Empirical derivation of the reference region for computing diagnostic sensitive
18fluorodeoxyglucose ratios in Alzheimer's disease based on the ADNI sample☆,☆☆
Jerod M. Rasmussena,⁎, Anita Lakatosa, Theo G.M. van Erpa, Frithjof Kruggelb, David B. Keatora,
James T. Fallona, Fabio Macciardia, Steven G. Potkina
for the Alzheimer's Disease Neuroimaging Initiative
aDepartment of Psychiatry and Human Behavior, University of California Irvine, Irvine, CA, USA
bDepartment of Biomedical Engineering, University of California Irvine, Irvine, CA, USA
a b s t r a c t a r t i c l e i n f o
Received 6 May 2011
Received in revised form 23 August 2011
Accepted 13 September 2011
Available online 19 September 2011
Careful selection of the reference region for non-quantitative positron emission tomography (PET) analyses is
critically important for Region of Interest (ROI) data analyses. We introduce an empirical method of deriving
the most suitable reference region for computing neurodegeneration sensitive18fluorodeoxyglucose (FDG)
PET ratios based on the dataset collected by the Alzheimer's Disease Neuroimaging Initiative (ADNI) study.
Candidate reference regions are selected based on a heat map of the difference in coefficients of variation
(COVs) of FDG ratios over time for each of the Automatic Anatomical Labeling (AAL) atlas regions normalized
by allother AAL regions. Visual inspection ofthe heat map suggests thatthe portionof the cerebellum and ver-
mis superior to the horizontal fissure is the most sensitive reference region. Analyses of FDG ratio data show
increases in significance on the order of ten-fold when using the superior portion of the cerebellum as com-
pared with the traditionally used full cerebellum. The approach to reference region selection in this paper
can be generalized to other radiopharmaceuticals and radioligands as well as to other disorders where brain
changes over time are hypothesized and longitudinal data is available. Based on the empirical evidence pre-
sented in this study, we demonstrate the usefulness of the COV heat map method and conclude that intensity
normalization based on the superior portion of the cerebellum may be most sensitive to measuring change
when performinglongitudinal analysesofFDG-PETratiosaswellasgroupcomparisonsinAlzheimer's disease.
This article is part of a Special Issue entitled: Imaging Brain Aging and Neurodegenerative disease.
© 2011 Published by Elsevier B.V.
Careful selection of the reference region for computing radio-
pharmaceutical and radioligand ratios in non-quantitative positron
emission tomography (PET) studies is critically important, in partic-
ular for diseases that may differentially affect brain regions, such as
Alzheimer's disease (AD). While several studies have shown that
data-driven selection of a normalization region for PET ratio group
comparison studies is superior to global mean normalization [1–4],
no study to date has empirically derived the optimal denominator
region for computing degeneration sensitive
(FDG) ratios. The longitudinal design of the Alzheimer's Disease Neu-
roimaging Initiative (ADNI) study is uniquely suited to this data
Absolute values of cerebral metabolic rates of glucose (CMRgl) as
measured by FDG-PET in the ADNI sample show large amounts of
intra- and inter-subject and site (scanner) variability that limit data
analyses. This unwanted variance is significantly reduced by averag-
ing over multiple acquisitions and by computing FDG ratios based
on reference regions . Averaging increases the signal to noise
ratio and computation of ratios to a reference region cancels out
unwanted intra- and inter-subject and site variability, thereby facili-
tating greater ability to significantly detect smaller group differences.
It is critically important to choose a reference region that robustly
removes the unwanted variance but not the effects of the disease
and/or its progression. This may be particularly challenging for neu-
rodegenerative disorders that differentially affect much of the brain.
A selective review of the PET FDG literature on Alzheimer's disease
and aging shows the use of a wide variety of normalization regions,
including the whole brain [6–9], the cortex , the calcarine cortex
Biochimica et Biophysica Acta 456 (2012) 457–466
☆ This article is part of a Special Issue entitled: Imaging Brain Aging and Neurodegener-
☆☆ Data used in preparation of this article were obtained from the Alzheimer¹s Disease
Neuroimaging Initiative (ADNI) database (adni.loni.ucla.edu). As such, the investigators
within the ADNI contributed to the design and implementation of ADNI and/or pro-
vided data but did not participate in analysis or writing of this report. A complete list-
ing of ADNI investigators can be found at: http://adni.loni.ucla.edu/wp-content/
⁎ Corresponding author at: Department of Psychiatry and Human Behavior, School of
Medicine, University of California, Irvine, 5251 California Avenue, Irvine, CA 92617,
USA. Tel.: +1 949 824 7464; fax: +1 949 824 3324.
E-mail address: firstname.lastname@example.org (J.M. Rasmussen).
0925-4439/$ – see front matter © 2011 Published by Elsevier B.V.
Contents lists available at SciVerse ScienceDirect
Biochimica et Biophysica Acta
journal homepage: www.elsevier.com/locate/bbadis
and basal ganglia , the cerebellum , the pons [1,12–14], and the
combined cerebellum, vermis and pons . The use of different nor-
malization regions may contribute to inconsistent and spurious find-
ings across studies. Borghammer and colleagues [16,17] have shown
that an undetected decrease in normalization regions, e.g., mean
whole brain FDG, can result in the creation of artificial hypermetabo-
lism in unaffected regions and spurious relationships between FDG
ratios and demographic or other clinical variables.
Given the importance of normalization region selection and
available longitudinal data, we set out to empirically derive the
most suitable normalization region for longitudinal studies in Alz-
heimer's disease. We employ the use of visual inspection of a heat
map depiciting differences in coefficients of variation across time
among all Automated Anatomical Labeling (AAL) atlas regions, nor-
malized by all other AAL atlas regions, as well as the criterion of the
largest change over time. The most suitable ratios computed should
show the largest amount of change over time in regions known to be
affected by Alzheimer's disease. To our knowledge, this is the first
PET study to empirically derive the optimal denominator region for
computing degeneration sensitive18fluorodeoxyglucose ratios lon-
gitudinally in Alzheimer's disease.
We first examined differences in annualized percent change in
absolute and un-normalized FDG uptake between healthy elderly
and Alzheimer's disease patients in order to get an initial estima-
tion of regions that may show change in FDG uptake over time.
Subsequently, we derived candidate normalization regions based
on an Alzheimer's disease versus healthy elderly control coefficient
of variation difference in the heat map. We predicted that normal-
ization using the optimal denominator region would show the
largest change in FDG PET ratios over time compared with normal-
ization based on other candidate regions or whole brain FDG. Ad-
ditionally, we explored whether regions that show the largest
mean annualized percent degeneration in Alzheimer's disease
also show the largest group differences between AD patients and
healthy elderly volunteers. Given that AD is a progressive neurode-
generative disorder, we predicted a positive association between
mean annualized percent change in FDG PET ratio in Alzheimer's
disease and mean group differences between AD patients and
healthy elderly volunteers. Finally, we hypothesized that the FDG
ratios that are the most sensitive to degeneration also show the
strongest group differences when comparing AD patients with
healthy elderly volunteers.
