Early Alzheimer’s Disease Progression Detection using

Multi-subnetworks of the Brain

Jaroslav Rokicki

1,2

, Hiyoshi Kazuko

1,3

, Francois-Benoit Vialatte

, Andrius Uˇsinskas

and

Andrzej Cichocki

Electrical Engineering Department, Vilnius Gediminas Technical University, Saul˙etekio av. 11, Vilnius, Lithuania

Laboratory of Advanced Brain Signal Processing, Brain Science Institute, RIKEN, Wako-shi, Saitama, Japan

Department of Functional Brain Imaging, Human Research Center,

Kyoto University Graduate School of Medicine, Kyoto, Japan

SIGnal Processing and MAchine Learning Laboratory, ESPCI ParisTech, Paris, France

Keywords:

Alzheimer’s Disease, Brain Atrophy, Segmentation of Brain Subnetworks, Hippocampus, Amygdala,

Entorhinal Cortex, Multi-volume, Classiﬁcation, LDA, Early Detection.

Abstract:

Alzheimer’s disease is neurodegenerative disorder believed to affect 24.3 million people worldwide. Pro-

posed MRI based disease progression markers have shown ability to perform the classiﬁcation between the

Alzheimer’s Disease (AD), Mild Cognitive Impariment (MCI) and Normal Cognitive (NC) subjects. We ex-

ploited two approaches, ﬁrst one is to use single sub-network volume as a feature, second to use a network of

most discriminative sub-networks. Multi-feature approach showed improvement by 4.5% in AD/NC classiﬁ-

cation case, and 1.5 % in MCI/NC case. Study was summarized for 48 AD, 119 MCI and 66 NC subjects.

1 INTRODUCTION

People hit by Alzheimer’s disease (AD) live in aver-

age around 8 years, but there are cases of surviving

up to 20 years. As for now there is no cure for the

Alzheimer disease, but a large number of new com-

pounds are constantly being developed to modify the

ﬂow of disease or to slow down its progression. Since

AD-related brain atrophy is irreversible, its early de-

tection is extremely important. This allows clinicians

to introduce new treatment as early as possible.

Currently diagnosis of Alzheimer relies largely on

medical documentation. There is no single test which

could show whether the subject already is struck by

Alzheimer’s disease. Therefore, a list of mental as-

signments is performed (Mini-Mental State Examina-

tion, Clinical Dementia Rating, Clock Drawing Test,

Hachinski Ischemic Score) and a complete test of the

medical history of the subject and his family members

is done. The earliest brain changes leading to devel-

opment of AD may begin up to 20 years before the

external symptoms appear. Therefore, we have cho-

sen MRI as a source for early AD progression mark-

ers. Such decision is motivated by a fact that MRI is

widely accessible, with standard protocols across dif-

ferent vendors. Therefore, we believe that MRI scan

will be an integral part of the annual health check rou-

tine, at least for elderly subjects. In such case, a huge

amount of data would overwhelm a single physician.

Therefore,automatic and physician friendly computer

analysis is urgently needed.

In this study we employed the data of 233 subjects

from ADNI database, to assess the automatic clas-

siﬁcation between the Alzheimer’s disease (AD) pa-

tients, mild cognitive impairment (MCI) and normal

cognitive (NC) subjects. We treat these groups as two

separate problems, AD versus NC, and MCI versus

NC classiﬁcation. Since according to in 6 years 80 %

of the MCI subjects are expected to developdementia,

we found it useless to try discriminate between MCI

converters and non-converters. Rather we treat MCI

itself as an early stage of Alzheimer’s disease.

We present a novel early AD detection technique

based on the multiple (9 regions, 1 ﬁg.) sub-network

volume descriptors extracted from the brain. This ap-

proach improves results compared to one’s based on

the single region of interest by 1.5-4.5 depending on

the stage of subject. Other novelty of this paper pro-

posed a new sub-network volume. Due to difﬁculty to

perform precise automatic segmentation between the

684

Rokicki J., Kazuko H., Vialatte F., Ušinskas A. and Cichocki A..

Early Alzheimer’s Disease Progression Detection using Multi-subnetworks of the Brain.

