalized onto the elderly brain template using the MRI image to define the normalization parameters as described in Study 1. Further, to determine whether the PET scans could be used in the absence of an MRI, the uncoregistered PET scans were directly normalized to the template (figure 2B).
The hippocampal ROIs and the HipMask were applied to the PET images to extract the hippocampal MRglc (μmol/100 g/minute) across all slices sampled with the values averaged across hemispheres.
We examined several additional questions. 1) The correspondence between the HipMask sampling and the ROI technique; 2) the HipMask susceptibility to malpositioning errors; and 3) the effect of atrophy correction26,27 on the hippocampal MRglc data (see appendix E-2, available online at www.neurology.org).
Statistical methods.
For all image analysis methods, the General Linear Model (GLM) univariate analysis was used to identify the MRglc effects across the three clinical groups, after controlling for age, sex, and pons metabolism. Pons MRglc was sampled at the center of a mid pontine slice at the level of the middle cerebral peduncles with a 16 × 16 mm box9 and was used to adjust for between-subject variations in the global MRglc.28
Results were considered significant at p < 0.05. For the ROI and the HipMask data, post hoc comparisons were performed with Scheffe' tests using SPSS 12.0. For VBA, post hoc comparisons were performed with t-tests as implemented in SPM99.29 The VBA results were assessed for all the voxels in the whole brain, with a subsequent correction for multiple comparisons. Only voxels with values greater than 80% of the whole brain mean MRglc were included in the analysis29 and only clusters exceeding an extent threshold of 30 voxels (i.e., >2 times the FWHM) were considered significant. Anatomic localization of hypometabolic areas was assessed according to the Talairach and Tournoux space.23 MRglc values were extracted from the brain regions showing group effects using MarsBar toolbox (http://marsbar.sourceforge.net/).
Diagnostic accuracy.
All significant regions from the above image analyses were examined with logistic regressions to assess their diagnostic accuracy in classifying the NL, MCI, and AD groups. Results were considered significant at p < 0.05.
Results.
Study 1.
Clinical data.
Analyses of variance (ANOVAs) showed no age, sex, or education differences among the NL, MCI, and AD groups. Group differences were observed for the MMSE scores (F[2,46] = 42.6, p < 0.001). Post hoc comparisons between groups showed lower MMSE scores for the AD group as compared to both the NL and MCI groups (p < 0.001) (see table 1).
MRI hippocampal volume data.
Analyses of covariance showed differences among NL, MCI, and AD groups for the pes hippocampus volume, after correcting for age, sex, and head size (F[2,43] = 20.8, p < 0.001). Post hoc comparisons between groups showed lower pes hippocampus volumes for the MCI (18%) and AD group (23%) as compared to NL (p < 0.001) (see table 1 for unadjusted means and SD).
Assessment of the normalization accuracy of the hippocampus.
Our results show that across all subjects only 3% of the pes hippocampus ROIs had 100% overlap. In other words, 97% of this anatomy was not completely shared after spatial normalization (see appendix E-1 for details).
Creation of the HipMask.
Our results show that, by maximizing the voxel-wise overlap for the hippocampus across subjects, a reliable HipMask can be created that is independent of sample size, yields accurate anatomic sampling as assessed by MRI (see appendix E-1 for details), and is highly correlated with the ROI gold standard (see appendix E-2).
Study 2.
Clinical data.
ANOVAs showed no differences among NL, MCI, and AD groups for age, sex, and education. Group differences were observed for the MMSE scores (F[2,34] = 16.0, p < 0.001). Post hoc comparisons between groups showed lower MMSE scores for the AD group as compared to both the NL and MCI groups (see table 1).
Comparisons between analytic methods.
An excellent correspondence was found between the HipMask and the ROI MRglc measures in all the clinical groups (ps < 0.001; figure 3 and see appendix E-2 for a detailed description). Hippocampus ROI MRglc was different across the three groups (F[2,33] = 8.0, p < 0.001). Post hoc Scheffe' tests for the paired groups showed MRglc reductions for AD (31%, p < 0.001) and MCI (14%, p < 0.05) relative to NL (figure 4) (see table 1 for unadjusted means and SD). Consistent with the ROI MRglc data, the HipMask MRglc was also different across the three clinical groups (F[2,33] = 9.6, p < 0.001). Post hoc Scheffe' tests for the paired groups showed HipMask MRglc reductions relative to NL that were nearly identical to those found with the ROI: AD (33%, p < 0.001) and MCI (10%, p < 0.05) (see figure 4, table 1). For neither ROI nor HipMask methods were differences found between MCI and AD.
