Article

Automatic analysis of medial temporal lobe atrophy from structural MRIs for the early assessment of Alzheimer disease

Dipartimento di Fisica, Università di Genova, 1-16146, Genova, Italy.
Medical Physics (Impact Factor: 2.64). 08/2009; 36(8):3737-47. DOI: 10.1118/1.3171686
Source: PubMed

ABSTRACT

The purpose of this study is to develop a software for the extraction of the hippocampus and surrounding medial temporal lobe (MTL) regions from T1-weighted magnetic resonance (MR) images with no interactive input from the user, to introduce a novel statistical indicator, computed on the intensities in the automatically extracted MTL regions, which measures atrophy, and to evaluate the accuracy of the newly developed intensity-based measure of MTL atrophy to (a) distinguish between patients with Alzheimer disease (AD), patients with amnestic mild cognitive impairment (aMCI), and elderly controls by using established criteria for patients with AD and aMCI as the reference standard and (b) infer about the clinical outcome of aMCI patients. For the development of the software, the study included 61 patients with mild AD (17 men, 44 women; mean age +/- standard deviation (SD), 75.8 years +/- 7.8; Mini Mental State Examination (MMSE) score, 24.1 +/- 3.1), 42 patients with aMCI (11 men, 31 women; mean age +/- SD, 75.2 years +/- 4.9; MMSE score, 27.9 +/- 1.9), and 30 elderly healthy controls (10 men, 20 women; mean age +/- SD, 74.7 years +/- 5.2; MMSE score, 29.1 +/- 0.8). For the evaluation of the statistical indicator, 150 patients with mild AD (62 men, 88 women; mean age +/- SD, 76.3 years +/- 5.8; MMSE score, 23.2 +/- 4.1), 247 patients with aMCI (143 men, 104 women; mean age +/- SD, 75.3 years +/- 6.7; MMSE score, 27.0 +/- 1.8), and 135 elderly healthy controls (61 men, 74 women; mean age +/- SD, 76.4 years +/- 6.1). Fifty aMCI patients were evaluated every 6 months over a 3 year period to assess conversion to AD. For each participant, two subimages of the MTL regions were automatically extracted from T1-weighted MR images with high spatial resolution. An intensity-based MTL atrophy measure was found to separate control, MCI, and AD cohorts. Group differences were assessed by using two-sample t test. Individual classification was analyzed by using receiver operating characteristic (ROC) curves. Compared to controls, significant differences in the intensity-based MTL atrophy measure were detected in both groups of patients (AD vs controls, 0.28 +/- 0.03 vs 0.34 +/- 0.03, P < 0.001; aMCI vs controls, 0.31 +/- 0.03 vs 0.34 +/- 0.03, P < 0.001). Moreover, the subgroup of aMCI converters was significantly different from controls (0.27 +/- 0.034 vs 0.34 +/- 0.03, P < 0.001). Regarding the ROC curve for intergroup discrimination, the area under the curve was 0.863 for AD patients vs controls, 0.746 for all aMCI patients vs controls, and 0.880 for aMCI converters vs controls. With specificity set at 85%, the sensitivity was 74% for AD vs controls, 45% for aMCI vs controls, and 83% for aMCI converters vs controls. The automated analysis of MTL atrophy in the segmented volume is applied to the early assessment of AD, leading to the discrimination of aMCI converters with an average 3 year follow-up. This procedure can provide additional useful information in the early diagnosis of AD.

Full-text

Available from: Roberto Bellotti
Automatic analysis of medial temporal lobe atrophy from structural MRIs
for the early assessment of Alzheimer disease
Piero Calvini
Dipartimento di Fisica, Università di Genova, I-16146, Genova, Italy and Istituto Nazionale di Fisica
Nucleare, Sezione di Genova, I-16146, Genova, Italy
Andrea Chincarini and Gianluca Gemme
a
Istituto Nazionale di Fisica Nucleare, Sezione di Genova, I-16146, Genova, Italy
Maria Antonietta Penco and Sandro Squarcia
Dipartimento di Fisica, Università di Genova, I-16146, Genova, Italy and Istituto Nazionale di Fisica
Nucleare, Sezione di Genova, I-16146, Genova, Italy
Flavio Nobili and Guido Rodriguez
Neurofisiologia Clinica, Dipartimento di Neuroscienze, Oftalmologia e Genetica, Azienda
Ospedale-Università S. Martino, Genova, I-16132, Genova, Italy
Roberto Bellotti
Dipartimento Interateneo di Fisica “M. Merlin” and TIRES, Università degli Studi di Bari, I-70126, Bari,
Italy and Istituto Nazionale di Fisica Nucleare, Sezione di Bari, I-70126, Bari, Italy
Ezio Catanzariti
Dipartimento di Scienze Fisiche, Università di Napoli, I-80126, Napoli, Italy and Istituto Nazionale
di Fisica Nucleare, Sezione di Napoli, I-80126, Napoli, Italy
Piergiorgio Cerello
Istituto Nazionale di Fisica Nucleare, Sezione di Torino, I-10125, Torino, Italy
Ivan De Mitri
Dipartimento di Fisica, Università del Salento, I-73100, Lecce, Italy and Istituto Nazionale di Fisica
Nucleare, Sezione di Lecce, I-73100, Lecce, Italy
Maria Evelina Fantacci
Dipartimento di Fisica, Università di Pisa, I-56127, Pisa, Italy and Istituto Nazionale di Fisica Nucleare,
Sezione di Pisa, I-56127, Pisa, Italy
The MAGIC-5 Collaboration
b
and
The Alzheimer’s Disease Neuroimaging Initiative ADNI
c
Received 24 December 2008; revised 22 May 2009; accepted for publication 17 June 2009;
published 13 July 2009
The purpose of this study is to develop a software for the extraction of the hippocampus and
surrounding medial temporal lobe MTL regions from T1-weighted magnetic resonance MR
images with no interactive input from the user, to introduce a novel statistical indicator, computed
on the intensities in the automatically extracted MTL regions, which measures atrophy, and to
evaluate the accuracy of the newly developed intensity-based measure of MTL atrophy to a
distinguish between patients with Alzheimer disease AD, patients with amnestic mild cognitive
impairment aMCI, and elderly controls by using established criteria for patients with AD and
aMCI as the reference standard and b infer about the clinical outcome of aMCI patients. For the
development of the software, the study included 61 patients with mild AD 17 men, 44 women;
mean age standard deviation SD, 75.8 years 7.8; Mini Mental State Examination MMSE
score, 24.1 3.1, 42 patients with aMCI 11 men, 31 women; mean age SD, 75.2 years 4.9;
MMSE score, 27.9 1.9, and 30 elderly healthy controls 10 men, 20 women; mean age SD,
74.7 years 5.2; MMSE score, 29.1 0.8. For the evaluation of the statistical indicator, 150
patients with mild AD 62 men, 88 women; mean age SD, 76.3 years 5.8; MMSE score,
23.2 4.1, 247 patients with aMCI 143 men, 104 women; mean age SD, 75.3 years 6.7;
MMSE score, 27.0 1.8, and 135 elderly healthy controls 61 men, 74 women; mean age
SD,
76.4 years 6.1. Fifty aMCI patients were evaluated every 6 months over a 3 year period to assess
conversion to AD. For each participant, two subimages of the MTL regions were automatically
extracted from T1-weighted MR images with high spatial resolution. An intensity-based MTL
atrophy measure was found to separate control, MCI, and AD cohorts. Group differences were
3737 3737Med. Phys. 36 8, August 2009 0094-2405/2009/368/3737/11/$25.00 © 2009 Am. Assoc. Phys. Med.