DOI: 10.5220/0004182806840691

In Proceedings of the 4th International Joint Conference on Computational Intelligence (SSCN-2012), pages 684-691

ISBN: 978-989-8565-33-4

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

hippocampus and amygdala, we propose to integrate

these two volumes.

2 BACKGROUND

All the automatic classiﬁcation MRI based methods

can be roughly divided into 3 big groups, depending

on the type of features used to perform the classiﬁ-

cation: direct approach or probability maps for three

tissue classes (white matter (WM); gray matter (GM)

and celebro spinal ﬂuid (CSF)); atlas based approach,

when atlas based segmentation is followed by extrac-

tion of volumes of interest; or single region of interest

analysis.

2.1 Direct Approach

In direct approach all the brain region voxels are

used as an input data. To align the subjects data is

smoothed (8 mm kernel is usually used) and regis-

tered with template. To diminish the data size, often

downsampling by averaging is used (Vemuri et al.,

2008). Moreover, each voxel is assigned to one of

3 anatomical classes (WM, GM or CSF). (Vemuri

et al., 2008) reports his algorithm obtained sensitiv-

ity of 88% and speciﬁcity of 86% for AD versus MCI

subjects discrimination with linear-SVM. Moreover

authors added age and gender to the feature vector im-

proving results up to sensitivity 88% and speciﬁcity

90%. (Kloppel et al., 2008) in his study proposed

two different approaches. One is based on extracted

hippocampal volume, while the other one uses sim-

ilar approach to (Vemuri et al., 2008), but this time

no downsampling of input data is reported. The ac-

curacy of classiﬁcation is 87.5% (sensitivity 95.0%,

speciﬁcity 80.0%). There was no signiﬁcant improve-

ment in performance if non-linear kernels were used

compared to linear-SVM (Kloppel et al., 2008). Sim-

ilar approach was taken by (Fan et al., 2008). The

difference is after the registration with template, re-

Figure 1: Analyzed brain sub-networks.

gional volumetric maps (so called RAVENS maps)

were created (Goldszal et al., 1998). Results of clas-

siﬁcation via leave-one-out cross validation method

were 94.3% for AD vs NC, 81.8% for MCI vs NC and

74.3% for MCI vs AD (Fan et al., 2008). (Davatzikos

et al., ) using the brain watershed-based clustering

method and other techniques described in (Goldszal

et al., 1998) tries to distinguish between the MCI

and NC subjects using only cross-sectional data. Ap-

plying leave-one-out procedure with SVM non-linear

classiﬁer for the 30 subjects he receives accuracy of

90%.

2.2 Atlas based Approach

Atlas based segmentation is common technique par-

cel brain into non-overlapping, anatomical regions.

An unseen image is registered with labeled atlas and

labels are transposed onto the unseen subject’s vol-

ume. (Magnin et al., 2009) in his study after register-

ing brains with MNI152 template parceled them into

90 regions of interest (ROI). The white matter for 34

most signiﬁcant sub-networks was selected and mod-

eled by Gaussian distribution. SVM with radial basis

function was used for classiﬁcation. (Magnin et al.,

2009) claims speciﬁcity 96.6%, sensitivity 91.5%.

(Desikan et al., 2009) after atlas registration, selected

entorhinal cortex thickness, hippocampal volume and

supra-marginalgyrus thickness as disease progression

markers. Authors report speciﬁcity of 91% and sen-

sitivity of 90% for the cohort based on the ADNI

database when he tried to separate the MCI and NC

subjects, and ideal results for the AD vs NC classiﬁ-

cation. (Kloppel et al., 2008) after performing atlas

registration, uses only the region cropped from vol-

ume around hippocampus region (12× 16× 12 mm)

to classify between the AD vs NC, which results with

sensitivity 75.8%, speciﬁcity 91.2%.