Figure 3. Correlation between hippocampus MRglc data extracted with the non-corrected regions of interest from the MRI-coregistered PET scans and with the HipMask from the spatially normalized PET scans (r = 0.89, p < 0.001). Healthy elderly = white circles; mild cognitive impairment = black circles; Alzheimer disease = triangles.
Figure 4. Hippocampus MRglc differences for healthy elderly (NL), mild cognitive impairment (MCI), and Alzheimer disease (AD) from the grand mean, after controlling for age, sex, and pons metabolism. Hippocampal MRglc reductions were found for both MCI ( p < 0.05) and AD ( p < 0.001) groups relative to NL using either the regions of interest (gray) or the HipMask (white) techniques.
VBA showed no hippocampal group differences (figure 5). Consistent with other VBA observations, patients with AD showed reduced MRglc within the bilateral posterior cingulate (PCC), inferior parietal (IPC), temporal (TC), and left inferior frontal cortices (IFC) as compared to the NL as well as to the MCI (p < 0.05, corrected for multiple comparisons) (see figure 5 and table E-1). No VBA differences were found between NL and MCI. For exploratory purposes, as done by previous VBA studies in MCI,5,10,30–32 we removed the correction for multiple comparisons and re-examined our data using less conservative probability thresholds. In MCI relative to NL two clusters of reduced MRglc were found: the left middle TC and the left PCC (see figure 5 and table E-1).
Figure 5. Results from voxel-based analysis with SPM99. (A) No MRglc differences were found between healthy elderly (NL) and mild cognitive impairment (MCI) at p < 0.05, corrected for multiple comparisons. (B) By resetting the probability threshold at a less restrictive value of p < 0.001, uncorrected, two clusters of MRglc reductions were found in the left posterior cingulate (PCC) (left side of figure) and the left superior temporal gyrus (right side of the figure) for the patients with MCI as compared to NL. (C) PET voxels showing MRglc reductions for the patients with Alzheimer disease (AD) as compared to NL, involving the parieto-temporal, PCC, and frontal cortices ( p < 0.05, corrected for multiple comparisons). (D) PET voxels showing MRglc reductions for the patients with AD as compared to MCI, involving the parieto-temporal, PCC, and frontal cortices ( p < 0.05, corrected for multiple comparisons). Such reductions have a regional distribution similar to that observed for AD vs NL but are less spatially extended, suggesting that cortical hypometabolism may be present also in some patients with MCI. Areas of hypometabolism (represented on a red to yellow color coded scale) are displayed onto the medial and left lateral views of a volume-rendered spatially normalized MRI image. These brain regions correspond to the brain atlas coordinates in table E-1.
Diagnostic accuracy.
Diagnostic accuracy is illustrated in table 2.
Table 2 Diagnostic classification accuracy, sensitivity, and specificity in %
NL vs MCI diagnostic classification.
The logistic regression analyses showed that the hippocampus ROI MRglc correctly classified 77% of the NL and MCI patients (χ2 [1] = 6.91, p < 0.01). Likewise, HipMask MRglc classified 78% of these cases (χ2 [1] = 5.76, p < 0.02). Based on the VBA exploratory study, the PCC MRglc reductions classified 68% of the cases (χ2 [1] = 3.59, p < 0.05). The TC MRglc was not significant. Entering either hippocampal ROI or HipMask measures at the second step of the regression model improved the PCC MRglc discrimination accuracy from 68% to 73%, yielding improved identification of the patients with MCI (82% sensitivity, at the same specificity level) (ROI: χ2 [2] = 6.49, p < 0.04; HipMask: χ2 [2] = 7.29, p < 0.03). These data underline the importance of hippocampal evaluation in the early detection of AD.
NL vs AD diagnostic classification.
The logistic regression analyses showed that the hippocampus ROI MRglc correctly classified 78% of the NL and AD patients (χ2 [1] = 11.2, p < 0.001). Likewise, HipMask MRglc classified 78% of these cases (χ2 [1] = 13.9, p < 0.001). The overall discrimination accuracy for the brain regions identified with VBA was 87% for the IPC MRglc (χ2 [1] = 9.78, p < 0.005), 83% for the TC (χ2 [1] = 14.4, p < 0.001) and IFC MRglc (χ2 [1] = 16.7, p < 0.001), and 74% for the PCC MRglc (χ2 [1] = 8.81, p < 0.005). Moreover, entering either hippocampal measure at the second step of the regression model boosted the overall discrimination accuracy from 74% to 83% for the PCC (χ2 [2] = 17.9, p < 0.001), from 83% to 87% for the TC (χ2 [2] = 18.6, p < 0.001), and from 83% to 91% for the FC (χ2 [2] = 23.3, p < 0.001). Only the IPC did not benefit from the added hippocampal data. These data show that hippocampal evaluation is important even after disease expression.