Page 1
assessed by using two-sample t test. Individual classification was analyzed by using receiver oper-
ating characteristic ROC curves. Compared to controls, significant differences in the intensity-
based MTL atrophy measure were detected in both groups of patients AD vs controls, 0.28 0.03
vs 0.34 0.03, P 0.001; aMCI vs controls, 0.31 0.03 vs 0.34 0.03, P 0.001. Moreover, the
subgroup of aMCI converters was significantly different from controls 0.27 0.034 vs 0.34 0.03,
P 0.001. Regarding the ROC curve for intergroup discrimination, the area under the curve was
0.863 for AD patients vs controls, 0.746 for all aMCI patients vs controls, and 0.880 for aMCI
converters vs controls. With specificity set at 85%, the sensitivity was 74% for AD vs controls, 45%
for aMCI vs controls, and 83% for aMCI converters vs controls. The automated analysis of MTL
atrophy in the segmented volume is applied to the early assessment of AD, leading to the discrimi-
nation of aMCI converters with an average 3 year follow-up. This procedure can provide additional
useful information in the early diagnosis of AD. © 2009 American Association of Physicists in
Medicine. DOI: 10.1118/1.3171686
Key words: magnetic resonance imaging, image analysis, Alzheimer disease, hippocampus
I. INTRODUCTION
Early and accurate diagnosis of Alzheimer disease AD is
challenging. In recent years, the early clinical signs of AD
have been extensively investigated, leading to the concept of
amnestic mild cognitive impairment aMCI, an intermediate
cognitive state between normal aging and dementia.
15
aMCI
patients experience memory loss to a greater extent than ex-
pected for their age, and they progress to a diagnosis of AD
at a faster rate than controls.
1,2
A challenge for modern neuroimaging is to help in the
diagnosis of early AD, particularly in aMCI patients. Three-
dimensional 3D magnetic resonance MR imaging with
high spatial resolution allows visualization of subtle ana-
tomic changes and thus can help detect brain atrophy in the
initial stages of the disease.
6
For this reason, sensitive neu-
roimaging measures have been investigated to quantify brain
structural changes in early AD which are automated enough
to permit large-scale studies of the disease and the factors
that affect it. To date, these studies have particularly focused
on assessing atrophy of medial temporal lobe MTL struc-
tures, including the hippocampus for a review of this topic
see Refs. 711.
Methods to assess hippocampal atrophy have largely been
based on volumetric measurements,
9,12,13
on mapping the
spatial distribution of atrophy in 3D scans,
1418
or on visual
rating scales.
19
Volumetric measurements typically rely on
manual outlining of the MTL structures on serial MR im-
ages, which is time consuming and prone to inter-rater and
intrarater variabilities. Thus, large-scale studies of AD-
related hippocampal atrophy are often impractical.
20
Visual
rating scales, although simple to use and suitable for a clini-
cal application, were not designed to detect atrophy progres-
sion on serial imaging; their quantized nature makes them
insensitive to change over time.
21
To accelerate and spread epidemiological studies and
clinical trials, some automated systems have been proposed
for hippocampal segmentation,
2228
but none is yet widely
used
29
due to the high computational burden,
30
unsatisfactory
results,
31
or poor generalization capability.
24
In order to overcome these difficulties, in this work we
describe the development of a simple, quick, and operator-
independent method to extract two fixed size 30 70
30 mm
3
, parallelepiped-shaped subimages from a MR
image. These subimages contain both the hippocampus and
the perihippocampal region; in the following they are de-
noted as hippocampal boxes HBs.
From the automatically extracted HBs we derive a novel,
intensity-based, statistical indicator, which carefully mea-
sures the MTL atrophy and is able to distinguish between
patients with AD, patients with aMCI, and elderly controls
and also between converters and nonconverters to AD within
the aMCI population with reasonably good accuracy.
II. DEVELOPMENT OF THE SOFTWARE
II.A. Subjects
At the development stage the study included 61 patients
with mild AD 17 men, 44 women; mean age standard de-
viation SD, 75.8 years 7.8; Mini Mental State Examina-
tion MMSE score, 24.1 3.1, 42 patients with aMCI 11
men, 31 women; mean age SD, 75.2 years 4.9; MMSE
score, 27.9 1.9, and 30 elderly healthy controls 10 men,
20 women; mean age SD, 74.7 years 5.2; MMSE score,
29.1 0.8兲共see Table I.
The diagnosis of aMCI was made according to the criteria
of Petersen et al.,
2
while the diagnosis of AD was made
according to the NINCDS-ADRDA Ref. 32 and the
DSM-IV criteria.
The presence of dementia was evaluated on the basis of
clinical interview with the patient and caregiver, question-
naires for the activities of daily living ADLs,
33
instrumen-
tal ADL IADL,
34
and clinical dementia rating CDR scale
the result was 0.5 in all aMCI patients and 1.0 in all AD
patients. General cognition was assessed by means of
MMSE.
35
All patients underwent a standard battery of blood count,
blood chemical examinations, and urinalysis, according to
the commonly followed rules in order to exclude secondary
causes of cognitive impairment. The presence of analpha-
betism, major vision disturbances, psychiatric illnesses, epi-
lepsy, major head trauma, Parkinsonism, previous stroke or
3738 Calvini et al.: Automatic analysis of MTL for AD early assessment 3738
Medical Physics, Vol. 36, No. 8, August 2009
Page 2
TIA, and brain masses was another exclusion criterion. A
mild depressive trait, as ascertained by the 15-item geriatric
depression scale GDS, was not an exclusion criterion. Neu-
ropsychiatric symptoms were assessed by interviewing the
informant with the neuropsychiatric inventory NPI.
36
Pa-
tients scoring higher than 0 on the delusion and the halluci-
nation NPI items were excluded.
MR imaging was performed in all patients by means of a
1.5 T equipment. Only patients with evidence of major stroke
were excluded, while white matter WM hyperintensities,
leukoaraiosis, and lacunae were not exclusion criteria.
The follow-up of the patients began with a clinical exami-
nation, also comprehensive of MMSE, ADL, and IADL
questionnaires and CDR, and this was repeated every 6
months. A follow-up time of at least 1 year was available for
all patients.
The 30 control subjects were recruited among a group of
healthy volunteers attending university courses dedicated to
elderly people. They all gave their informed consent. Their
healthy condition was carefully checked by means of general
medical history and clinical examination, and the same ex-
clusion criteria as for patients were applied, with the excep-
tion of cognitive complaints. MMSE was performed, and
only subjects with a normal score i.e., 26 were consid-
ered for this study. Moreover, only subjects with a CDR of 0
were included. The enrolled controls underwent brain MR
imaging and the same neuropsychological battery as the pa-
tients. The protocol received the approval of the local ethics
committee.
II.B. Image acquisition
T1-TFE volumetric MR imaging was performed in all pa-
tients using 1.5 T equipment Gyroscan Intera, Philips Medi-
cal Systems, Best, The Netherlands to acquire a sagittal se-
quence with the following parameters: TRs of 8.7 ms, TEs of
4.1 ms, flip of 8, FOVs of 256 mm, matrix of 256 256, 150
sagittal slices, and voxel size of 0.98 0.98 1.6 mm
3
. All
images were resampled to make them isotropic with voxel
size of 1 1 1mm
3
.
II.C. Extraction of the hippocampal boxes
The extraction method relies on the fact that the gray level
contrast displayed by the hippocampal formation and adja-
cent structures is unique all over the brain. Therefore, a pro-
cedure can be developed to identify the hippocampal region
unambiguously. Neuroanatomical considerations suggested
the size of a HB as a 30 70 30 mm
3
parallelepiped-
shaped box sizes of right-to-left, posterior-to-anterior, and
inferior-to-superior directions, respectively. The extraction
of the 266 HBs 133 right and 133 left was performed with
an automatic procedure, that required minimal interactive in-
tervention. Here we illustrate only the process for the extrac-
tion of the right HBs, the procedure for the extraction of the
left ones being the same.