2.3 Volume of Interest Analysis

Methods of this group are closely related to the at-

las based approach. In most of cases atlas registra-

tion is used to segment the desired subnetwork from

the other brain tissues. The difference is, that in sub-

network of interest approach whole classiﬁcation pro-

cedure is dependent upon the features extracted from

one region solely, rather than on network of differ-

ent regions. Moreover, the registration step is often

used only as an initialization step for more complex

algorithm. Most authors rely on subnetwork of hip-

pocampus solely. Nevertheless, according to (Braak

and Braak, 1997) there is an evidence that early AD

pathologymay start in entorhinal cortex and only then

EarlyAlzheimer'sDiseaseProgressionDetectionusingMulti-subnetworksoftheBrain

685

progress to hippocampus.

Labor intensive methods involving manual region

delineation of hippocampus and entorhinal cortex

were performed by (Pennanen et al., 2004), (Juotto-

nen et al., 1998) and (Du et al., 2001). For AD/NC

classiﬁcation accuracy was in range from 86-91 % if

hippocampus was used and 82-83 % if entorhinal cor-

tex volume was as an input. In MCI/NC case accu-

racies were 60-70 % for hippocampus and 66 % for

entorhinal cortex volume. Important conclusion was

that incorporating few subnetworks together can im-

prove overall results, for example in AD case if both

sub-network volumes were incorporated accuracy im-

proved by +3 % (Du et al., 2001).

(Fan et al., 2008) in the second part of his study

used the volume of hippocampus (left and right)

against the entorhinal cortex (left and right) after nor-

malization by intra-cranial volume. The received ac-

curacy using linear-SVM and leave-one-outcross val-

idation was 82.0% for AD vs NC 76.0% for MCI/NC,

and AD/MCI 58.3% respectively. (Chupin et al.,

2009) proposed fully automatic approach based on

anatomical priors for the hippocampus region extrac-

tion. Leave-one-out approach was used for testing.

This method proved to be accurate in 76-80 % in

NC/AD classiﬁcation case. (Lotjonen et al., 2011)

proposes to perform the automatic hippocampus ex-

traction based on multi-atlas segmentation frame-

work, adding the partial volume effect correction

(Tohka et al., 2004). The basic idea is to register non-

rigidly the unseen data with a template and to select

the most similar atlas compared to the registered un-

seen data. Tissue class having the highest probabil-

ity in voxel is chosen as feature for ﬁnal segmenta-

tion. Simplest linear classiﬁer was used with accuracy

75.0 % in AD/NC case.

(Gerardin et al., 2009) uses hippocampus segmen-

tation method provided by (Chupin et al., 2009) as

a ﬁrst step, followed by hippocampus shape descrip-

tion by spherical harmonics coefﬁcients. It is a math-

ematical approachto represent surfaces with spherical

topology, which can be seen as 3D Fourier series ex-

pansion (Gerardin et al., 2009). The best combination

of parameters gave sensitivity 96 % and speciﬁcity of

92 % in AD vs NC classiﬁcation. MCI vs NC case

best result was sensitivity 83 %, speciﬁcity 84 %. Re-

sults were validated for only 23 AD and 23 MCI sub-

jects.

3 SUBJECTS

We studied and analyzed data of 48 patients

with AD (25 males, 23 females, age±standard-

deviation (SD)=76.6 ± 6.3, Mini Mental State Ex-

amination (MMSE)±(SD)=23.5 ± 1.9), 119 subjects

with MCI (79 males, 40 females, age±(SD)=75.1±

7.4, (MMSE)±(SD)=27.2 ± 1.6) and 66 NC sub-

jects (40 males, 26 females, age±(SD)=76.3 ±

4.6, (MMSE)±(SD)=29.2± 0.9) recruited for ADNI

study. Provided MMSE scores correspond the

ﬁrst screening at the hospital. Detailed subject

list with ID for ADNI database can be found at

http://pages.cs.wisc.edu/˜hinrichs/

. All the

subjects were followed up for two years, therefore

data was collected during the ﬁrst visit and the same

procedures repeated in two years.

ADNI eligibility criteria in detail are described at

http://www.adni-info.org

. Brieﬂy the subjects

are 55-90 years old, if they havea memory complains.