MCI vs AD diagnostic classification.
No hippocampal MRglc differences were found between MCI and AD with either the ROI or HipMask techniques. The logistic regression analyses showed that the overall discrimination accuracy for the brain regions identified with VBA was 83% for the IPC MRglc (χ2 [1] = 7.23, p < 0.01), 78% for the TC MRglc (χ2 [1] = 16.6, p < 0.001), and 70% for the FC MRglc (χ2 [1] = 3.97, p < 0.05). Although PCC MRglc was not a sensitive group discriminator, it showed an advantage when added to the TC measures by boosting the accuracy from 78% to 91% (χ2 [2] = 18.9, p < 0.001). These results show that the differentiation between MCI and AD is largely determined by the involvement of the neocortex.
Discussion.
Previous FDG-PET studies that used VBA have not reported hippocampal metabolic reductions in AD or MCI. Reports of hippocampal metabolic reductions come only from studies using coregistered MRI scans and individually guided MRI-ROI placement.6–10 This had led to the paradoxical conclusion that hippocampal MRglc reductions are not found in AD. Our results show hippocampal metabolic reductions in both MCI and AD and that these reductions can be detected using either a new automated sampling technique, the HipMask, or by the ROI method. Our study demonstrates that the time saving VBA-based HipMask method is precise enough for hippocampus sampling on PET scans. This will enable the future automated analysis of large data sets.
The present study directly addresses the paradoxical conclusions drawn from the bulk of the FDG-PET literature as compared with those from both pathology and MRI studies. Hippocampal pathology and neuronal atrophic changes have been widely reported33–35 and MRI studies show gross hippocampus atrophy in AD and MCI.36 On the other hand, only a handful of over 200 FDG-PET studies in AD or MCI demonstrate hippocampal dysfunctions.6–10 Thus, most evidence indicates absent hippocampal MRglc reductions in AD and MCI. Several technical explanations exist as to why the hippocampus on PET was ignored or inadequately studied. Earlier generations of PET instruments had relatively poor (7 to 10 mm) spatial resolution, which precluded analyses of small brain regions like the hippocampus. Other and more recent ROI studies performed with high-resolution PET scanners still yielded negative results presumably because they did not use MRI coregistration procedures37,38 or used axial scan sampling procedures where the dimensions of the hippocampus and slice thickness work against accurate isotope recovery.37,39 This possibly resulted38 in a missed target. Other studies used the fully automated VBA that allows statistical comparisons of the whole brain at the single voxel level.12,13 Although it is well known that VBA incorporates a series of pre-processing steps that may affect the detection of MRglc changes from smaller regions, the spatial normalization accuracy of the hippocampus has not been studied.22 VBA studies presumably failed to report such hippocampal abnormalities due to spatial normalization errors. Only one VBA study40 reported a correlation between hippocampal MRglc and memory performance, but it too failed to observe hippocampal group differences.
By directly comparing the VBA to the gold standard coregistered coronal image ROI, we find large discrepancies between the two methods. Our results for the ROI sampling method show, as previously described,8–10 hippocampal MRglc reductions in both MCI (18%) and AD (27%). These differences remained significant after atrophy correction (see table 1). However, our VBA findings show no effects for the hippocampus, whereas they replicate previous reports of MRglc reductions in the association cortex for AD and MCI relative to NL.1–5 Our estimate of the normalization errors in the intersubject averaging of the hippocampus, as done with the VBA method, indicates inadequate spatial alignment of this structure in the aging and diseased brain. To overcome this obstacle we developed an automated and anatomically validated technique to sample the hippocampus on PET (HipMask). With the hippocampus sampled in an anatomically correct way, the HipMask showed virtually identical hippocampus MRglc estimates as compared with the ROI technique (see table 1).
Our results show that hippocampal MRglc measures, as obtained with both the ROI and the HipMask methods, are accurate and equivalent discriminators of NL and MCI groups. Both techniques identified 18% hippocampal MRglc reductions in MCI relative to NL with classification accuracies of 77% and 78%. Diagnostically, the hippocampal MRglc reduction in AD significantly added to neocortical MRglc measurements in the separation of NL and AD groups, although the cortical measures were the best discriminators. These findings are consistent with the universal observation that, by the time a patient presents with symptoms of dementia, substantial cortical MRglc reductions are already evident. Our data indicate that the assessment of hippocampus PET abnormalities may be particularly useful for characterizing the MCI stage, when memory loss is evident but functional deficits and neocortical damage are still minimal.