First, all MR scans were labeled and denoted as MR
i
i =1, ...,133. Then all images were spatially normalized to
stereotactic space ICBM152 via a 12-degrees-of-freedom
affine transformation,
37
which normalizes the brains in terms
of dimensions, position, and spatial orientation. Conse-
quently, all hippocampi share similar positions and orienta-
tions. Three slices cut from a 3D MR image after spatial
normalization and the outline of the hippocampal ROI are
shown in Fig. 1.
After spatial normalization, the first HB was manually
extracted by an expert reader from MR
1
, the scan of a
healthy control displaying minimal atrophy. Particular care
was applied in the positioning of the box boundaries in order
TABLE I. Ensemble properties of our “development” subjects. The error is one standard deviation.
AD MCI Controls
No. of subjects 61 17 men, 44 women 42 11 men, 31 women 30 10 men, 20 women
Age years 75.87.8 75.2 4.9 74.7 5.2
MMSE score 24.1 3.1 27.9 1.9 29.1 0.8
FIG.1. a Axial, b sagittal, c and coronal views of an image after align-
ment with the ICBM152 template. On the three slices an outline of the right
hippocampal box is also shown.
3739 Calvini et al.: Automatic analysis of MTL for AD early assessment 3739
Medical Physics, Vol. 36, No. 8, August 2009
Page 3
to place the hippocampal formation in the inner portion of
the box. The next step of the procedure consists in extracting
the second HB from the remaining 132 MR images.
The extraction procedure is based on the registration of
the first HB the fixed image onto the remaining 132 images
the moving images via a six-degrees-of-freedom rigid
transformation three translational and three rotational de-
grees of freedom.
A definition of distance between two HBs drives the reg-
istration procedure in that it provides a measure of how well
the transformed moving image matches the fixed image. In
this way a quantitative criterion is assigned for finding the
optimal values of the transformation parameters. We adopted
a definition of distance based on the normalized correlation
coefficient C. Assigning the HBs A and B, each one consist-
ing of N voxels N= 63 000 in our case, the corresponding C
is given by
C
A,B
=
=1
N
A
A
¯
兲共B
B
¯
=1
N
A
A
¯
2
=1
N
B
B
¯
2
, 1
where voxel intensities A
and B
of HBs A and B, respec-
tively, are labeled by a single index Greek letter following
lexicographic ordering, and the average intensity I
¯
is given
by
I
¯
=
1
N
=1
N
I
, 2
where I= A , B. From this quantification of similarity, one de-
rives the following definition of distance between A and B:
d
A,B
=1−C
A,B
. 3
This distance is scale and shift invariant and it produces a
cost function with sharp peaks and well defined minima. It
has a relatively small capture radius,
38
but this fact does not
represent a severe limitation in our case because all images
are aligned in the same stereotactic space and thus the search
for the hippocampal formations requires the exploration of a
small parameter space.
The success of the registration of each moving image onto
the fixed image is quantified by the minimum reached in
distance values,
. With a moderate computational effort, one
could extract all the 132 remaining right HBs by using the
first manually defined HB alone, but the quality of the results
is not homogeneously good. In fact the fixed image is suc-
cessful in extracting the HBs which are not too dissimilar
from it. However, due to the ample morphological variability
contained in the population of MR images, some HBs exist
which are unsatisfactorily extracted or not found at all.
Therefore, a more complete approach is required. The
population of the remaining 132 MR images is registered
onto the fixed image, i.e., the HB extracted from MR
1
.
Thus, for each given value of index j 2 j 133 this op-
eration produces the value of the score,
1,j
, and of the six
geometrical parameters, three translations and three rotation
angles. Such values are stored in the first row of seven
strictly upper triangular matrices. We emphasize that no ac-
tual HB extraction is performed at this stage.
On the basis of the presently available score list the first
row of matrix
, the second box is extracted from MR
j
where j
is the index of the minimum not vanishing value
of
1,j
. In our application we obtain j
=13. Now the extrac-
tion of the new HB is really performed by using the corre-
sponding values of the geometrical parameters
x
1,13
, y
1,13
,z
1,13
,....
Once the second HB is available, the remaining 131 mov-
ing images are registered onto this new fixed image, and a
new set of scores and geometrical parameters are obtained
and stored in the second row of the seven matrices. The third
extracted HB is selected from MR
j
ⴱⴱ
where now j
ⴱⴱ
is given
by the index of the minimum not vanishing value of
1,j
and
2,j
. In this search for the minimum score, the entries
corresponding to the already extracted HBs must be skipped.
The procedure for the progressive extraction of all HBs
follows this scheme and the extension to an increasing num-
ber of HB examples is obvious. The procedure stops once the
whole sample of 133 images has been processed and the 133
right HBs have been obtained.
Except for the extraction of the first HB, the whole pro-
cess runs automatically, without any manual intervention,
and no appreciable drift affecting hippocampus orientation or
positioning in the HBs is noticeable during the extraction
process. Visual inspection of the whole set of 133 HBs
shows that the level of spatial registration of similar anatomi-
cal structures is very high. Such stability is not surprising if
one considers the way the whole procedure works. At the
beginning, the early extractions exhaust the set of the HBs
which are very similar to the manually defined HB. Then, the
procedure starts extracting HBs which are progressively dif-
ferent from the first ones, but diversity creeps into the grow-
ing HB database very slowly, thanks to the relevant size of
the population of the available MR images. Thus, the orien-
tation and position of the essential geometrical features of
the hippocampal formation are preserved during the whole
process of HB extraction.
The same procedure was applied to the left hippocampi. A
first example of a left HB is manually generated and, then,
the whole process just described is run on the left side of the
brain. The result consists in the generation of the matrices of
scores and of geometrical parameters and in the extraction of
the 133 left HBs.
II.D. Selection of templates
The procedure described in Sec. II C to generate the 266
HBs is rather demanding and it is unreasonable to run over it
again for extracting the two HBs of any new MR image. In
this section we show that the extraction can be successfully
performed by a smaller number of properly chosen HBs, in
the following denoted as HB templates HBTs.
The basic idea of the HBT selection process is to create
groups of HBs, or clusters, in such a way that the HBs in the
same cluster are near and the HBs belonging to different
clusters are far. In general, let n be the number of the avail-
3740 Calvini et al.: Automatic analysis of MTL for AD early assessment 3740
Medical Physics, Vol. 36, No. 8, August 2009
Page 4
able HBs n = 133 in our case. Here we denote by d
i,j
the
distance between HB
i
and HB
j
as defined in Eq. 3. All
d
i,j
are known and can be considered as the entries of a
symmetric n n matrix, with vanishing diagonal elements.
Thus, each HB can be considered as a point belonging to an
N-dimensional space N =63 000 and whose distances from
all other HBs are known.
The classification of the n HBs in homogeneous clusters
is performed by means of the k-means algorithm.
39
This al-
gorithm tries to find a partition of the whole set of data into
k clusters that are as compact and well separated as possible.
Each cluster is defined by its members and by its centroid,
i.e., the HB having a minimal sum of distances from all
members of the cluster.
For a given value of k, the algorithm needs the initial
centroid position as input. These are fixed selecting k ran-
domly chosen HBs among the whole population of n HBs
k n. Since the final result depends on the initial choice,
the algorithm is repeated ten times with different initial cen-
troid selections, and the best result is kept. The required set
of k HBTs is represented by the centroids of the optimal
partition.