Have to be ﬂuent in Spanish or English, accompany-

ing person has to be present. Speciﬁc psychoactive

medications are excluded (Fennema-Notestine et al.,

2009). In this study all the ADNI subjects will be di-

vided into 3 groups, based on criteria as follows:

• Normal Cognitive (NC) - Mini-Mental State

Examination (MMSE) scores between 24 and

30 (maximal), CDR of 0 (3 maximal), non-

depressed, no memory complains.

• Mild Cognitive Impairment (MCI) - MMSE

scores between 24 and 30, memory complaint,

preferably corroborated by an informant, objec-

tive memory loss measured by a education ad-

justed scores on Wenchsler Memory Scale Logi-

cal Memory, a CDR of 0.5, absence of signiﬁcant

levels of impairment in other cognitive domains,

essentially preserved activities of daily living.

• Alzheimer’s Disease Subject (AD) - MMSE

scores between 20 and 26, CDR 0.5 or 1, mem-

ory complaint.

4 METHODS

Volumetric measures were created in 7 subcortical re-

gions: hippocampus, amygdala, caudate, thalamus,

pallidus, putamen, and two brain ventrical networks,

namely lateral and 3rd ventricle. In addition, two gray

matter structures temporal pole and entorhinal cortex.

All together 10 different brain regions were investi-

gated. Automatic 3D whole-brain segmentation pro-

cess was based on publicly available FreeSurfer soft-

ware package.

All experiments were performed using the sta-

ble release v5.1.0 with a HP Z800 workstation (pa-

rameters: 94.6 GB of RAM memory, 64-bit Intel



Xeon

 Six-Core Processor X5670 with a processing

IJCCI2012-InternationalJointConferenceonComputationalIntelligence

686

speed 2.93 GHz, 12 MB cache). During all the ex-

periments Ubuntu 11.10, 64-bit operating system was

used. The average processing time for one subject

was 12 hours. After automatic processing was ﬁn-

ished all the subjects were checked manually for the

artifacts and segmentation errors.

Before we can measure individual volumes for

the different subcortical structures we should decide

“how to normalize brain size variability among dif-

ferent subjects?”. According to our measurements

registering all the brains with a MNI152 template

(a 9 degrees of freedom afﬁne transformation was

used; transformation matrix was calculated using

FreeSurfer software) gave better results, than struc-

ture normalization by cortical area. By better we

mean, that differences between the subject groups in

ﬁrst case were larger compared to normalization by

cortical volume, leading to better AUC scores.

We analyzed two approaches. First, each extracted

sub-network volume was used solely as an input fea-

ture. Each sub-network was described by 3 features:

• V

,mm

– structure volume at the base line visit

(month M

), for the structure name SN.

• V

,mm

– structure volume after 24 months had

passed from the base line visit, for the structure

name SN

• V

∆

,% – structure volume change in 24 month.

This criteria was added to have one dimensionless

descriptor for the each volume.

Moreover, we proposeone new sub-network. Hip-

pocampus and amygdala are close to indistinguish-

able using only the intensity information available in

MRI (Fischl et al., 2002). Therefore, automatic soft-

ware performs the segmentation based on the spatial

information provided by the atlas and local spacial re-

lationship, such as “posterior amygdala is frequently

superior to anterior hippocampus, but never inferior

to it” (Fischl et al., 2002). Therefore, we propose to

integrate these volumes, since they are neighbors and

the border between them is hard to deﬁne. We call our

new marker HA or hippocampus+amygdala.

Each feature was checked for its suitability to be

employed for the automatic classiﬁcation. This was

done by evaluating the AUC for each feature, together

with Fisher score (equation 1), due to it’s close rela-

tion with LDA. The main advantage of AUC over the

Fisher score is that it provides a non-parametric rep-

resentation of the diagnostic accuracy of the feature.

Moreover, features from completely different sources

or studies can be compared to each other by comput-

ing

F =

. (1)

Here, in a 2 class case, S

presents the so called be-

tween class scatter of the original feature vectors:

= µ

− µ

. (2)

- is the mean for each class i, while S

presents

within class scatter:

= σ

+ σ

. (3)

where, σ

- is a measure of variability in each class i.