2. Materials and methods
The ADNI study was approved by each of the participating sites' Institutional Review Boards (IRBs) and complied with the Code of Ethics of
the World Medical Association (Declaration of Helsinki). Written informed consent was obtained from all participants after they had received a
complete description of the study.
2.2. The Alzheimer's Disease Neuroimaging Initiative (ADNI)
ADNI was launched in 2003 by the National Institute on Aging (NIA), the National Institute of Biomedical Imaging and Bioengineering (NIBIB),
the Food and DrugAdministration (FDA),private pharmaceutical companies, andnon-profit organizations as a $60 million, 5-year public–private
partnership. The primary goal of ADNI is to test whether serial magnetic resonance imaging (MRI), positron emission tomography (PET), other
biological markers, such as cerebrospinal fluid (CSF) markers, APOE status and full-genome genotyping via blood sample, and clinical and neu-
ropsychological assessments can be combined to measure the progression of mild cognitive impairment (MCI) and AD. Determination of sensi-
tive and specific markers of very early AD progression is intended to: (1) aid in the development of new treatments, (2) increase the ability to
monitor their effectiveness, and (3) reduce the time and cost of clinical trials.
The principal investigator of the initiative is Michael W. Weiner, M.D., of the Veteran's Affairs Medical Center and University of California,
San Francisco. ADNI is the result of efforts of many co-investigators from a broad range of academic institutions and private corporations, and
participants have been recruited from over 50 sites across the U.S. and Canada. ADNI participants range in age from 55 to 90 years and include
approximately 200 cognitively normal elderly followed for 3 years, 400 elderly with MCI followed for 3 years, and 200 elderly with early AD
followed for 2 years. Participants are evaluated at baseline, 6, 12, 18 (for MCI only), 24, and 36 months (AD participants do not have a
36 month evaluation). Baseline and longitudinal follow-up structural MRI scans are collected on the full sample and 11C-labeled Pittsburgh
Compound-B (11C-PIB) and18FDG PET scans are collected on a subset every 6–12 months (for study details see http://www.adni-info.org)
2.3. Study samples
In this study, data analyses were conducted on two sub-samples derived from the complete ADNI sample. Both sub-samples were confined to
individuals with mild probable Alzheimer's disease (AD) and healthy control (HC) samples as determined at the baseline assessment. The first
sample comprised 117 individuals (60 AD and 57 HC) with baseline, 6, 12 and 24-month follow-up18FDG PET assessments. The second larger
sample comprised 199 individuals (97 AD and 102 HC) who had baseline18FDG PET assessments. General eligibility criteria for participation in
the ADNI study include age between 55 and 90 years, English or Spanish speaking, as well as the willingness and ability to undergo all test pro-
cedures including neuroimaging and longitudinal follow up assessments. Participants were assessed with the Mini Mental State Examination
(MMSE) , the Clinical Dementia Rating sum-of-boxes (CDR-sob) , and the Alzheimer's Disease Assessment Scale-cognitive subscale
(ADAScog) . Diagnoses of AD were based on NINCDS/ADRDA criteria  including MMSE scores between 20 and 26 and CDR scores 0.5
or 1.0. Healthy controls have MMSE scores between 24 and 30 and a CDR of 0, are non-depressed, non-MCI, and non-demented (for details
Comparisons of demographic data showed lower ADAS-cog scores, lower MMSE scores, higher CDR global values, similar mean age, similar
mean socioeconomic status, similar smoking history, as well as similar sex and handedness distributions in the AD compared with the HC groups
in both study samples. However, the AD groups had a lower education level and disproportionately higher APOE4 allele frequencies compared
with the HC groups (Table 1).
J.M. Rasmussen et al. / Biochimica et Biophysica Acta 456 (2012) 457–466
2.4. Image acquisition and preprocessing
High-resolution structural MRI and two pre-processed sets of18FDG PET baseline, 6, 12, and 24-month follow-up assessment brain scans were
adni.loni.ucla.edu/research/protocols/mri-protocols). The18FDG pet datasets included absolute and intensity normalized scans. The absolute scans
were co-registered and averaged by ADNI, while intensity normalized scans were also standardized with regard to image resolution, voxel size, and
smoothness. The procedures are detailed on the ADNI website (see points 2 and 4 of the “Processed PET image data” section at http://adni.loni.ucla.
edu/about-data-samples/image-data/). Briefly, all raw PET images were converted to a standard file format. Six five-minute frames 30–60 minute
post18FDG-injection were extracted from the image file. The extracted frames were registered to the first frame (acquired at 30–35 min post-injec-
tion) and averaged to yield a single 30-minute average PET image in “native” space with a resolution of 128×128×63 voxels and a 2 mm isotropic
resolution.The average image of the baseline scan was reoriented such that the anterior–posterior axis of the image grid is parallel to theAC-PC line
and the image was re-sampled to a resolution of 160×160×96 voxels with a 1.5 mm isotropic resolution. All individual frames from each PET scan,
baseline and follow-up, were then co-registered to this reoriented baseline reference scan and an average image for each assessment point was cre-
ated. Each average image was intensity normalized using a subject-specific mask, in that the average intensity of voxels within the mask becomes
exactly one. Each voxel intensity value is therefore an FDG/mean (of whole brain) ratio. Datasets were smoothed to a uniform isotropic resolution
of 8 mm full width half maximum, the approximate resolution of the lowest resolution scanners used in ADNI. The specific filter settings were de-
termined from the Hoffman phantom PET scans that were acquired during the ADNI certification process. Prior to image analysis all structural and
PET scans were converted from DICOM to Brain Image Analysis (BRIAN) format.