With the purpose of finding the appropriate value for the
number k of HBTs, some tests were performed by increasing
the number of clusters from k =3 up to k = 10. The quality of
the eight solutions can be evaluated in a quantitative way by
means of the corresponding average silhouette values. For a
k-clusters partition, the silhouette value for point i is defined
as
S
i
=
min
k
b
i,l
a
i
maxa
i
,min
k
b
i,l
兲兲
, 4
where a
i
is the average distance of point i from the other
points in its cluster and b
i,l
is the average distance of point i
from the points belonging to cluster l. This quantity is a
measure of how similar that point is to points in its own
cluster compared to the points in other clusters. Its value
ranges from 1 to +1 and is given by
S
¯
=
1
n
i=1
n
S
i
, 5
where the sum is over the whole population of n points. The
plot of S
¯
for k =3 to k = 10 is shown in Fig. 2 and indicates
that average silhouette values begin to decrease for k 6.
The main objective of the above illustrated clustering pro-
cedure consists in selecting a small number of k HBTs as
representative of the ample morphological variability con-
tained in the whole MR data set and, finally, in finding an
appropriate value for k. Thus, the performance of such small
sets of HBTs in extracting all the 133 right HBs was evalu-
ated. The test consisted of extracting all HBs from all MR
images by means of every set of k HBTs. Each MR image
was registered to all the k HBTs and the actual HB extraction
was performed on the basis of the best score obtained
among the k available scores. The test was repeated for k
=3 10. The procedure generated eight sets of 133 HBs to be
compared to the original set of HBs extracted by means of
the general procedure described in Sec. II C. In order to
evaluate the performance of the eight HBT sets in extracting
all HBs, we calculated, according to Eq. 3, the average
distance between the newly extracted HBs and the original
ones. The result is shown in Fig. 3 and displays average
distance values less than 0.1 for all values of k=3 10. Ob-
viously such values decrease when the number k of HBTs
increases.
On the basis of the results obtained by the previous tests
and shown in Figs. 2 and 3, k=5 was chosen for the number
of HBTs giving the best trade-off between representative ca-
pability and dimension of the template dictionary. A sagittal
slice of the five right templates is shown in Fig. 4. The same
procedure and the same tests were applied to the left hippoc-
ampi see Figs. 2 and 3, and also in this case the choice k
=5 represents an excellent trade-off.
FIG. 2. Average silhouette for k =3 to k =10 for right circles and left hip-
pocampi squares.
FIG. 3. Average distance for k =3 to k =10 for right circles and left hip-
pocampi squares.
3741 Calvini et al.: Automatic analysis of MTL for AD early assessment 3741
Medical Physics, Vol. 36, No. 8, August 2009
Page 5
III. APPLICATION OF THE SOFTWARE
III.A. Subjects
In order to test the HBT extraction efficiency on images
acquired with different scanners and/or with different system
settings we processed n= 532 images, partly acquired by the
National Institute of Cancer IST on patients clinically
treated at the Clinical Neurophysiology unit in Genoa,
Italy, and partly downloaded from the Alzheimer
Disease Neuroimaging Initiative ADNI database
www.loni.ucla.edu/ADNI.
40
The ADNI was launched in
2003 by the National Institute on Aging NIA, the National
Institute of Biomedical Imaging and Bioengineering
NIBIB, the Food and Drug Administration FDA, private
pharmaceutical companies, and nonprofit organizations as a
$60 million, 5 year public-private partnership. ADNI is the
result of efforts of many coinvestigators from a broad range
of academic institutions and private corporations, and sub-
jects have been recruited from over 50 sites across the US
and Canada. The initial goal of ADNI was to recruit 800
adults, ages 55–90, to participate in the research—
approximately 200 cognitively normal older individuals to be
followed for 3 years, 400 people with MCI to be followed
for 3 years, and 200 people with early AD to be followed for
2 years.
The study included 150 patients with mild AD 62 men,
88 women; mean age SD, 76.3 years 5.8; MMSE score,
23.2 4.1, 247 patients with aMCI 143 men, 104 women;
mean age SD, 75.3 years 6.7; MMSE score, 27.0 1.8,
and 135 elderly healthy controls 61 men, 74 women; mean
age SD, 76.4 years 6.1; MMSE score, 29.0 0.8兲共see
Table II. For a subset of 50 aMCI patients clinical follow-up
data were available 3 year average follow-up time. Among
them 25 converted to AD and 5 reverted to normalcy or
developed pathologies different from AD, whereas the re-
maining 20 were still affected by memory deficit but without
dementia.
The images are 1.5 T screening scans, acquired on scan-
ners manufactured by Philips, Siemens, and GE, consisting
of a straight sagittal 3D MPRAGE sequence. The images
were not corrected for gradient nonlinearity, B1 nonunifor-
mity, and intensity nonuniformity to better represent average
scans obtained in “real world” clinical practice. Each image
was registered onto the five HBTs and the extraction was
successfully carried out as confirmed by visual inspection.
III.B. The -box score
In this section we introduce a novel statistical indicator,
which measures MTL atrophy and is computed on the inten-
sities in the automatically extracted HBs. Because the abso-
lute intensity of a T1-weighted image is meaningless on its
own, after the extraction of the HBs a preliminary equaliza-
tion of the gray levels was necessary in order to map the
mean cerebrospinal fluid CSF, gray matter GM, and WM
intensities of each HB to common references. This prepro-
cessing, which required a rough segmentation in CSF, GM,
and WM, was necessary for obtaining reliable results from
voxel based operations and comparisons among HBs.
The next step consisted in preparing the average HBs rep-
resentative for controls and for AD cohorts. In order to maxi-
mize intragroup similarities and intergroup differences, out-
liers in each group were preliminarily discarded. For each
group a distribution of distances was built after defining a
FIG. 4. Sagittal sections of the five right HBTs.
TABLE II. Ensemble properties of our “test” subjects. The error is one standard deviation.
AD MCI Controls
No. of subjects 150 62 M, 88 W 247 143 M, 104 W 135 61 M, 74 W
Age years 76.3 5.8 75.3 6.7 76.4 6.1
MMSE score 23.2 4.1 27.0 1.8 29.0 0.8
-box score 0.28 0.03 0.31 0.03 0.34 0.03
3742 Calvini et al.: Automatic analysis of MTL for AD early assessment 3742
Medical Physics, Vol. 36, No. 8, August 2009
Page 6
metric among HBs in terms of the usual scalar product. Out-
liers for each group were selected and removed from the
least significant 5% of the corresponding distance distribu-
tions just obtained. After the outlier removal, the average
HBs representative for controls and AD cohorts were pre-
pared by taking the voxel-by-voxel median for each group.
Subsequently, the voxel-by-voxel difference between these
two medians was performed and denoted as the box. The
projection of every HB onto the box, evaluated by means
of the scalar product, produces the required intensity-based
measure, named hereafter -box score. Age detrending was
applied to our variable to ensure that the analysis was not
biased by age confounding effects.
III.C. Results
Group differences were assessed by using Student’s two-
sample t test. Significant differences of MTL atrophy, mea-
sured by the -box method, were detected both in AD and in
aMCI cohorts AD vs controls, 0.28 0.03 vs 0.34 0.03,
P 0.001; aMCI vs controls, 0.31 0.03 vs 0.34 0.03, P
0.001. MTL atrophy in the subgroup of 25 aMCI convert-
ers was similar to the one from the AD group and was sig-
nificantly different from the one of the controls 0.27 0.03
vs 0.34 0.03, P 0.001.
Individual classification on the basis of the -box score
was also analyzed by using the receiver operating character-
istic ROC curves, which indicate the relationship between
sensitivity and 1-specificity for each intergroup discrimina-
tion. The area under the curve was 0.863 for AD patients vs
controls Fig. 5, 0.746 for aMCI patients vs controls Fig.
