REGIONAL ANALYSIS OF FDG-PET FOR USE IN THE CLASSIFICATION OF
ALZHEIMER’S DISEASE
Katherine R. Graya , Robin Wolza , Shiva Keihaninejadb,c , Rolf A. Heckemannc,d ,
Paul Aljabara , Alexander Hammersc,d , Daniel Rueckerta
Department of Computing, Imperial College London, London, UK
a
b
Dementia Research Centre, UCL Institute of Neurology, London, UK
c
MRC Clinical Sciences Centre, Imperial College London, London, UK
d
Neurodis Foundation, CERMEP-Imagerie du Vivant, Lyon, France
ABSTRACT
We present the first use of multi-region FDG-PET data for clas-
sification of subjects from the Alzheimer’s Disease Neuroimaging
Initiative. Image data were obtained from 69 healthy controls, 71
AD patients, and 147 patients with a baseline diagnosis of MCI.
Anatomical segmentations were automatically generated in the na-
tive MRI-space of each subject, and the mean signal intensity per
cubic millimetre in each region was extracted from the FDG-PET
images. Using a support vector machine classifier, we achieve ex-
cellent discrimination between AD patients and HC (accuracy 82%),
and good discrimination between MCI patients and HC (accuracy
70%). Using FDG-PET, a technique which is often used clinically in
the workup of dementia patients, we achieve results which are com-
parable with those obtained using data from research-quality MRI,
or biomarkers obtained invasively from the cerebrospinal fluid.
Index Terms— Alzheimer’s disease; Mild cognitive impair-
ment; Positron emission tomography; Classification; Support vector
machines
1. INTRODUCTION
Alzheimer’s disease (AD) is the most common cause of dementia in
the elderly and, although there is currently no cure, research is on-
going into the development of new treatments. Since these would
be of greatest benefit to pre-symptomatic patients, their early iden-
tification is important, and patients with mild cognitive impairment
(MCI), who are at increased risk of developing AD, are of interest.
Numerous FDG-PET studies (such as [1]) have shown that
both MCI and AD are associated with significant reductions in the
cerebral metabolic rate of glucose (CMRgl) in brain regions pref-
erentially affected by AD. Reductions in AD patients can predict
both their cognitive decline and histopathological diagnosis [2],
and those in MCI patients can predict their conversion to AD [3].
FDG-PET, in conjunction with other neuroimaging methods and
clinical biomarker measurements, may therefore be a useful tool for
the early diagnosis of AD, and for monitoring its progression.
We investigate the use of a support vector machine (SVM) to
classify subjects, based on their baseline FDG-PET scans, as healthy
controls (HC); AD patients; or as having a baseline diagnosis of
MCI, which may either convert to AD (progressive MCI; p-MCI), or
remain stable (s-MCI). Machine learning techniques based on FDG-
PET data have been successful in discriminating AD patients from
HC (for example [4]), but there are many more classification stud-
ies based on structural MRI data. Various MRI-based methods, in-
978-1-4244-4128-0/11/$25.00 ©2011 IEEE 1082
cluding SVM classifiers, have been applied for the discrimination
of AD patients from HC, with reported accuracies generally ranging
between 80% and 95% (for example [4, 5, 6]).
Many classification methods rely on voxel-wise comparisons, in
which each voxel is considered as a feature that may be exploited
for group discrimination. As well as requiring that all images be
transformed into a common space, these methods can fail to ac-
count for the spatial structure resulting from correlations between
neighbouring voxels. We propose a region-based analysis which
uses anatomical segmentations that are automatically generated in
the native space of each subject, rather than in the space of a single
reference image.
When performing comparisons of FDG-PET images across sub-
jects, normalization is required to account for inter-subject variabil-
ity in overall radioactivity. Normalization is often performed relative
to the cerebral global mean (CGM) but, due to the nature of the dis-
ease, both MCI and AD patients have a lower CMRgl than HC across
the whole brain. CGM normalization therefore results in under-
estimation of the relative hypometabolism in patients compared to
HC [7], as well as the observation of apparent hypermetabolism in
relatively preserved regions of the brain [8]. Recent work suggests
that improved group discrimination can be achieved by using these
relatively preserved regions for normalization [9], and our analysis
makes use of this “reference cluster” method.