Correcting for multiple comparisons, our VBA results showed no cortical MRglc reductions in MCI as compared to NL. After removal of this conservative procedure, two clusters of hypometabolism became evident in the left posterior cingulate and left temporal cortex of the patients with MCI. Although these results are exploratory, they raise the important question of the utility of cortical deficits to characterize MCI and mild AD. Actually, several studies showed that cortical MRglc measures do not accurately classify patients with MCI,8,9,41 which can be interpreted as evidence for the relative absence of neocortical disease in early AD.5,38,42–47 On the other hand, studies that include amnestic patients with MCI report cortical MRglc deficits.5,10,30,40 Among our patients with MCI, only 5/13 fulfilled the diagnostic criteria for amnestic MCI.48 Therefore, our sample of patients may have been clinically less impaired, thus explaining the absence of severe cortical hypometabolism in comparison to NL. Our exploratory VBA results show greater PCC and TCx MRglc reductions in the five amnestic MCI as compared with the eight other MCI patients (PCC: t11 = 2.30, p = 0.04; TCx: t11 = 2.29, p = 0.04). The hippocampal MRglc was reduced in both MCI subgroups and did not differentiate them. Overall, the main observation from our MCI study was hippocampal MRglc reductions in the relative absence of cortical defects. These data stress the importance of hippocampal evaluations in the very early diagnosis of AD and suggest that hippocampal MRglc changes are early and may lose their discrimination value in MCI-AD comparisons, where other regions become more informative.
We have demonstrated that the HipMask method is suitable for the analysis of NL and patients with MCI and mild to moderate AD. Possible limitations to the HipMask approach, which may include generalizability to other patient samples and very atrophic hippocampi, are discussed in appendix E-2. Further studies are needed to test the mask in different groups of patients and in other clinical situations that can affect hippocampal size, spatial displacement, and function; for instance, vascular disease, space occupying lesions, or epilepsy.
The demographics of aging and recent US Medicare reimbursement approval for FDG-PET suggest a great need for rapid and accurate diagnostic tools to enable early diagnosis and to assist development of dementia prevention strategies. The anatomically validated hippocampal masking procedure we developed may be of use here, as it would be expected to improve the anatomic precision and speed of anatomic sampling for the hippocampal MRglc thus making it useful with large datasets. This approach could potentially be further developed for use with other brain regions. Prior work has suggested that a generally superior diagnostic accuracy is achieved using hippocampal FDG-PET measurements as compared with MRI volume measures in the classification of normal, MCI, and AD subjects.9 Further, the clinical diagnostic accuracy of PET is higher than that achieved with other clinical measures.49,50 The direct comparison between the PET HipMask and other established clinical measures would thus be of considerable interest in determining the optimal diagnostic assessment of early AD. The results of the present study also stimulate renewed interest in prior VBA-PET studies where hippocampal metabolic deficits were not found. This list would include previous PET studies of conversion to AD, genetic risk factors, and differential diagnosis.
Acknowledgment
The authors thank Joanna Fowler, David Schlyer, and Gene Jack Wang from the BNL PET team for support of the PET studies.
Footnotes
-
Additional material related to this article can be found on the Neurology Web site. Go to www.neurology.org and scroll down the Table of Contents for the June 14 issue to find the title link for this article.
Supported by NIH-NIA grants AG12101, AG13613, and AG08051.
Received October 1, 2004. Accepted in final form February 8, 2005.
References
Letters: Rapid online correspondence
REQUIREMENTS
If you are uploading a letter concerning an article:
You must have updated your disclosures within six months: http://submit.neurology.org
Your co-authors must send a completed Publishing Agreement Form to Neurology Staff (not necessary for the lead/corresponding author as the form below will suffice) before you upload your comment.
If you are responding to a comment that was written about an article you originally authored:
You (and co-authors) do not need to fill out forms or check disclosures as author forms are still valid
and apply to letter.
Submission specifications:
- Submissions must be < 200 words with < 5 references. Reference 1 must be the article on which you are commenting.
- Submissions should not have more than 5 authors. (Exception: original author replies can include all original authors of the article)
- Submit only on articles published within 6 months of issue date.
- Do not be redundant. Read any comments already posted on the article prior to submission.
- Submitted comments are subject to editing and editor review prior to posting.