6, and 0.880 for aMCI converters vs controls Fig. 7. With
specificity set at 85%, the sensitivity was 74% for AD vs
controls, 45% for aMCI vs controls, and 83% for aMCI con-
verters vs controls.
IV. DISCUSSION
While several automated hippocampal segmentation
methods have been proposed,
25,4145
most of them rely either
FIG.5. Boxplotof-box score for controls and AD. Lines represent the median, boxes the interquartile range, and whiskers the range; stars=outliers. The
area under the ROC curve is 86.3%.
FIG. 6. Box plot of -box score for controls and aMCI. The area under the ROC curve is 74.6%.
3743 Calvini et al.: Automatic analysis of MTL for AD early assessment 3743
Medical Physics, Vol. 36, No. 8, August 2009
Page 7
on the manual identification of several hippocampal land-
marks on each scan
16,25,46,47
or on algorithms based on the
intensities and spatial anatomical relationship of different
brain structures to guide hippocampal outlining.
24,48
Webb et
al.
49
devised an automated method to detect hippocampal
atrophy in patients with temporal lobe epilepsy based on the
analysis of the image intensity differences between patients
and controls within a volume of interest centered on the hip-
pocampus. Thompson et al.
18
generated color-coded maps to
visualize the hippocampal atrophy rate using 3D parametric
mesh models of manually segmented hippocampal regions
on serial scans. In addition, an automatic measure of hippoc-
ampal atrophy rates has been derived using regional fluid
registration
23
or by calculating the regional boundary shift
integral.
22
However, both methods require manual segmenta-
tion of the baseline hippocampal region. Rusinek et al.
50
used the boundary shift integral analysis applied to a volume
of interest centered on the hippocampus to calculate the rate
of MTL atrophy. Compared to these methods our procedure
requires no prerequisites for automation.
Moreover, few of these methods have been applied to
patients with AD and/or MCI and rarely did researchers re-
port the accuracy of their techniques in the differentiation
among MCI, AD, and controls. Carmichael et al.
20
assessed
the performance of automated atlas-based segmentation by
using several freely available registration methods auto-
mated image registration University of California at Los
Angeles, Los Angeles, CA, statistical parametric mapping
Wellcome Department of Imaging Neuroscience, functional
MR imaging linear image registration tool University of
Oxford, Oxford, England, and a fully deformable approach
in AD and MCI patients. They came to the conclusion that
these approaches are less precise when applied to AD pa-
tients than controls but this should be tempered by the fact
that these techniques were not specifically designed for this
task. Fischl et al.
24
proposed a general method, derived from
a probabilistic atlas, to automatically label different noncor-
tical structures, including the hippocampus, and applied this
technique to patients with mild and questionable AD. The
method helped identify significant group differences in terms
of hippocampal volume but the authors did not investigate
the classification of individual participants. Csernansky et
al.
51
used the high-dimensional brain mapping approach, on
the basis of fluid registration with a template, to obtain hip-
pocampal volumes and hippocampal shape differences be-
tween patients with very mild AD and controls. By using a
classification on the basis of volume and shape features, they
achieved a sensitivity of 83% and a specificity of 78%. By
using a similar high-dimensional brain mapping approach,
Hsu et al.
46
compared automated and manual segmentations
in AD and cognitively impaired patients. They reported good
correlations between manual and automated measurements
of the hippocampal volume. However, they did not investi-
gate the accuracy of this technique for the classification of
individual patients. Colliot et al.
52
evaluated the accuracy of
automated hippocampal volumetry to help distinguish be-
tween patients with AD, patients with MCI, and elderly con-
trols. Individual classification on the basis of hippocampal
volume resulted in 84% correct classification sensitivity,
84%; specificity, 84% between AD patients and controls and
73% correct classification sensitivity, 75%; specificity, 70%
between MCI patients and controls. Ridha et al.
21
compared
an automated intensity-based measure of medial temporal
lobe atrophy ATLAS with volumetric and visually based
methods. Their measure differentiates patients with AD from
controls at cross-sectional and longitudinal levels. At base-
line, for a specificity of 85%, the sensitivity of hippocampal
volume measurement and visual rating scale
19
were similar
84% vs 86%, whereas the sensitivity of the ATLAS mea-
sure was lower at 73%.
In this work we introduced a novel approach, based on the
measure of a new statistical indicator, the -box score, able
to separate the AD, aMCI, and controls cohorts. Since it is
well known that MTL atrophy is associated with declining
cognitive function,
13
we showed that our method is able to
capture differences between subgroups of interest with dif-
ferent stages of cognitive impairment, with comparable dis-
criminating capability between aMCI converters and controls
FIG.7. Boxplotof-box score for controls and aMCI converters. The area under the ROC curve is 88%.
3744 Calvini et al.: Automatic analysis of MTL for AD early assessment 3744
Medical Physics, Vol. 36, No. 8, August 2009
Page 8
and between AD patients and controls. This result should be
considered with caution owing to the relatively small number
of converters. Anyway, this is in agreement with several
studies regarding manual segmentation of hippocampus,
which have reported that baseline hippocampal volume is an
indicator of future progression to AD.
13,5356
This is also in
agreement with studies based on visual rating, which clearly
found MTL atrophy in patients who subsequently converted
to AD.
5759
Compared to other methods of hippocampal or MTL at-
rophy measurement, our method, which does not directly
tackle the objective of hippocampus segmentation, is fully
automated, allowing the analysis of large sets of data, and
requires relatively moderate image postprocessing and pre-
requisites for automation. Therefore, it could be a good can-
didate for being more widely used than other automatic
methods.
In conclusion, we report a novel procedure for assessing
MTL atrophy based on intensity measurement in a standard-
ized perihippocampal volume using established T1-weighted
volumetric scans. This measure significantly differentiates
patients with AD from controls and MCI converters from
controls. The technique is simple to use and may be of value
in clinical practice for an early diagnosis of AD, without the
need for expert assessment or labor intensive manual mea-
sures.
ACKNOWLEDGMENTS
The authors thank Mr. E. Deseri for image acquisition and
Dr. C. E. Neumaier for MR image reporting. Data collection
and sharing for this project was funded by the ADNI Prin-
cipal Investigator: Michael Weiner; NIH Grant No. U01
AG024904. ADNI is funded by the National Institute on
Aging, the National Institute of Biomedical Imaging and
Bioengineering NIBIB, and through generous contributions
from the following: Pfizer Inc., Wyeth Research, Bristol-
Myers Squibb, Eli Lilly and Co., GlaxoSmithKline, Merck &
Co. Inc., AstraZeneca AB, Novartis Pharmaceuticals Corpo-
ration, Alzheimer Association, Eisai Global Clinical Devel-
opment, Elan Corporation plc, Forest Laboratories, and the
Institute for the Study of Aging, with participation from the
US Food and Drug Administration. Industry partnerships are
coordinated through the Foundation for the National Insti-
tutes of Health. The grantee organization is the Northern
California Institute for Research and Education, and the
study is coordinated by the Alzheimer 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. This work benefited
from the use of the Insight Segmentation and Registration
Toolkit
ITK, an open source software developed as an ini-
tiative of the US National Library of Medicine and available
at http://www.itk.org.
60
a
Author to whom correspondence should be addressed. Electronic mail:
gianluca.gemme@ge.infn.it
b
This work was funded by INFN within the MAGIC-5 research project
and by MIUR within Research Program No. 2005020135.
c
Data used in the preparation of this article were in part obtained from the
ADNI database www.loni.ucla.edu/ADNI. As such, the investigators
within the ADNI contributed to the design and implementation of ADNI
and/or provided data but did not participate in analysis or writing of this
report. ADNI investigators complete listing available at
www.loni.ucla.edu/ADNI/Collaboration/ADNI_Authorship_list.pdf
1
R. C. Petersen, G. E. Smith, S. C. Waring, R. J. Ivnik, E. G. Tangalos, and
E. Kokmen, “Mild cognitive impairment: Clinical characterization and
outcome,” Arch. Neurol. 563, 303–308 1999.