We present the first use of multi-region FDG-PET data for clas-
sification of subjects from the Alzheimer’s Disease Neuroimaging
Initiative (ADNI). Segmentations into 83 anatomical regions were
generated in the native MRI-space of each subject, and multiplied
by a grey matter mask as appropriate. The FDG-PET signal inten-
sity in each region was sampled, and normalized relative to the signal
intensity in a reference cluster derived from an independent dataset.
2. METHODS
2.1. Image Data
Data used in this work were obtained from the ADNI database1 . Par-
ticipants were recruited from over 50 sites across the United States
and Canada, to include 229 HC, 380 MCI patients, and 210 AD pa-
tients. Subjects undergo regular cognitive assessments, as well as
neuroimaging with MRI, FDG-PET and PIB-PET.
We use FDG-PET scans from 287 ADNI subjects, whose group-
wise characteristics are provided in Table 1. The MCI subjects were
1 www.loni.ucla.edu/ADNI
ISBI 2011
 Table 1. Group characteristics for the subjects whose images are used in this study.
Age Gender (%)
(mean ± std. dev.) Male Female
HC (n = 69) 75.6 ± 5.0 61
s-MCI (n = 85) 76.0 ± 6.9 73
p-MCI (n = 62) 75.2 ± 6.9 65
AD (n = 71) 76.2 ± 7.0 59
divided into p-MCI and s-MCI based on changes in status occurring
over 24 ± 11 (range 0 – 36) months. Although 404 baseline FDG-
PET scans are available in total, 89 were excluded from our analysis
due to quantification issues with the Siemens HRRT and BioGraph
HiRez scanners. Also excluded were 16 subjects whose scans could
not be processed as required, and 12 whose diagnoses did not clearly
fall into one of the four clinical categories.
The 287 baseline FDG-PET images were downloaded from the
LONI Image Data Archive in their original forms, along with pre-
processed versions of the corresponding 1.5T MRI. Full details of
the acquisition protocols may be found in the ADNI technical pro-
cedures manuals.
2.2. Image Processing
The independently derived reference cluster required for FDG-PET
image normalization was provided in MNI-space, and transformed
into the native MRI-space of each subject. The anatomical seg-
mentations required for regional sampling were generated in native
MRI-space, and the reference-cluster-scaled FDG-PET images, co-
registered with the corresponding native space MRI, were sampled
in this higher resolution space. An overview of the image processing
pipeline is provided in Fig. 1.
Fig. 1. Processing steps required to obtain the images needed for
regional sampling. Horizontal arrows indicate image registration and
re-slicing steps.
2.2.1. FDG-PET Image Pre-Processing
Since FDG-PET acquisition was performed according to one of three
standard protocols (30-minute static, 30-minute dynamic, 60-minute
dynamic), standardization was necessary before the images could be
compared. The 30-minute dynamic scans were corrected for patient
motion by using tools from the Image Registration Toolkit (IRTK)2
to rigidly register an image’s first frame with each of the five subse-
quent frames. The resulting co-registered frames were averaged to
2 www.doc.ic.ac.uk/∼dr/software
1083
MMSE Score CDR (%)
(mean ± std. dev.) 0 0.5 1
39 29.0 ± 1.1 100 0 0
27 27.5 ± 1.7 0 100 0
35 26.8 ± 1.7 0 100 0
41 23.3 ± 2.2 0 38 62
produce a single 30-minute static image. For the 60-minute quanti-
tative scans, the final six 5-minute frames were extracted, and con-
catenated into a static image in the same way.
2.2.2. Co-Registration of FDG-PET with MRI
Each of the pre-processed FDG-PET images was co-registered with
its corresponding native space MRI, again using tools from IRTK.
Rigid registration parameters were estimated, and used as a start-
ing point for the estimation of a twelve-parameter affine registration.
These affine registration parameters were then applied to transform
the FDG-PET image into the space of its corresponding MRI using
a linear interpolation. An affine transformaton was preferred over a
rigid one, because it can account for any scaling or voxel size errors
which may remain after phantom correction of the MRI [10].