At the ﬁrst stage, we performed classiﬁcation us-

ing each feature separately. Then we tested a new,

multi-volume based approach. It utilizes a group of

the subnetwork based features, rather than on any sin-

gle feature. The vector was constructed by sorting

the features according their Fisher score in descend-

ing fashion, similar to:

V = [V

M00

M24

,∆

M00

,...,∆

LastVolume

], (4)

where, Hp stands for the hippocampus and Am for

amygdala.

Starting with a single feature, with each subse-

quent iteration we includedextra feature as an input to

the classiﬁcator. Since we used 9 subnetworks(amyg-

dala and hippocampus were integrated) in our study

and 3 descriptors per each volume in the ﬁnal itera-

tion there was 27 features used as an input data.

5 RESULTS

The best discriminative abilities have volumes of the

hippocampus (AUC = 0.86), amygdala (AUC = 0.85)

and entorhinal cortex (AUC = 0.80) in both AD and

MCI cases (ﬁg. 2) indicating that disease mostly pro-

gresses in medial temporal lobe and it’s subnetworks.

Figure 2: The most discriminative regions in AD and MCI

cases: 1) Hippocampus and Amygdala; 2) Entorhinal Cor-

tex; 3) Lateral Ventricle (only in AD case, in MCI case sub-

stituted by pallidus).

EarlyAlzheimer'sDiseaseProgressionDetectionusingMulti-subnetworksoftheBrain

687

Table 1: AUC scores for investigated structures (V

and

are given in mm

, andV

∆

in % from V

) in AD/NC

case.

Feature Name V

∆

Hipp.+Amyg. 0.88 0.94 0.81

Hippocampus 0.86 0.92 0.82

Amygdala 0.85 0.91 0.70

Entorhinal Cortex 0.80 0.92 0.78

Lateral Ventricle 0.68 0.76 0.61

Putamen 0.60 0.75 0.71

Caudate 0.57 0.47 0.60

Thalamus 0.57 0.70 0.61

Temporal Pole 0.56 0.71 0.74

3rd Ventricle 0.51 0.58 0.53

Pallidus 0.50 0.62 0.61

Table 2: AUC scores for investigated structures (V

and

are given inmm

, andV

∆

in % fromV

) in MCI/NC

case.

Feature Name V

∆

Hipp.+Amyg. 0.71 0.74 0.67

Hippocampus 0.71 0.74 0.66

Amygdala 0.68 0.70 0.62

Entorhinal Cortex 0.64 0.69 0.59

Pallidus 0.62 0.59 0.51

Thalamus 0.61 0.61 0.53

Putamen 0.59 0.60 0.57

Temporal Pole 0.56 0.58 0.55

Caudate 0.55 0.52 0.53

Lateral Ventricle 0.54 0.56 0.57

3rd Ventricle 0.53 0.54 0.51

The results for the AUC scores are presented in ta-

bles 1 for AD/NC case and 2 for MCI/NC case. Also

ﬁgure 3 presents all 3 ROC curves for the AD/NC

case. In both cases the best discriminative ability had

been shown by integrated hippocampusand amygdala

volume. Therefore, for our classiﬁcation task instead

of separate amygdala and hippocampus volumes, we

will use the integrated one. Second remark would

be, that the discriminative strength increases as time

passes. It’s due to fact, that in AD (same as MCI) the

subject’s brain deteriorates faster compared to a NC

subject. Thus, differences between structures become

bigger in two years. Therefore, they have more dis-

criminative power.

Finally, the last remark would be, that in both AD

and MCI cases, the cross-sectional differences V

and V

, show better AUC scores, compared to the

longitudinal changes in volume, presented by V

∆

This indicates, that incubating period for the disease is

long. Therefore, volume change in 2 years is weaker

compared to ones, happened before the subject has

0 0.2 0.4 0.6 0.8 1

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

False positive rate (1−Specificity)

True positive rate (Sensitivity)

ROC curve for Hippocampus+Amygdala (NC vs AD)

ROC at M00 (AUC=0.88)

ROC at M24 (AUC=0.94)

ROC for Delta24 (AUC=0.81)

Random classifier

Cut−off point

Figure 3: ROC curve in AD/NC case for the best MRI data

based feature – Hippocampus+Amygdala. V

– red line,

– blue line, V

∆

– black line.

been included to the study. The other way around, it

can happen that some NC subjects are on a way to

develop dementia, and the shrinkage of some brain

subnetworks just started recently. Therefore, while

the absolute volumes of the subnetworks are still rela-

tively large, they start to shrink with intensity compa-

rable to the MCI or AD groups, so we shouldn’t rely

on the volume changes solely.