2
R. C. Petersen, R. Doody, A. Kurz, R. C. Mohs, J. C. Morris, P. V.
Rabins, K. Ritchie, M. Rossor, L. Thal, and B. Winblad, “Current con-
cepts in mild cognitive impairment,” Arch. Neurol. 5812, 1985–1992
2001.
3
B. Dubois and M. L. Albert, “Amnestic MCI or prodromal Alzheimers
disease?,” Lancet Neurol. 34, 246–248 2004.
4
R. C. Petersen, “Mild cognitive impairment as a diagnostic entity,” J.
Intern. Med. 2563, 183–194 2004.
5
B. Winblad, K. Palmer, M. Kivipelto, V. Jelic, L. Fratiglioni, L.-O. Wahl-
und, A. Nordberg, L. Bäckman, M. Albert, O. Almkvist, H. Arai, H.
Basun, K. Blennow, M. de Leon, C. DeCarli, T. Erkinjuntti, E. Giacobini,
C. Graff, J. Hardy, C. Jack, A. Jorm, K. Ritchie, C. van Duijn, P. Visser,
and R. C. Petersen, “Mild cognitive impairment–beyond controversies,
towards a consensus: Report of the International Working Group on Mild
Cognitive Impairment,” J. Intern. Med. 2563, 240–246 2004.
6
K. Kantarci and C. R. Jack, “Quantitative magnetic resonance techniques
as surrogate markers of Alzheimers disease,” NeuroRx: The Journal of
the American Society for Experimental NeuroTherapeutics 112, 196
205 2004.
7
M. Atiya, B. T. Hyman, M. S. Albert, and R. Killiany, “Structural mag-
netic resonance imaging in established and prodromal Alzheimer disease:
A review,” Alzheimer Dis. Assoc. Disord. 173, 177–195 2003.
8
G. Chetelat and J.-C. Baron, “Early diagnosis of Alzheimers disease:
Contribution of structural neuroimaging,” Neuroimage 182, 525–541
2003.
9
K. Kantarci and C. R. Jack, “Neuroimaging in Alzheimer disease: An
evidence-based review,” Neuroimaging Clin. N. Am. 132, 197–209
2003.
10
V. C. Anderson, Z. N. Litvack, and J. A. Kaye, “Magnetic resonance
approaches to brain aging and Alzheimer disease-associated neuropathol-
ogy,” Top. Magn. Reson Imaging 166, 439–452 2005.
11
MICCAI 2008 Workshop on the Computational Anatomy and Physiology
of the Hippocampus unpublished, http://picsl.upenn.edu/caph08.
12
C. R. Jack, R. C. Petersen, P. C. O’Brien, and E. G. Tangalos, “MR-based
hippocampal volumetry in the diagnosis of Alzheimers disease,” Neurol-
ogy 421, 183–188 1992.
13
C. R. Jack, R. C. Petersen, Y. C. Xu, P. C. O’Brien, G. E. Smith, R. J.
Ivnik, B. F. Boeve, S. C. Waring, E. G. Tangalos, and E. Kokmen, “Pre-
diction of AD with MRI-based hippocampal volume in mild cognitive
impairment,” Neurology 527, 1397–1403 1999.
14
L. G. Apostolova, P. H. Lu, S. Rogers, R. A. Dutton, K. M. Hayashi, A.
W. Toga, J. L. Cummings, and P. M. Thompson, “3D mapping of mini-
mental state examination performance in clinical and preclinical Alzhe-
imer disease,” Alzheimer Dis. Assoc. Disord. 204, 224–231 2006.
15
L. G. Apostolova, I. D. Dinov, R. A. Dutton, K. M. Hayashi, A. W. Toga,
J. L.Cummings, and P. M. Thompson, “3D comparison of hippocampal
atrophy in amnestic mild cognitive impairment and Alzheimer’s disease,”
Brain: A Journal of Neurology 129, 2867–2873 2006.
16
J. G. Csernansky, S. Joshi, L. Wang, J. W. Haller, M. Gado, J. P. Miller,
U. Grenander, and M. I. Miller, “Hippocampal morphometry in schizo-
phrenia by high dimensional brain mapping,” Proc. Natl. Acad. Sci.
U.S.A. 9519, 11406–11411 1998.
17
G. B. Frisoni, F. Sabattoli, A. D. Lee, R. A. Dutton, A. W. Toga, and P. M.
Thompson, “In vivo neuropathology of the hippocampal formation in AD:
A radial mapping MR-based study,” Neuroimage 321, 104–110 2006.
18
P. M. Thompson, K. M. Hayashi, G. I. De Zubicaray, A. L. Janke, S. E.
Rose, J. Semple, M. S. Hong, D. H. Herman, D. Gravano, D. M. Dod-
drell, and A. W. Toga, “Mapping hippocampal and ventricular change in
Alzheimer disease,” Neuroimage 224, 1754–1766 2004.
19
P. Scheltens, D. Leys, F. Barkhof, D. Huglo, H. C. Weinstein, P. Vermer-
sch, M. Kuiper, M. Steinling, E. C. Wolters, and J. Valk, “Atrophy of
medial temporal lobes on MRI in ‘probable’ Alzheimers disease and
normal ageing: Diagnostic value and neuropsychological correlates,” J.
Neurol., Neurosurg. Psychiatry 5510, 967–972 1992.
3745 Calvini et al.: Automatic analysis of MTL for AD early assessment 3745
Medical Physics, Vol. 36, No. 8, August 2009
Page 9
20
O. T. Carmichael, H. A. Aizenstein, S. W. Davis, J. T. Becker, P. M.
Thompson, C. C. Meltzer, and Y. Liu, “Atlas-based hippocampus seg-
mentation in Alzheimers disease and mild cognitive impairment,” Neu-
roimage 274, 979–990 2005.
21
B. H. Ridha, J. Barnes, L. A. van de Pol, J. M. Schott, R. G. Boyes, M.
M. Siddique, M. N. Rossor, P. Scheltens, and N. C. Fox, “Application of
automated medial temporal lobe atrophy scale to Alzheimer disease,”
Arch. Neurol. 646,8498542007.
22
J. Barnes, R. I. Scahill, R. G. Boyes, C. Frost, E. B. Lewis, C. L. Rossor,
M. N. Rossor, and N. C. Fox, “Differentiating AD from aging using
semiautomated measurement of hippocampal atrophy rates,” Neuroimage
232, 574–581 2004.
23
W. R. Crum, R. I. Scahill, and N. C. Fox, “Automated hippocampal
segmentation by regional fluid registration of serial MRI: Validation and
application in Alzheimers disease,” Neuroimage 135, 847–855 2001.
24
B. Fischl, D. H. Salat, E. Busa, M. Albert, M. Dieterich, C. Haselgrove,
A. van der Kouwe, R. Killiany, D. Kennedy, S. Klaveness, A. Montillo,
N. Makris, B. Rosen, and A. M. Dale, “Whole brain segmentation: Au-
tomated labeling of neuroanatomical structures in the human brain,” Neu-
ron 333, 341–355 2002.
25
R. E. Hogan, K. E. Mark, L. Wang, S. Joshi, M. I. Miller, and R. D.
Bucholz, “Mesial temporal sclerosis and temporal lobe epilepsy: MR im-
aging deformation-based segmentation of the hippocampus in five pa-
tients,” Radiology 2161, 291–297 2000.