2.2.3. MRI Anatomical Segmentation
Automatic segmentations of the MRI were generated in native
MRI-space using multi-atlas propagation with enhanced registration
(MAPER), which has been previously described and validated for
use in subjects with AD [11]. Atlas data required for MAPER con-
sisted of 30 manually segmented T1-weighted MRI. To match the
requirements of the MAPER procedure, additional pre-processing
was applied for brain extraction and tissue classification.
For brain extraction, binary masks covering both intracranial
white matter and grey matter (WM+GM) were available as the
starting point. These had been generated using MIDAS, a semi-
automatic procedure that is described elsewhere [12]. Each mask
was extended to cover the intracranial region generously, using
dilation and hole-filling, and FSL FAST3 was then applied to iden-
tify CSF within this pre-masked region. The original WM+GM
mask was extended by the resulting CSF mask to obtain a complete
intracranial mask.
Individual tissue probability maps for CSF, GM and WM were
obtained using FSL FAST. For FDG-PET image analysis, the GM
portion within each cortical label is of relevance. Masked segmenta-
tions are therefore employed, in which all regions except ventricles,
central structures, cerebellum and brainstem have been masked with
a GM label, and the lateral ventricles have been masked with a CSF
label.
2.2.4. FDG-PET Image Normalization
We obtained a MNI-space image of the reference cluster used in
[9] from the author. Using the “Segment” module of SPM54 , each
MRI was linearly and non-linearly deformed [13] to the MNI tem-
plate. The inverse transformation parameters generated by this pro-
cess were then used to transform the MNI-space cluster into the na-
3 www.fmrib.ox.ac.uk/fsl
4 www.fil.ion.ucl.ac.uk/spm
tive MRI-space of each subject. The cluster was also re-sampled to
the higher resolution of the MRI.

2.3. Image Analysis

Each of the native MRI-space FDG-PET images was overlaid with
its corresponding masked anatomical segmentation, and the signal
intensity per mm3 was determined for all 83 regions. Global vari-
ations in the CMRgl between subjects were accounted for by nor-
malization to the signal intensity per mm3 in the reference cluster of
each subject.

2.4. Classification
We applied a SVM classifier using LIBSVM5 . Robust estimates of
classifier performance were obtained via a bootstrapping approach,
assessing the classification rates between pairs of clinical groups.
The mean accuracy, sensitivity and specificity were evaluated over
1000 runs, where 75% of the subjects were randomly selected for
training, and the remaining 25% for testing. Receiver operating char-
acteristic (ROC) curves were also plotted, and the area under the
curve (AUC) evaluated, as this provides an overall measure of the
discriminative ability of a classifier.
A two-class SVM aims to construct a hyperplane that maximises
the margin, which is the distance between the closest points on either
side of the boundary. Since it is unlikely that real-world data would
be linearly separable, we applied a soft-margin formulation of the
SVM, in which the trade-off between maximising the margin and
minimising the training error is controlled by the penalty parameter
C. This is the “C-SVC” formulation from LIBSVM. We selected a
radial basis function kernel, whose width is defined by γ, to allow
the data to be mapped into a higher dimensional space. As part of the
training process, it was necessary to optimise the parameters C and
γ, and this was achieved by performing a grid-search using five-fold
cross-validation.
3. RESULTS
Classification results are presented in Table 2, with corresponding
ROC curves displayed in Fig. 2.
Table 2. Classification accuracy, sensitivity, and specificity, ex-
pressed as mean (standard error) over the 1000 runs.
Acc. (%) Sens. (%) Spec. (%)
AD/HC 81.6 (0.2) 82.7 (0.3) 80.4 (0.3)
MCI/HC 70.2 (0.2) 73.8 (0.3) 62.3 (0.6)
AD/MCI 68.2 (0.2) 58.3 (0.6) 73.0 (0.3)
p-MCI/s-MCI 56.4 (0.2) 51.6 (0.6) 60.1 (0.4)
Regional t-values for comparisons between AD patients and HC,
as well as MCI patients and HC are shown in Fig. 3. In the com-
parison between AD patients and HC, the ten highest t-values were
observed bilaterally in the posterior cingulate gyrus, hippocampus,
posterior temporal lobe and parietal lobe, and in the left parahip-
pocampal gyrus and middle and inferior temporal gyri. In the com-
parison between MCI patients and HC, the ten highest t-values were
observed bilaterally in the hippocampus and parietal lobe, and in the
left parahippocampal gyrus, amygdala, posterior temporal lobe, pos-
terior cingulate gyrus, insula, and pre-subgenual frontal cortex.