2.5. Extraction of mean FDG for AAL atlas ROIs
To compensate for structural differences between diagnoses during registration, separate anatomical group atlas templates were generated
for the AD patients and healthy elderly controls. Briefly, baseline T1-weighted MRI scans were aligned with the stereotaxic coordinate system
. A quality check was performed on the successful conversion and registration. The outer hulls of the brain were removed by a registration
approach , yielding the “peeled” brain in the intracranial space (IC). Subsequently, all brain datasets were registered with the MNI 152 brain
using a recent approach for nonlinear registration based on fluid dynamics . All registered brains were averaged, correcting for their mean
whole brain intensity, to yield the ADNI template 1. All brain datasets were registered to the ADNI template 1, and averaged voxelwise to yield
the ADNI template 2. All brain datasets were registered with their group specific ADNI template 2, resulting in a deformation field for each sub-
ject. Preprocessed PET data were registered with the baseline “peeled” MRI brain using rigid body registration and normalized mutual informa-
tion as similarity criterion. The deformation fields were used to warp the18FDG PET data into MNI space. The ADNI template 2 in MNI space was
divided into ROIs using the AAL atlas  and absolute as well as whole-brain normalized FDG means for each of the AAL template ROIs were
2.6. Longitudinal analysis of absolute FDG data
To determine whether glucose metabolism in the various reference regions changed over time between comparison groups, we performed a
longitudinal analysis on the absolute18FDG PET ROI data. Absolute data was examined to take advantage of the unique repeated measures de-
sign of the ADNI PET sample while uncoupling the measurements from the whole brain normalization step that uses a scale factor not publicly
available from ADNI. Given that the absolute data is not normalized by any region (whole brain or otherwise) and therefore has a high degree of
variability, we computed the annualized percent signal change during the two year follow-up period based on the three available follow up as-
sessments Eq. (1). As a ratio, this metric removes site and subject variance and averages visit-related noise effects. The annualized percent signal
change (δ) was computed using the mean signal (S) in each subject (N) and ROI (j).
Sample 1 and sample 2 demographics.
Sample 1 Sample 2
Mean years of education
Scan resolution (low/high)
0/59/1 0/56/1FET1.0 1/99/21/94/2FET 0.99
Mean±standard deviation (SD) and the frequency of each category are represented with test statistics. ** Denotes significance at pb.005.
FET = Fisher's Exact Test (2-tailed).
J.M. Rasmussen et al. / Biochimica et Biophysica Acta 456 (2012) 457–466
Eq. (1). Annualized Percent Signal Change in Absolute18FDG by Subject and ROI.
4 S6 N;j
4 S12 N;j
2 S24 N;j
Group means and standard errors for AD patients and healthy elderly controls were calculated for every ROI. To remove extreme outliers, the
data was systematically trimmed prior to the annualized percent change calculation. The optimal trim threshold was set at 10% of the time point
absolute differences in each group to prevent rejection of true effect related outliers and was robust within 3% of this threshold. A larger thresh-
old rejected true effects and drove the baseline up, while a lower threshold included extreme outliers and introduced instability in the baseline
value. All trimmed outliers reflected nonsensical changes greater than 100% annually.
2.7. Empirical derivation of FDG ratio denominator region
To identify the reference region(s) that will result in18FDG ratios that are most sensitive to small changes in metabolism in key brain regions
known to be affected by Alzheimer's disease, we computed coefficients of variation (COVs) for each of the 116 ROIs normalized by all of the 116
ROIs. COVs are capable of capturing linear and non-linear dispersion of FDG ratios over time. Only ROI data derived from the intensity-normal-
ized18FDG PET data were used for the computations. The approach taken is formalized in the following paragraphs.
First, we created a matrix Aij, comprised of 116 unique FDG ratios of one ROI to another for each of the 116 AAL atlas ROIs. This matrix com-
prises 116 rows by 116 columns and was created for each subject and time point in the sample data set.
Eq. (2). Ratio matrix definition.
Second, we concatenated the 2-dimensional matrix A over time (baseline, 6, 12, and 24 months) resulting in a 3-dimensional matrix for each
subject with 116 rows by 116 columns by four time points. The COV, ci,j, in each element of the matrix was defined by collapsing across time
using Eq. (3), where μ and σ are the mean and standard deviation across time, respectively.
Eq. (3). COV.
Third, we concatenated matrix c across subjects into the resulting matrix C with 116 rows by 116 columns by Nsubjects,where each cell con-
tains a COV. This matrix is collapsed once more across AD patients and healthy elderly controls by taking the mean for each group. This results in
the two final heat maps, one for each group, with 116 rows and 116 columns containing the mean COVs for each of the AAL regions normalized
by all other AAL regions. The matrix bCOVN is the mean normalized dispersion over time across subjects for a diagnostic group. This measure
reflects the generalized dynamic nature of a given ROI ratio over time.
Eq. (4). bCOVN, heat map.
Finally, to assess the differences in temporal dynamics between the groups, AD and HC were contrasted through a simple subtraction bCOVΔN=
2.8. Longitudinal analysis of FDG ratios based on candidate reference regions
Based on the empirical results of COV and longitudinal analysis of absolute FDG data, candidate reference regions were identified and exam-
ined for their respective sensitivities to relative metabolic change in the AD and HC samples. Using whole brain as a reference region, percent
change relative to baseline was calculated for each subject and time point in a region traditionally considered to be stable across time (bilateral
basal ganglia sans caudate) as well as a region typically implicated in AD (bilateral posterior cingulate). Each time point was averaged across
subjects, keeping diagnosis separate. This calculation was repeated in three reference regions in addition to whole brain that were chosen to
best reflect reference regions popular in current literature while carrying forward the results of the COV heat map: basal ganglia without cau-
date, entire cerebellum, and bilateral superior cerebellum including vermis. The superior cerebellum ROI referenced here and throughout this
work is defined by those areas superior to the horizontal fissure and include AAL atlas regions: cerebellum III, IV, V, VI as well as the entire
2.9. Association between annualized percent change and group difference
Longitudinal changes in CMRgl and cross-sectional group differences in CMRgl were examined to assess the inferences they may have on one
another. Group difference was defined and calculated as the percent difference between AD and HC relative to HC in all AAL regions, normalized
by bilateral superior cerebellum and vermis. The normalization region was chosen to reflect the region identified by the COV heat map analysis.
This was done for every time point and averaged in each ROI. Annualized percent change was defined relative to baseline and was calculated
J.M. Rasmussen et al. / Biochimica et Biophysica Acta 456 (2012) 457–466
with the same reference region as the group difference metric. These were then plotted against each other and correlated to yield a Pearson's
correlation coefficient, r.
2.10. Statistical analyses
In the longitudinal analysis of absolute18FDG data, two sample t-tests were performed for every ROI to determine differential changes in
18FDG uptake over time between AD patients and healthy elderly controls. In the COV analysis, candidate denominator reference regions
were selected based on regions showing the largest difference in COV by visual inspection of the AD minus HC COV subtraction heat map
(Fig. 6). In addition, the prevalence of denominator regions used in the established Alzheimer's disease PET literature as well as evidence of
no differential change over time based on the longitudinal analysis of absolute FDG uptake were taken into account. Where multiple AAL
ROIs were combined to form the denominator region, weighted means based on ROI intensities and ROI volumes were computed.