26
S. Powell, V. A. Magnotta, H. Johnson, V. K. Jammalamadaka, R. Pier-
son, and N. C. Andreasen, “Registration and machine learning-based au-
tomated segmentation of subcortical and cerebellar brain structures,”
Neuroimage 391, 238–247 2008.
27
L. Wang, F. Beg, T. Ratnanather, C. Ceritoglu, L. Younes, J. C. Morris, J.
G. Csernansky, and M. I. Miller, “Large deformation diffeomorphism and
momentum based hippocampal shape discrimination in dementia of the
Alzheimer type,” IEEE Trans. Med. Imaging 264, 462– 470 2007.
28
P. A. Yushkevich, J. Piven, H. C. Hazlett, R. G. Smith, S. Ho, J. C. Gee,
and G. Gerig, “User-guided 3D active contour segmentation of anatomi-
cal structures: significantly improved efficiency and reliability,” Neuroim-
age 313, 11161128 2006.
29
J. H. Morra, Z. Tu, L. G. Apostolova, A. E. Green, C. Avedissian, S. K.
Madsen, N. Parikshak, X. Hua, A. W. Toga, C. R. Jack, M. W. Weiner,
and P. M. Thompson, “Validation of a fully automated 3D hippocampal
segmentation method using subjects with Alzheimers disease mild cog-
nitive impairment, and elderly controls,” Neuroimage 431, 59–68
2008.
30
J. T. Becker, S. W. Davis, K. M. Hayashi, C. C. Meltzer, A. W. Toga, O.
L. Lopez, and P. M. Thompson, “Three-dimensional patterns of hippoc-
ampal atrophy in mild cognitive impairment,” Arch. Neurol. 631, 97–
101 2006.
31
D. L. Collins, P. Neelin, T. M. Peters, and A. C. Evans, “Automatic 3D
intersubject registration of MR volumetric data in standardized talairach
space,” J. Comput. Assist. Tomogr. 182, 192–205 1994.
32
G. McKhann, D. Drachman, M. Folstein, R. Katzman, D. Price, and E. M.
Stadlan, “Clinical diagnosis of Alzheimers disease: Report of the
NINCDS-ADRDA work group under the auspices of department of health
and human services task force on Alzheimers disease,” Neurology 347,
939–944 1984.
33
S. Katz, T. D. Downs, H. R. Cash, and R. C. Grotz, “Progress in devel-
opment of the index of ADL,” Gerontologist 101, 20–30 1970.
34
M. P. Lawton and E. M. Brody, “Assessment of older people: Self-
maintaining and instrumental activities of daily living,” Gerontologist
93, 179–186 1969.
35
M. F. Folstein, S. E. Folstein, and P. R. McHugh, “Mini-mental state. A
practical method for grading the cognitive state of patients for the clini-
cian,” J. Psychiatr. Res. 123, 189–198 1975.
36
J. L. Cummings, M. Mega, K. Gray, S. Rosenberg-Thompson, D. A.
Carusi, and J. Gornbein, “The neuropsychiatric inventory: Comprehen-
sive assessment of psychopathology in dementia,” Neurology 4412,
2308–2314 1994.
37
A. C. Evans, M. Kamber, D. L. Collins, and D. MacDonald, “An MRI–
based probabilistic atlas of neuroanatomy,” Magnetic Resonance Scan-
ning and Epilepsy, NATO Advance Studies Institute, Series A: Life Sci-
ences Vol. 264, edited by S. Shorvon, D. Fish, F. Andermann, G. Bydder,
and H. Stefan Plenum, New York, 1994, pp. 263–274.
38
L. Ibanez, W. Schroeder, L. Ng, and J. Cates, The ITK software guide,
2nd ed., 2005.
39
G. A. F. Seber, Multivariate Observations Wiley, New York, 1984.
40
Alzheimers Disease Neuroimaging Initiative ADNI, http://
www.loni.ucla.edu/ADNI/, 2007.
41
A. Kelemen, G. Székely, and G. Gerig, “Elastic model-based segmenta-
tion of 3-D neuroradiological data sets,” IEEE Trans. Med. Imaging
1810, 828–839 1999.
42
J. Yang and J. S. Duncan, “3D image segmentation of deformable objects
with joint shape-intensity prior models using level sets,” Med. Image
Anal. 83, 285–294 2004.
43
A. Pitiot, H. Delingette, P. M. Thompson, and N. Ayache, “Expert
knowledge-guided segmentation system for brain MRI,” Neuroimage 23,
S85–S96 2004.
44
D. Shen, S. Moffat, S. M. Resnick, and C. Davatzikos, “Measuring size
and shape of the hippocampus in MR images using a deformable shape
model,” Neuroimage 152, 422–434 2002.
45
S. Duchesne, J. Pruessner, and D. L. Collins, “Appearance-based segmen-
tation of medial temporal lobe structures,” Neuroimage 172, 515–531
2002.
46
Y.-Y. Hsu, N. Schuff, A.-T. Du, K. Mark, X. Zhu, D. Hardin, and M. W.
Weiner, “Comparison of automated and manual MRI volumetry of hip-
pocampus in normal aging and dementia,” J. Magn. Reson Imaging 163,
305–310 2002.
47
J. W. Haller, G. E. Christensen, S. C. Joshi, J. W. Newcomer, M. I. Miller,
J. G. Csernansky, and M. W. Vannier, “Hippocampal MR imaging mor-
phometry by means of general pattern matching,” Radiology 1993, 787–
791 1996.
48
K. M. Gosche, J. A. Mortimer, C. D. Smith, W. R. Markesbery, and D. A.
Snowdon, “An automated technique for measuring hippocampal volumes
from MR imaging studies,” AJNR Am. J. Neuroradiol. 229, 1686–1689
2001.
49
J. Webb, A. Guimond, P. Eldridge, D. Chadwick, J. Meunier, J. P. Thirion,
and N. Roberts, “Automatic detection of hippocampal atrophy on mag-
netic resonance images,” Magn. Reson. Imaging 178, 1149–1161
1999.
50
H. Rusinek, S. De Santi, D. Frid, W.-H. Tsui, C. Y. Tarshish, A. Convit,
and M. J. de Leon,“Regional brain atrophy rate predicts future cognitive
decline: 6-year longitudinal MR imaging study of normal aging,” Radi-
ology 2293, 691–696 2003.
51
J. G. Csernansky, L. Wang, S. Joshi, J. P. Miller, M. Gado, D. Kido, D.
McKeel, J. C. Morris, and M. I. Miller, “Early DAT is distinguished from
aging by high-dimensional mapping of the hippocampus. Dementia of the
Alzheimer type,” Neurology 5511, 1636–1643 2000.
52
O. Colliot, G. Chételat, M. Chupin, B. Desgranges, B. Magnin, H. Benali,
B. Dubois, L. Garnero, F. Eustache, and S. Lehéricy, “Discrimination
between Alzheimer disease, mild cognitive impairment, and normal aging
by using automated segmentation of the hippocampus,” Radiology
2481, 194–201 2008.
53
T. Tapiola, C. Pennanen, M. Tapiola, S. Tervo, M. Kivipelto, T. Hänninen,
M. Pihlajamäki, M. P. Laakso, M. Hallikainen, A. Hämäläinen, M. Van-
hanen, E.-L. Helkala, R. Vanninen, A. Nissinen, R. Rossi, G. B. Frisoni,
and H. Soininen, “MRI of hippocampus and entorhinal cortex in mild
cognitive impairment: A follow-up study,” Neurobiol. Aging 291, 31–38
2008.