5 www.csie.ntu.edu.tw/∼cjlin/libsvm
Fig. 2. ROC curves, with the AUC for each group comparison given
in brackets.
4. DISCUSSION
We investigate the classification of AD and MCI subjects using a
SVM classifier based on data obtained from a regional analysis of
baseline FDG-PET scans. We achieve excellent discrimination of
AD patients from HC (accuracy 82%). This result is comparable
with other published studies, and approaches the limit achievable
in the best clinical centres, since commonly used diagnostic criteria
themselves have an accuracy of around 90%. Reports that some of
the ADNI AD patients may have a pattern of glucose metabolism
that is more consistent with frontotemporal dementia than with
AD [14] indicate a potential confounding factor which could also be
limiting the classification accuracy.
The majority of research surrounding the classification of AD
and MCI focuses on the discrimination of AD patients from HC and,
in fewer cases, MCI patients from HC. We propose the first use of
multi-region FDG-PET data for classification on the ADNI cohort,
and further aim to discriminate between p-MCI and s-MCI patients.
Good group discrimination is achieved between MCI patients and
HC (accuracy 70%), as well as between AD and MCI patients (ac-
curacy 68%), with the results again comparable with other published
studies (for example [6]). The discrimination between p-MCI and s-
MCI patients is relatively poor (accuracy 56%) but it is important
to consider that clinical follow-up data are still being acquired for
the ADNI subjects, and therefore those currently in the s-MCI group
may yet convert to AD. A recent comparison of ten high-dimensional
classification methods [15] applied to the ADNI MRI data found no
classifier to be significantly better than chance in distinguishing p-
MCI and s-MCI patients. The potential to improve on these results,
perhaps by combining information from FDG-PET and MRI, would
therefore be an interesting avenue to pursue in the future.
In comparisons between AD patients and HC, as well as between
MCI patients and HC, regional t-values indicate significant group
differences across most of the brain. The most significant regions
include those known to be affected in AD, consistent with previous
voxel-wise t-tests performed on the ADNI data [1]. The incorpo-
ration of feature selection into the classification process, such that
only the most statistically significant regions are used in discrimi-
nating between groups, could be another interesting area for further
study, since this could improve the discriminative power by exclud-
ing redundant features.

1084
Fig. 3. Regional t-values superimposed onto a maximum probability brain atlas for comparisons between (left) AD patients (n = 71) and HC
(n = 69), and (right) MCI patients (n = 147) and HC (n = 69). R: right, L: left, A: anterior, P: posterior, S: superior, I: inferior.
The reference cluster normalization used was proposed as a
data-driven method, with the cluster derived directly from the image
data. However, we have chosen to use an independently derived
cluster for normalization to avoid introducing bias into the clas-
sification process. It was important to first assess the validity of
this approach, by determining whether the regions identified as rel-
atively preserved in [9] are also valid for the ADNI dataset. We
therefore derived a reference cluster using the ADNI FDG-PET im-
ages (according to the methods described in [9]), and calculated the
intraclass correlation coefficient (ICC) between the values obtained
by sampling this ADNI derived cluster and those obtained by sam-
pling the independently derived cluster. The resulting ICC of 0.95
suggests that the area of the brain identified is reliably preserved
across early AD and MCI, and thus is likely to provide a robust
and portable reference region for image normalization. In addition,
using the ADNI derived cluster for normalization was found to have
no appreciable effect on the observed classification rates.
Using FDG-PET, a technique which is often used clinically
in the workup of dementia patients, we achieve classification re-
sults which are comparable with those obtained using data from
research-quality MRI, or biomarkers obtained invasively from the
cerebrospinal fluid.