To identify the reference region that resulted in the best separation of AD patients and healthy elderly controls, univariate ANCOVAs predict-
ing FDG ratios for each selected reference region (whole brain, basal ganglia, cerebellum, cerebellar vermis) with diagnosis (AD, HC), covarying
for age, sex, and race, were performed. Group contrasts (t-tests) were performed to assess which regions showed significantly different FDG ra-
tios between AD and HC. A Bonferroni correction based p-value (b0.0004) was used to control for multiple comparisons (116 ROIs). The AAL
atlas regions showing significant group contrast for each of the reference regions are presented on a standard brain (Fig. 6).
3.1. Longitudinal analysis of absolute18FDG data
The longitudinal analysis showed that FDG uptake in superior cer-
ebellum regions did not decrease in either AD patients or healthy el-
derly controls during the two-year follow-up period (Fig. 1 top,
middle). The comparison of annualized absolute percent signal
change between healthy controls and patients showed non-signifi-
cant global declines in whole brain FDG update in AD patients
(3.3%) and healthy elderly controls (0.3%). Group contrasts showed
several cerebral ROIs that had significantly larger rates of decline in
AD patients compared with controls (Fig. 1 bottom). However, none
of these differences survived Bonferroni correction for multiple com-
parisons. AD patients did show significantly larger decreases in FDG
uptake in the supratentorial (3.5%) when compared with the infra-
tentorial (1.6%) region (pb0.0001). Comparable rates in controls
were 0.38% vs. 0.15%. There was some regional variation in the level
of decline among the supratentorial regions, with the basal ganglia
sans caudate at 2.4%, the parietal lobe at 3.1%, the occipital lobe at
3.3%, the frontal lobe at 3.3%, the limbic areas at 4.1%, and the tempo-
ral lobes at 4.4% decrease in glucose metabolism per year.
3.2. Coefficient of variation analysis
Visual inspections of the group COV heat maps show a clear global
COV increase in AD when compared to HC (Fig. 2). Despite the global
difference, there are a number of similar features between the two
maps. Notably, the map bound by the cerebellum regions vertically
and other regions horizontally (AD .044+/−.007, CTRL .036+/−.006)
as well as the limbic regions vertically and other regions horizontally
(AD .045+/−.008, CTRL .038+/−.006) show increased COV relative
to the rest of the map (AD .036+/−.009, CTRL .029+/−.007). Two re-
gions in both groups stand out (the bilateral mid-orbital frontal and bi-
lateral cerebellum 10 AAL regions) as showing especially large
agonal that have high spatial correlation, are expectedly smaller than
interlobe pixels, those regions that are found well off diagonal with
low spatial correlation. An exception to this observation can be seen
within a subsection of the infratentorial region (bilateral cerebellum
Fig. 1. Annualized percent change (+/− standard error) in un-normalized CMRgl between baseline, six, twelve and twenty-four-month follow-up across all AAL atlas regions of
interest. Top row: Healthy Controls (HC), middle row: probable Alzheimer's disease (AD), bottom row: significance difference between groups. Each dot represents a single ROI
defined in the AAL atlas.
J.M. Rasmussen et al. / Biochimica et Biophysica Acta 456 (2012) 457–466
10). Cooler regions well off the diagonal are reflective of low variability
relative to the mean, due to temporal CMRgl correlation between the
numerator and denominator that is independent of spatial correlation.
Visual inspection of the (AD–HC) COV contrast map addresses the
primary aim of the COV analysis: to differentiate, between groups,
which normalizing regions are the most reflective of both linear and
non-linear regional changes over time. Hotter features reflect in-
creased CMRgl temporal dynamics in AD compared to HC. Because
each pixel's temperature is a reflection of one region normalized by
another, the map is symmetric. As an example of this, it is clear that
the temporal lobe has a higher COV whether it is the numerator or
the denominator, demonstrated by the hot bars in the temporal
lobe both vertically (temporal as the denominator) and horizontally
(temporal as the numerator). AD subjects have larger COV values
across 95% of all brain regions, a testament to the degenerative nature
of the disease. Those regions that are preferentially targeted by the
disease show larger COV increases from HC to AD, in particular tem-
poral lobe, parietal lobe and the limbic system. Specific features moti-
vate the selection of reference regions in Section 2.8. Candidate
reference regions are composed of several hot regions with the
exclusion of cool regions within the area (jump out boxes in Fig. 3).
Hot regions include the basal ganglia (excluding caudate which is
cool), a subset of cerebellum including regions III, IV, V and VI with
the entirety of the vermis, the full cerebellum and the whole brain.
3.3. Longitudinal analysis of FDG ratios based on candidate
Analysis of percent change over time in two ROIs, one implicated
in disease and one thought to be spared by the disease, showed
marked differences between the perceived rate of change in CMRgl
based on the reference region used to compute the ratios. The poste-
rior cingulate was identified by all four reference regions to be signif-
icantly different from AD at the 24-month visit. However, only full
and partial cerebellum reference ROIs were able to wean out this ef-
fect by the one year point and only by using the partial cerebellum
as the denominator region were we able to show group differences
at the earliest follow up visit, 6 months. In addition to being the
most significant, partial cerebellum proved to be the most sensitive
to relative CMRgl changes, having an average decline of 8.7% in
2 years for bilateral posterior cingulate ROIs. The average control sub-
ject declined in CMRgl at all three follow-up time points relative to all
reference regions, with a maximum of no more than 1.4% per year.
Examining a stable region yielded dramatically different results. Re-
dundantly, basal ganglia as the reference region showed no change
over time, essentially a null result as this region should roughly nor-
malize to unity at all time points. Provocatively, basal ganglia regions
were shown to be significantly decreased relative to full and partial
cerebellum at 12 and 24 months respectively. CMRgl decline was as
great as 2.8% over 2 years using partial cerebellum as the denomina-
tor. Whole brain normalization, on the other hand, suggested neuro-
degeneration in the basal ganglia to be non-existent in AD (Fig. 4). On
the contrary it was shown to significantly increase by the two year
mark by 2.7%.
3.4. Association between annualized percent change and mean effect size
In order to make inferences about AD from longitudinal changes in
CMRgl it is important to know the relationship between it and cross-
sectional group differences. Annualized percent changes relative to
partial cerebellum in all ROIs were in the range of −4.8% to +1.0%.
The measured group difference, relative to HC, had a range of
−19.2% to +1.8%, reflecting the cumulative toll of the disease on neu-
rodegeneration. Plotting and correlating these two metrics (Fig. 5)
showed a significantly strong correlation between the two (r=.84).
There is suggestion of an asymptote at the 0% mark, populated mostly
by those ROIs involved in normalization.