54
D. P. Devanand, G. Pradhaban, X. Liu, A. Khandji, S. De Santi, S. Segal,
H. Rusinek, G. H. Pelton, L. S. Honig, R. Mayeux, Y. Stern, M. H. Tabert,
and M. J. de Leon, “Hippocampal and entorhinal atrophy in mild cogni-
tive impairment: Prediction of Alzheimer disease,” Neurology 6811,
828836 2007
55
A. Convit, J. de Asis, M. J. de Leon, C. Y. Tarshish, S. De Santi, and H.
Rusinek, “Atrophy of the medial occipitotemporal, inferior, and middle
temporal gyri in non-demented elderly predict decline to Alzheimers
disease,” Neurobiol. Aging 211, 19–26 2000.
56
C. R. Jack, M. M. Shiung, S. D. Weigand, P. C. O’Brien, J. L. Gunter, B.
F. Boeve, D. S. Knopman, G. E. Smith, R. J. Ivnik, E. G. Tangalos, and R.
C. Petersen, “Brain atrophy rates predict subsequent clinical conversion
in normal elderly and amnestic MCI,” Neurology 658, 1227–1231
2005.
57
C. DeCarli, G. B. Frisoni, C. M. Clark, D. Harvey, M. Grundman, R. C.
Petersen, L. J. Thal, S. Jin, C. R. Jack, and P. Scheltens, “Qualitative
estimates of medial temporal atrophy as a predictor of progression from
mild cognitive impairment to dementia,” Arch. Neurol. 641, 108–115
2007.
58
C. Geroldi, R. Rossi, C. Calvagna, C. Testa, L. Bresciani, G. Binetti, O.
3746 Calvini et al.: Automatic analysis of MTL for AD early assessment 3746
Medical Physics, Vol. 36, No. 8, August 2009
Page 10
Zanetti, and G. B. Frisoni, “Medial temporal atrophy but not memory
deficit predicts progression to dementia in patients with mild cognitive
impairment,” J. Neurol., Neurosurg. Psychiatry 7711, 1219–1222
2006.
59
M. J. de Leon, J. Golomb, A. E. George, A. Convit, C. Y. Tarshish, T.
McRae, S. De Santi, G. Smith, S. H. Ferris, and M. Noz., “The radiologic
prediction of Alzheimer disease: The atrophic hippocampal formation,”
AJNR Am. J. Neuroradiol. 144, 897–906 1993.
60
T. S. Yoo, M. J. Ackerman, W. E. Lorensen, W. Schroeder, V. Chalana, S.
Aylward, D. Metaxes, and R. Whitaker, in Proceedings of Medicine Meets
Virtual Reality Conference, edited by J. Westwood IOS, Amsterdam,
2002, pp. 586–592.
3747 Calvini et al.: Automatic analysis of MTL for AD early assessment 3747
Medical Physics, Vol. 36, No. 8, August 2009
Page 11
  • Source
    • "Additional studies have used other modalities such as functional MRI [99], only FDG-PET [100], diffusion tensor imaging [101, 102] , MRI and magnetic resonance spectroscopy [103], MRI and vitamin E [104], MRI and functional MRI [105], as well as MRI and magnetoencephalography [106]. The latter studies, in addition to other studies that used a lower number of subjects than our limit of 100 subjects per study107108109110111, or used classical statistical methods rather than multivariate or machine learning approaches [71,112113114, or performed only MCI classification [115, 116], have not been discussed here. In this article, we have presented key areas of multivariate analysis and machine learning including feature extraction, feature selection, classification, validation, and cohorts. "
    [Show abstract] [Hide abstract] ABSTRACT: Machine learning algorithms and multivariate data analysis methods have been widely utilized in the field of Alzheimer's disease (AD) research in recent years. Advances in medical imaging and medical image analysis have provided a means to generate and extract valuable neuroimaging information. Automatic classification techniques provide tools to analyze this information and observe inherent disease-related patterns in the data. In particular, these classifiers have been used to discriminate AD patients from healthy control subjects and to predict conversion from mild cognitive impairment to AD. In this paper, recent studies are reviewed that have used machine learning and multivariate analysis in the field of AD research. The main focus is on studies that used structural magnetic resonance imaging (MRI), but studies that included positron emission tomography and cerebrospinal fluid biomarkers in addition to MRI are also considered. A wide variety of materials and methods has been employed in different studies, resulting in a range of different outcomes. Influential factors such as classifiers, feature extraction algorithms, feature selection methods, validation approaches, and cohort properties are reviewed, as well as key MRI-based and multi-modal based studies. Current and future trends are discussed.
    Full-text · Article · Apr 2014 · Journal of Alzheimer's disease: JAD
  • Source
    • "CSF tau, phospho-tau, and amyloid measurements are in development [2] [3] but require lumbar puncture; therefore, a noninvasive imaging marker is more appealing for screening outpatients without hospital admission. With this purpose, volumetric MRI measures of mesiotemporal atrophy demonstrated to have some prognostic value [4] [5]. Compared to these volumetric measures of macrostructural damage, diffusion tensor imaging (DTI) indexes of microstructural damage within mesiotemporal lobe have shown to better discriminate MCI from controls [6] [7] and to better detect MCI converters [8] [9] [10]. "
    [Show abstract] [Hide abstract] ABSTRACT: Hippocampal damage, by DTI or MR volumetry, and PET hypoperfusion of precuneus/posterior cingulate cortex (PC/PCC) were proposed as biomarkers of conversion from preclinical (MCI) to clinical stage of Alzheimer's disease (AD). This study evaluated structural damage, by DTI and MR volumetry, of hippocampi and tracts connecting hippocampus to PC/PCC (hipp-PC/PCC) in 10 AD, 10 MCI, and 18 healthy controls (CTRL). Normalized volumes, mean diffusivity (MD), and fractional anisotropy (FA) were obtained for grey matter (GM), white matter (WM), hippocampi, PC/PCC, and hipp-PC/PCC tracts. In hippocampi and hipp-PC/PCC tracts, decreased volumes and increased MD were found in AD versus CTRL (P < .001). The same results with lower significance (P < .05) were found in MCI versus CTRL. Verbal memory correlated (P < .05) in AD with left hippocampal and hipp-PC/PCC tract MD, and in MCI with FA of total WM. Both DTI and MR volumetry of hippocampi and hipp-PC/PCC tracts detect early signs of AD in MCI patients.
    Full-text · Article · Jan 2012 · Neurology Research International
  • Source
    • "Those brain changes are prevalent in MCI (Apostolova et al., 2006; Bombois et al., 2008; Calvini et al., 2009; Jack et al., 1999) and have been associated with different memory deficit patterns (Nordahl et al., 2005; Nordlund et al., 2007; Villeneuve et al., 2011). Therefore, they may cause different memory changes in MCI and contribute to cognitive heterogeneity . "
    [Show abstract] [Hide abstract] ABSTRACT: The main goal of this study was to assess vulnerability to proactive interference and memory binding capacity, the ability to combine different information into a single coherent memory event, in persons with mild cognitive impairment (MCI). We also examined whether hippocampal atrophy and vascular burden were differentially related to these memory capacities in MCI. We further assessed whether memory performance and brain changes differ as a function of later development (or not) of dementia and whether they can predict progression to dementia. The study included 77 participants, 49 meeting the criteria for MCI and 28 healthy older adults. Results showed binding deficits and greater vulnerability to proactive interference in persons with MCI compared with healthy older adults. Hippocampal volume was associated with binding capacity, whereas vascular burden was associated with resistance to interference in persons with MCI. Follow-up analyses indicated that binding deficits predict progression from MCI to dementia. In conclusion, binding deficits and vulnerability to proactive interference are present in persons with MCI and are associated with different brain markers. However, only binding deficits predict progression to dementia.
    Full-text · Article · Nov 2011 · Neurobiology of aging
Show more