5. ACKNOWLEDEGEMENTS
Dr. Igor Yakushev (University of Mainz, Germany) kindly provided
a MNI-space image of the reference cluster from [9]. Katherine Gray
received a studentship from the Engineering and Physical Sciences
Research Council. Rolf Heckemann was supported by a research
grant from the Dunhill Medical Trust, UK.
6. REFERENCES
[1] JB Langbaum, K Chen, and W Lee et al., “Categorical and
correlational analyses of baseline fluorodeoxyglucose positron
emission tomography images from the Alzheimer’s Disease
Neuroimaging Initiative (ADNI),” NeuroImage, vol. 45, no.
4, pp. 1107–1116, 2009.
[2] DHS Silverman, GW Small, and CY Chang et al., “Positron
emission tomography in evaluation of dementia - regional brain
metabolism and long-term outcome,” JAMA, vol. 286, no. 17,
pp. 2120–2127, 2001.
[3] D Anchisi, B Borroni, and M Franceschi et al., “Heterogeneity
of brain glucose metabolism in mild cognitive impairment and
clinical progression to Alzheimer disease,” Arch Neurol, vol.
62, no. 11, pp. 1728–1733, 2005.
[4] C Hinrichs, V Singh, and L Mukherjee et al., “Spatially aug-
mented LPboosting for AD classification with evaluations on
1085
the ADNI dataset,” NeuroImage, vol. 48, no. 1, pp. 138–149,
2009.
[5] P Vemuri, JL Gunter, and ML Senjem et al., “Alzheimer’s
disease diagnosis in individual subjects using structural MR
images: Validation studies,” NeuroImage, vol. 39, no. 3, pp.
1186–1197, 2008.
[6] O Colliot, G Chételat, and M Chupin et al., “Discrimina-
tion between Alzheimer’s disease, mild cognitive impairment
and normal aging by using automated segmentation of the hip-
pocampus,” Radiology, vol. 248, no. 1, pp. 194–201, 2008.
[7] I Yakushev, C Landvogt, and HG Buchholz et al., “Choice of
reference area in studies of Alzheimer’s disease using positron
emission tomography with fluorodeoxyglucose-F18,” Psychiat
Res-Neuroim, vol. 164, no. 2, pp. 143–153, 2008.
[8] K Herholz, E Salmon, and D Perani et al., “Discrimination
between Alzheimer dementia and controls by automated anal-
ysis of multicenter FDG PET,” NeuroImage, vol. 17, no. 1, pp.
302–316, 2002.
[9] I Yakushev, A Hammers, and A Fellgiebel et al., “SPM-based
count normalization provides excellent discrimination of mild
Alzheimer’s disease and amnestic mild cognitive impairment
from healthy aging,” NeuroImage, vol. 44, no. 1, pp. 43–50,
2009.
[10] MJ Clarkson, S Ourselin, and C Nielsen et al., “Comparison of
phantom and registration scaling corrections using the ADNI
cohort,” NeuroImage, vol. 47, no. 4, pp. 1506–1513, 2009.
[11] RA Heckemann, S Keihaninejad, and P Aljabar et al., “Im-
proving intersubject image registration using tissue-class infor-
mation benefits robustness and accuracy of multi-atlas based
anatomical segmentation,” NeuroImage, vol. 51, no. 1, pp.
221–227, 2010.
[12] PA Freeborough, NC Fox, and RI Kitney, “Interactive algo-
rithms for the segmentation and quantitation of 3-D MRI brain
scans,” Comput Methods Programs Biomed, vol. 53, no. 1, pp.
15–25, 1997.
[13] J Ashburner and KJ Friston, “Unified segmentation,” Neu-
roImage, vol. 26, no. 3, pp. 839–851, 2005.
[14] F. Thiele, F. Wenzel, and S. Young et al., “Metabolic hetero-
geneity in ADNI PET subjects,” JNM Meetings Abstracts, vol.
50, pp. 429, 2009.
[15] R. Cuingnet, E. Gérardin, and J. Tessieras et al., “Automatic
classification of patients with Alzheimer’s disease from struc-
tural MRI: A comparison of ten methods using the ADNI
database,” NeuroImage, vol. Epub ahead of print - June 11,
2010.