3.5. GLM analysis of group differences
As a final metric for comparison of reference regions, t-tests of the
ANCOVA model beta values were performed to highlight similarities
and differences in identifying significant group differences between
the regions used in normalization (Fig. 6). Using the whole brain as
a reference region, as done in previously published ADNI studies,
large portions of the cerebellum and basal ganglia were uniquely
identified as being affected by the disease. The regions deemed signif-
icant no matter the choice of reference regions, were limited to part of
the temporal lobe, angular gyri, hippocampus and posterior cingulate.
The reference region identified in this analysis that revealed the larg-
est ROI group differences was the superior cerebellum and vermis. In
addition to identifying all regions identified by basal ganglia and the
full infratentorial ROI, this specific reference ROI identified multiple
regions of the frontal lobe, as well as the R mid temporal pole, R ante-
rior cingulate, and superior parietal and occipital lobes. Overall, 94 of
the 116 regions were identified as affected by AD with one of the
Fig. 3. Coefficient of Variance (COV) heat map, AD–NC. Each pixel represents an effect
size ROI normalized by a reference ROI. Note decreased COV in AAL Cerebellum 1–2
and 7–10 relative to the rest of the cerebellum, as well as the smaller COV present in
the caudate region of basal ganglia. Both of these observations are motivating factors
for the modified reference regions used in the statistical analysis portion of the
Fig. 2. COV heat maps of NC and AD groups. Hotter pixels reflect greater relative time
variability, indicative of increased sensitivity to metabolic change differences in the
AD population. Map similarities include COV hyperintensities in cerebellum and limbic
regions. Cooler regions indicate spatial correlation (those regions near to the diagonal)
and/or temporal correlation (those regions well off diagonal). Temporal correlation
suggests coherence in rate of CMRgl change over time. CV: cerebellar vermis, C: cere-
bellum, TL: temporal lobe, BG: basal ganglia, PL: parietal lobe, OL: occipital lobe, LS:
limbic system, and FL: frontal lobe.
J.M. Rasmussen et al. / Biochimica et Biophysica Acta 456 (2012) 457–466
above reference regions. As the most sensitive region for differentiat-
ing groups was the superior cerebellum which accounted for 68% of
the 94 regions. This is in contrast to 54%, 33% and 44% for infratentor-
ial, basal ganglia, and whole brain respectively. In addition to the in-
creased spatial extent of group differences, ROIs typically implicated
in the disease show large increases in significance when using the su-
perior portion of the cerebellum relative to the cerebellum in full. For
example, L posterior cingulate and L hippocampus showed 20- and
50-fold gains in significance respectively.
In the attempt to empirically derive the optimal reference region
for detecting neurodegenerative metabolic decline in Alzheimer's dis-
ease, five principal findings emerged: 1) while large extents of the ce-
rebrum and inferior cerebellum do decrease in CMRgl over time in
AD, the superior cerebellum including vermis does not, 2) FDG ratios
normalized by superior cerebellum and basal ganglia regions have the
highest COVs in Alzheimer's disease patients when compared to
healthy elderly volunteers, 3) among the candidate normalization re-
gions, FDG ratios normalized with the superior cerebellum show the
steepest rate of decline over time when compared with whole
brain, basal ganglia and full cerebellum reference regions 4) mean
annualized percent signal change in AD patients is strongly correlated
with mean cross-sectional AD-HC group differences and 5) among all
candidate reference regions, superior cerebellum is the most sensitive
in identifying AD–HC group differences.
In order to express regional changes over time independent of all
other regions it is imperative to use absolute, non-normalized data.
ADNI's preprocessing pipeline includes measures to account for dif-
ferences in site, subject and visit. This includes whole-brain normali-
zation, a preprocessing step that renders data no longer independent
of other voxels. Because of this, absolute unprocessed data is needed
if one chooses to make statements regarding absolute changes over
time free of influence from other voxels. In addition to site differ-
ences, the unprocessed data is susceptible to physiological and visit
differences within subject. Variation in FDG PET determined CMRgl
has been shown to be tightly coupled to CBF , making it a relevant
factor in examining day to day variations in PET quantification. Sub-
ject variation in CBF is a function of both age and sex, among other
factors, and can have a normal CBF range of 43–69 (ml/100 g/min)
when measured by Arterial Spin Labeling (ASL) . Visit differences
are a function of, among other things, day-to-day variation in CBF.
These fluctuations have been shown to be roughly 9% between one-
week session intervals in ASL [28,29]. Using a percent absolute signal
change within subject, across time, one can divide out both the site
and subject level variances. However, visit level variance will still be
Systematic trimming played a deterministic role in the baseline
value of absolute annualized percent change. As a numerical example
without trimming, HC subjects were observed to have a greater than
3% decrease on average per year in CMRgl in all ROIs, while 87 of the
116 AD ROIs had an overall annual increase in CMRgl. These non-
trimmed results were viewed with skepticism considering they in-
cluded visits with annualized percent changes greater than 200%
across the entire brain. As the trim threshold was increased to 10%,
baseline values converged asymptotically to those shown in Fig. 1.
All values trimmed in this manuscript were extreme outliers of annu-
alized percent change greater than 100% across the brain between
visits. For trim thresholds above the ten percent used in this study,
the baseline AD values slowly increase. The trimmed values in this
case were directionally biased in the negative direction, likely due
to true disease effects. The final trim factor of 10% was selected
based on the ability to remove extreme outlying values while retain-
ing meaningful disease effects.
When determining the extent of longitudinal percent signal change,
it is necessary to have greater temporal variability than visit variability.
Fig. 5. Mean effect size as measured by the difference between normalized HC and AD,
averaged across all four time points, plotted against annualized18FDG percent change
for all 116 AAL ROIs. A strong correlation between the cross-sectional effect size and
longitudinal rate of neurodegeneration is seen.
Fig. 4. Longitudinal example of relative neurodegeneration by denominator region, the ROI on the left (LR combined posterior cingulate) is traditionally associated with AD, while
the ROI on the right (LR combined pallidum, putamen and thalamus) is traditionally thought to be spared by the disease. HC and AD represent blue and red lines respectively. Com-
bined cerebellum III, IV, V, VI and vermis as a reference region was shown to be the most sensitive to neurodegeneration, while whole brain normalization significantly identified
the basal ganglia regions to be increasing with disease progression.
J.M. Rasmussen et al. / Biochimica et Biophysica Acta 456 (2012) 457–466
Because of this, effect sizes between contrasts need to be large. Despite
the large variability present in the longitudinal data, two significant
trends from the analysis were clear: 1) cerebrum CMRgl decrease over
time was greater in AD than in NC and 2) cerebrum CMRgl decrease
over time was greater than superior cerebellum CMRgl decrease over
time, implying that the superior cerebellum is relatively stable. In addi-
tion, rates of change in typically disease-preferred areas are in confor-
mance with previously published findings. These findings demonstrate
the heterogeneity present in the dynamics of the disease progression.