5.1 Classiﬁcation

To evaluate automatic classiﬁcation results we used

Linear Discriminant Analysis (LDA) and Quadratic

Discriminant Analysis (QDA). These methods max-

imize ratio of between-class variance to the within-

class variance in any particular data set and guar-

antees maximal separability. These classiﬁcators in

contrast to Support Vector Machines (SVM) are pa-

rameter free. Therefore it is easy to interpret results.

Leave-one-outwas used as a cross-validation strategy.

The accuracy in the ﬁgures 4(a), 4(b), 5(a) and

5(b) is presented by means of balanced Accuracy

(b − ACC) in y-axis, which represents the averaged

sensitivity and speciﬁcity value:

b− ACC =

SEN+

SPE, (5)

here, SEN - sensitivity, SPE - speciﬁcity.

Single Feature, AD vs NC. The results of cross val-

idation are presented in ﬁgure 4(a). None of the sin-

gle features is able to perform ideal classiﬁcation.

The best MRI based marker performance was ob-

tained by hippocampusat M

with b−ACC = 84.4%,

while integrated hippocampus and amygdala volumes

IJCCI2012-InternationalJointConferenceonComputationalIntelligence

688

0.45

0.5

0.55

0.6

0.65

0.7

0.75

0.8

0.85

0.9

Structure name

b−ACC

AD vs NC for Single Feature

Hipp+Amyg

Hipp

Amyg

Entor

Pal

Thal

Put

TempP

Caud

LatV.

3Vent

M00 LDA

M00 QDA

M24 LDA

M24 QDA

D24 LDA

D24 QDA

(a)

0.4

0.45

0.5

0.55

0.6

0.65

0.7

0.75

Structure name

b−ACC

MCI vs NC for Single Feature

Hipp+Amyg

Hipp

Amyg

Entor

Pal

Thal

Put

TempP

Caud

LatV.

3Vent

M00 LDA

M00 QDA

M24 LDA

M24 QDA

D24 LDA

D24 QDA

(b)

Figure 4: Accuracy for LDA (dashed line) and QDA (dotted line) in Single-Feature case, for V

- red, V

- blue, V

∆

black lines in (a) AD/NC (b) MCI/NC cases.

were right behind, with b − ACC = 83.9% at M

But at M

best b − ACC belongs already to hip-

pocampus+amygdala with b − ACC = 85.4% (SEN

= 80.3%, SPE = 87.5%), while hippocampus solely

scores b − ACC = 84.4%. Generally speaking most

of MRI based features improved as time passed, ex-

cept for caudate and 3rd ventricle. While the relative

change V

∆

shows better results only in the case of

the temporal pole, possibly indicating that the sub-

network is involved in the disease at a later stages.

Moreover, in the most cases LDA showed better clas-

siﬁcation results than QDA.

Single Feature, MCI vs NC. The same procedure

was repeated for the MCI vs NC subjects. The

b − ACC results are summarized in the ﬁgure 4(b).

The best discriminative ability at V

is shown by

entorhinal cortex with b− ACC = 67.9%, followed by

hippocampus (66.2%) and hippocampus and amyg-

Table 3: Five best features, represented by Sensitivity and

Speciﬁcity values with LDA classiﬁer for investigated fea-

tures in AD/NC case (HA - hippocampus+amygdala, Hp -

hippocampus, Am - amygdala, EC - entorhinal cortex, LV

- lateral venricle).