Not only is this limited to the cerebrum, but to inferior portions of the
cerebellum as well, a region classically thought to be relatively spared
by the disease. It is precisely this heterogeneity that makes whole
brain normalization biased and suggests superior cerebellum as being
Finding the ideal reference region has been cast here as a data-
driven problem. COV is a measure of relative stability over time; a
small COV suggests little relative change over time, while a large
COV points to dynamic behavior. Using a brute force method allows
immediate insight into the problem by visual inspection of a single
heat map. Initially, based on the longitudinal analysis, one may as-
sume the cerebellum to have the smallest COV, however this is not
the case. Recall that these are the ratios of one ROI to another; there-
fore a small COV often indicates that there is a collinear relationship
between two ROIs. Adversely, a large COV indicates that one ROI is
changing relative to another, an ideal condition for normalization.
As an example of known collinearity, one can look at the square fea-
tures centered on the matrix diagonal. These are regions that are cor-
related with tissue very close in space and function to itself. Further
visual examination of the features in the group contrast COV map
yields immediate insight into the effectiveness of both individual
and grouped ROIs. For example, in line with the longitudinal analysis
there is a clear difference in using inferior versus superior regions of
the cerebellum as a normalizing region, demonstrated by higher in-
tensities horizontally in superior versus inferior cerebellum.
It should be emphasized however, that COV is not a change in
overall decreased activity, but the rate of change of activity. Because
the numerator of COV is standard deviation, this can be increased
noise just as easily as it can be a linear or nonlinear decrease. Noise
sources can be attributed to, among other things, mis-registration.
Additionally, the denominator of the COV is mean CMRgl across
time, which one can argue is smaller in AD patients, particularly in
the late stages of the disease. However, this does not discount the
metric since the phenotype used in further analysis will benefit
from the same scaling bias. Finally, the quantitative analysis of the
COV heat maps was largely limited in scope to simple statistics
while more sophisticated pattern recognition and clustering methods
may yield additional important observations.
In order to validate the assumptions made by qualitative visual in-
spection of the COV heat maps and inherently noisy non-normalized
longitudinal data we chose to evaluate two ROIs using four reference
regions identified as candidates for sensitive normalization. The two
ROIs chosen, bilateral posterior cingulate and basal ganglia without
the caudate, reflected what we believed to be strong decline due to
disease and relative sparing respectively. Looking only at posterior
cingulate, we demonstrated further motivation for using only the su-
perior portion of the cerebellum. Not only did it prove more signifi-
cant than all of the other three candidate reference regions in
identifying disease related changes, it did so earlier, significantly
identifying AD from HC after only a six month time span. Looking at
the normalization effects on basal ganglia corroborated the pitfalls
of using whole brain normalization in Alzheimer's disease as reported
by Borghammer and colleagues [16,17]. The authors found basal
ganglia whole-brain ratios in Alzheimer's disease patients appear to
increase with age, which is most likely due to larger decreases in
whole brain FDG uptake over time compared with decreases in
basal ganglia FDG uptake over time.
To focus on the COV methodology the principal findings of this
study were limited to the groups with the clearest diagnosis: healthy
controls versus probable AD. However, the ADNI sample contains a
group of potential early onset subjects labeled as Mildly Cognitive Im-
paired (MCI). Over the course of the study, subjects that are later di-
agnosed as AD are considered converters, those that are not, are
labeled as non-converters. While the interpretation and findings
using the COV method on this sample are thought to be beyond the
focus of this manuscript, a brief summary of the results can be
found in the supplementary material found accompanying this man-
uscript, demonstrating degenerative traits found in the MCI group
midway between HC and AD. The figures found in this material sup-
port the use of COV as a metric for identifying relative dynamic differ-
ences among groups and demonstrate the dynamics found within the
MCI sub-sample for posterior cingulate, occipital lobe and superior
The ADNI study is unique in that it provides a large longitudinal
sample with which to make inference on the dynamics of the dis-
ease. With absolute longitudinal analysis, COV techniques, and nor-
malized longitudinal analysis, we have shown the superior
cerebellum to consistently be more sensitive in identifying dynamic
changes brought on by the disease than all other AAL ROIs. While
dynamic changes are important in understanding the disease, most
clinical visits are inherently cross-sectional. Because of this, linking
longitudinal changes in each ROI to cross-sectional effects already
present at the time of appointment is crucial. We demonstrated
strong correlations between the mean cross-sectional effect size
and longitudinal percent change. This can be interpreted as, at
least in the mild AD stage of the disease, no ROI having detectable
Fig. 6. AAL atlas regions showing significantly lower FDG ratios in Alzheimer's disease patients compared with healthy elderly volunteers by denominator region. Colors reflect sig-
nificant differences for a given reference region. As an example, ROIs shown to be significant whether using the subset of cerebellum (AAL cerebellum III–VI and vermis) or the
entire infratentorial space as a reference region are shown in red. Images shown are in neurological orientation, with the MNI axial coordinate shown in blue.
J.M. Rasmussen et al. / Biochimica et Biophysica Acta 456 (2012) 457–466
saturation in decline. If an ROI were to have declined to the point of
saturation before the AD stage of disease it would elicit a large cross-
sectional effect size while showing little or no percent change over
Demonstrating annualized percent signal change in AD patients as
significantly associated with mean group difference is consistent with
the fact that Alzheimer's disease is a progressive neurodegenerative
disease. In addition it demonstrates that normalization with the supe-
of decline with age, but also shows the largest group differences when
comparing PET FDG ratios between Alzheimer's disease patients and
othergroups. ThecomparisonbetweenAD patients andhealthyelder-
ly volunteers on all AAL atlas regions using the full GLM confirms this.
The pitfalls of normalization with mean whole brain FDG are consis-
tent with recentfindings by Langbaumand colleagues, Alzheimer's
disease patients showed higher bilateral cerebellum, and sensory
motor cortex FDG-whole brain ratios when compared with controls.
Conversely, the largest group differences are observed when normal-
izing with the superior cerebellum as bisected by the horizontal fis-
sure. Coupling this finding with demonstrated stability over time
and increased sensitivity to dynamic disease effects provides strong
evidence for the use of superior cerebellum as a reference region.
The superior cerebellum defined here can be interpreted as the
spinocerebellum and associated vermis together with the cerebellum
VI superior to the horizontal fissure. Functionally these areas reflect
limb positioning and speech articulation. More importantly, these
areas are unique in terms of their blood supply. The superior cerebel-
lum receives its blood supply from the superior cerebellar artery
(SCA), whereas the areas inferior to the horizontal fissure receive
their supply from the anterior inferior (AICA) and posterior inferior
cerebral arteries (PICA). The SCA is adjacent to the posterior cerebral
artery (PCA) and might in fact have compensatory reactivity to block-
age in the PCA which supplies precuneus, posterior and anterior cin-
gulate. These are among the primary areas along the midline cerebral
hemispheres affected in AD. The blood flow from the AICA and PICA
are far more related to additional vertebral and basilar artery circula-
tion than the SCA and PCA, which are more closely related to addi-
tional Circle of Willis circulation. It is possible that the reduced flow
in medial cerebrum regions due to cerebral amyloid angiopathy (re-
duced FDG uptake in medial cerebrum regions) results in a compen-
satory increase in blood flow to the superior cerebellum with a
corresponding relative increase in FDG uptake in the superior cere-
We are aware that many recent studies have used the pons as a
reference area, this study was limited to regions that currently have
a standard definition in the AAL atlas. Atlas based delineations of
ROIs have considerable advantages in standardization and analyses
of large data sets. Future work on reference region selection should
consider using standard atlases that do include the pons or indepen-
dently generate a pons reference region based on the template, de-
spite the systematic error this may incur between studies. In
addition, the relatively small size and scarcity of gray matter in the
pons makes it susceptible to increased variability. Another technique
commonly used is to restrict the regions to gray matter tissue. While
we feel that the hypothesis behind this technique is strong, the stan-
dard filter of 8 mm isotropic used by ADNI largely obfuscates gray and
white matter introducing partial voluming effects and increases the
necessity of accurate registrations to the template, potentially creat-
ing a group bias.
This work has systematically isolated the superior portion of the
cerebellum, including vermis, as being the most sensitive normaliz-
ing region for detecting both rates of decline and baseline deficits in
Alzheimer's disease. The compiled evidence includes: 1) superior
cerebellum was demonstrated to be stable over time in AD patients
while the majority of the cerebrum is in decline (including the infe-
rior cerebellum), 2) COV analysis isolated the basal ganglia along
with the cerebellum (particularly superior cerebellum) as being
the most sensitive normalizing region over time, 3) grouping ROIs,
superior cerebellum detected group differences 6 months before
full cerebellum and 18 months before whole brain and basal ganglia
normalization, 4) rates of decline were significantly correlated with
and therefore indicative of already present disease effects when
using the superior cerebellum, and 5) the extent and magnitude of
baseline differences between AD and HC were greatest when nor-
malizing by superior cerebellum.
The authors wish to thank the patients and healthy volunteers
who participated in the study. This study was supported by grants
tothe Transdisciplinary Imaging
RR020837-01), the Alzheimer's Disease Neuroimaging Initiative
(ADNI U01 AG024904-01), and supplement (3U01AG024904-03S5),
the National Institute of Aging, the National Institute of Biomedical
Imaging and Bioengineering (NIH), the Functional Imaging Biomedi-
cal Informatics Research Network (FBIRN U24-RR021992, National
Center for Research Resources), the NIH through the following
NCRR grant: the Biomedical Informatics Research Network (1 U24
RR025736-01), commercial support from Vanda Pharmaceuticals,
and private support from an anonymous Foundation and anonymous
donations. Additional contributions made through the Foundation for
the NIH from Merck & Co. Inc., Pfizer, Inc., and Gene Network Sci-
ences, Inc. partially supported the genotyping results reported here.
Data collection and sharing for this project was funded by the Alzhei-
mer's Disease Neuroimaging Initiative (ADNI; Principal Investigator:
Michael Weiner; NIH grant and supplement). ADNI is funded by the
National Institute on Aging, the National Institute of Biomedical Imag-
ing and Bioengineering (NIBIB), and through generous contributions
from the following: Pfizer Inc., Wyeth Research, Bristol-Myers Squibb,
Eli Lilly and Company, GlaxoSmithKline, Merck & Co. Inc., AstraZeneca
AB, Novartis Pharmaceuticals Corporation, Alzheimer's Association,
Eisai Global Clinical Development, Elan Corporation plc, Forest Labo-
ratories, and the Institute for the Study of Aging, with participation
from the U.S. Food and Drug Administration. Industry partnerships
are coordinated through the Foundation for the National Institutes
of Health. The grantee organization is the Northern California Insti-
tute for Research and Education, and the study is coordinated by the
Alzheimer's Disease Cooperative Study at the University of California,
San Diego. ADNI data are disseminated by the Laboratory of Neuro
Imaging at the University of California, Los Angeles. The funders
had no role in study design, data collection and analysis, decision to
publish, or preparation of the manuscript. R Grant Number UL1
Appendix A. Supplementary data
Supplementary data to this article can be found online at doi:10.
 S. Minoshima, K.A. Frey, N.L. Foster, D.E. Kuhl, Preserved pontine glucose metab-
olism in Alzheimer disease: a reference region for functional brain image (PET)
analysis, J. Comput. Assist. Tomogr. 19 (1995) 541–547.
 J.L. Andersson, How to estimate global activity independent of changes in local
activity, Neuroimage 6 (1997) 237–244.
 I. Yakushev, A. Hammers, A. Fellgiebel, I. Schmidtmann, A. Scheurich, H.G.
Buchholz, J. Peters, P. Bartenstein, K. Lieb, M. Schreckenberger, SPM-based
count normalization provides excellent discrimination of mild Alzheimer's dis-
ease and amnestic mild cognitive impairment from healthy aging, Neuroimage
44 (2009) 43–50.
J.M. Rasmussen et al. / Biochimica et Biophysica Acta 456 (2012) 457–466
 P. Borghammer, J. Aanerud, A. Gjedde, Data-driven intensity normalization of PET Download full-text
group comparison studies is superior to global mean normalization, Neuroimage
46 (2009) 981–988.
 P.T. Fox, M.A. Mintun, E.M. Reiman, M.E. Raichle, Enhanced detection of focal
brain responses using intersubject averaging and change-distribution analysis
of subtracted PET images, J. Cereb. Blood Flow Metab. 8 (1988) 642–653.
 M. Kushner, M. Tobin, A. Alavi, J. Chawluk, M. Rosen, F. Fazekas, J. Alavi, M. Reivich,
Cerebellar glucose consumption in normal and pathologic states using fluorine-
FDG and PET, J. Nucl. Med. 28 (1987) 1667–1670.
 R. Mielke, U. Pietrzyk, A. Jacobs, G.R. Fink, A. Ichimiya, J. Kessler, K. Herholz, W.D.
parison of perfusion and metabolic pattern, Eur. J. Nucl. Med. 21 (1994) 1052–1060.
 J.B. Langbaum, K. Chen, W. Lee, C. Reschke, D. Bandy, A.S. Fleisher, G.E. Alexander,
N.L. Foster, M.W. Weiner, R.A. Koeppe, W.J. Jagust, E.M. Reiman, Categorical and
correlational analyses of baseline fluorodeoxyglucose positron emission tomog-
raphy images from the Alzheimer's Disease Neuroimaging Initiative (ADNI),
Neuroimage 45 (2009) 1107–1116.
 W.S. Liang, K. Chen, W. Lee, K. Sidhar, J.J. Corneveaux, A.N. Allen, A. Myers, S. Villa,
B. Meechoovet, J. Pruzin, D. Bandy, A.S. Fleisher, J.B. Langbaum, M.J. Huentelman,
K. Jensen, T. Dunckley, R.J. Caselli, S. Kaib, E.M. Reiman, Association between GAB2
haplotype and higher glucose metabolism in Alzheimer's disease-affected brain
regions in cognitively normal APOEepsilon4 carriers, Neuroimage 54 (2010)
 A. Loessner, A. Alavi, K.U. Lewandrowski, D. Mozley, E. Souder, R.E. Gur, Regional
cerebral function determined by FDG-PET in healthy volunteers: normal patterns
and changes with age, J. Nucl. Med. 36 (1995) 1141–1149.
 C. Messa, D. Perani, G. Lucignani, A. Zenorini, F. Zito, G. Rizzo, F. Grassi, A. Del Sole,
M. Franceschi, M.C. Gilardi, et al., High-resolution technetium-99m-HMPAO
SPECT in patients with probable Alzheimer's disease: comparison with fluorine-
18-FDG PET, J. Nucl. Med. 35 (1994) 210–216.
 K. Kantarci, M.L. Senjem, V.J. Lowe, H.J. Wiste, S.D. Weigand, B.J. Kemp, A.R. Frank,
M.M. Shiung, B.F. Boeve, D.S. Knopman, R.C. Petersen, C.R. Jack Jr., Effects of age on
the glucose metabolic changes in mild cognitive impairment, AJNR Am. J. Neuror-
adiol. 31 (2010) 1247–1253.
 K.B. Walhovd, A.M. Fjell, J. Brewer, L.K. McEvoy, C. Fennema-Notestine, D.J. Hagler
Jr., R.G. Jennings, D. Karow, A.M. Dale, Combining MR imaging, positron-emission
tomography, and CSF biomarkers in the diagnosis and prognosis of Alzheimer
disease, AJNR Am. J. Neuroradiol. 31 (2011) 347–354.
 K.B. Walhovd, A.M. Fjell, A.M. Dale, L.K. McEvoy, J. Brewer, D.S. Karow, D.P. Salm-
on, C. Fennema-Notestine, Multi-modal imaging predicts memory performance
in normal aging and cognitive decline, Neurobiol. Aging 31 (2011) 1107–1121.
 S.M. Landau, D. Harvey, C.M. Madison, R.A. Koeppe, E.M. Reiman, N.L. Foster, M.W.
Weiner, W.J. Jagust, Associations between cognitive, functional, and FDG-PET
measures of decline in AD and MCI, Neurobiol. Aging (2009).
 P. Borghammer, K.Y. Jonsdottir, P. Cumming, K. Ostergaard, K. Vang, M. Ashkanian,
M. Vafaee, P. Iversen, A. Gjedde, Normalization in PET group comparison studies—
the importance of a valid reference region, Neuroimage 40 (2008) 529–540.
 P. Borghammer, P. Cumming, J. Aanerud, A. Gjedde, Artefactual subcortical hyper-
perfusion in PET studies normalized to global mean: lessons from Parkinson's dis-
ease, Neuroimage 45 (2009) 249–257.
 S.G. Mueller, M.W. Weiner, L.J. Thal, R.C. Peterson, J. Clifford, W.J. Jagust, J.Q.
Trojanowski, A.W. Toga, L. Becket, Alzheimer's Disease Neuroimaging Initiative,
in: A. Fisher, M. Memo, F. Stocchi, I. Hanin (Eds.), Advances in Alzheimer's and
Parkinson's Disease: Insight Progress and Perspectives, vol. 57, Springer, 2008,
 M.F. Folstein, S.E. Folstein, P.R. McHugh, “Mini-mental state”. A practical method
for grading the cognitive state of patients for the clinician, J. Psychiatr. Res. 12
 L. Berg, Clinical dementia rating (CDR), Psychopharmacol. Bull. 24 (1988)
 W.G. Rosen, R.C. Mohs, K.L. Davis, A new rating scale for Alzheimer's disease, Am.
J. Psychiatry 141 (1984) 1356–1364.
 G. McKhann, D. Drachman, M. Folstein, R. Katzman, D. Price, E.M. Stadlan, Clinical
diagnosis of Alzheimer's disease: report of the NINCDS-ADRDA work group
under the auspices of Department of Health and Human Services Task Force on
Alzheimer's Disease, Neurology 34 (1984) 939–944.
 F. Kruggel, D. Yves von Cramon, Alignment of magnetic-resonance brain data-
sets with the stereotactical coordinate system, Med. Image Anal. 3 (1999)
 S. Hentschel, F. Kruggel, Determination of the intracranial volume: a registration
approach, in: T. Jiang (Ed.), International Workshop on Medical Imaging and
Augmented Reality (MIAR 2004), Lecture Notes in Computer Science, vol. 3150,
Springer, Berlin, 2004, pp. 253–260.
 G. Wollny, F. Kruggel, Computational cost of nonrigid registration algorithms
based on fluid dynamics, IEEE Trans. Med. Imaging 21 (2002) 946–952.
 N. Tzourio-Mazoyer, B. Landeau, D. Papathanassiou, F. Crivello, O. Etard, N. Delcroix,
B.Mazoyer, M. Joliot, Automated anatomical labeling of activations in SPM using
a macroscopic anatomical parcellation of the MNI MRI single-subject brain,
Neuroimage 15 (2002) 273–289.
 M. Bentourkia, A. Bol, A. Ivanoiu, D. Labar, M. Sibomana, A. Coppens, C. Michel,
G. Cosnard, A. De Volder, Comparison of regional cerebral blood flow and glu-
cose metabolism in the normal brain: effect of aging, J. Neurol. Sci. 181 (2000)
 L.M. Parkes, W. Rashid, D.T. Chard, P.S. Tofts, Normal cerebral perfusion measure-
ments using arterial spin labeling: reproducibility, stability, and age and gender
effects, Magn. Reson. Med. 51 (2004) 736–743.
 Y. Chen, D.J. Wang, J.A. Detre, Test–retest reliability of arterial spin labeling with
common labeling strategies, J. Magn. Reson. Imaging 33 (2011) 940–949.
J.M. Rasmussen et al. / Biochimica et Biophysica Acta 456 (2012) 457–466