Name V

∆

SEN SPE SEN SPE SEN SPE

HA 80.3 87.5 83.3 87.5 84.8 68.8

Hp 83.3 85.4 83.3 85.4 84.8 66.7

Am 81.8 77.1 86.4 83.3 72.7 60.4

EC 75.8 72.9 83.3 81.2 81.8 64.6

LV 72.7 60.4 75.8 66.7 62.1 56.2

dala combined volumes (64.7%). While after two

years, the best discriminative abilities shift to hip-

pocampus+amygdala (69.7% with QDA classiﬁca-

tor), followed by hippocampus (68.4%) and amyg-

dala (66.4%), while entorhinal cortex declines to

66.7%. Relative changes based features presented re-

sults were worse compared to the absolute volume

data, except for putamen and lateral ventricle, but in

both cases results were close to random.

Multi-feature, AD vs NC. The second approach was

to combine the strongest markers together. First we

use training set, to sort features in descending order,

according to their Fisher score. The 7 most discrimi-

native features in AD/NC were:

V = [V

Ent

∆

Lat

]. (6)

Starting from the most important feature, we in-

clude less important ones in each iteration. So, if

Table 4: Five best features represented by their Sensitivity

and Speciﬁcity values with LDA classiﬁer for investigated

features in MCI/NC case (HA - hippocampus+amygdala,

Hp - hippocampus, Am - amygdala, EC - entorhinal cortex,

Pl - pallidus).

Name M

∆

SEN SPE SEN SPE SEN SPE

HA 69.7 59.7 72.7 63.0 66.7 57.1

Hp 72.7 59.7 71.2 65.5 68.2 54.6

Am 68.2 59.7 69.7 63.0 59.1 56.3

EC 72.7 63.0 65.2 68.1 60.6 47.1

Pl 60.6 60.5 57.6 56.3 43.9 48.7

EarlyAlzheimer'sDiseaseProgressionDetectionusingMulti-subnetworksoftheBrain

689

5 10 15 20 25

0.83

0.84

0.85

0.86

0.87

0.88

0.89

Nr of Features

b−ACC

AD vs NC Multi Features

LDA

QDA

(a)

5 10 15 20 25

0.62

0.63

0.64

0.65

0.66

0.67

0.68

0.69

0.7

0.71

0.72

Nr of Features

b−ACC

MCI vs NC Multi Features

LDA

QDA

(b)

Figure 5: Accuracy for LDA (red) and QDA (blue) in Multi-Feature case (a) AD/NC (b) MCI/NC cases.

in ﬁrst iteration we use one feature, namely hip-

pocampus V

and the result is exactly the same as

if single feature would be used. Next iteration we

add entorhinal cortex, and the feature vector becomes

Ent

], this is continued until all the features are

included. Using 9 volumes (volumes of hippocampus

and amygdala were combined together), we have 27

features. Results for the full feature vector are pre-

sented in ﬁgure 5(a) for AD/NC case and 5(b) for

MCI/NC case.

The best obtained accuracy was, when ﬁrst 3 fea-

tures were used, and is equal to b − ACC = 88.5%

(SEN=87.9%, SPE=89.1%). Results obtained by the

multi-approach are better than based on single feature

approach by 4.5%.

Multi-feature, AD vs MCI. The best accuracy was

71.2% (SEN = 72.7%, SPE = 69.8%) when 6 fea-

tures were used (LDA). 7 most discriminative features

were:

V = [V

∆

Ent

]. (7)

5 from the 6 strongest features in MCI case be-

long, to hippocampus, amygdala and entorhinal cor-

tex regions. This conﬁrms one more time that disease

starts it’s progression in the sub-networks of medial

temporal lobe.

Compared with single feature case, where the best

accuracy score was from 69.7 % for V

, result im-

proved by 1.5% and became 71.2%.

6 CONCLUSIONS

1. We propose the integrated marker, consisting of

hippocampus and amygdala volumes improves

the AUC for M

and M

. Proposed marker

improves ROC curve compared to hippocam-

pus or amygdala volumes separately in both AD

(AUC

= 0.88, AUC

= 0.94, AUC

∆24

0.81) and MCI (AUC

= 0.71, AUC

= 0.74,

AUC

∆24

= 0.67) cases.

2. The best score for a single feature was obtained

when hippocampus+amygdala marker was used

with b − ACC = 85.4% (SEN = 80.3%, SPE =

87.5%) in AD and 67.9% (SEN = 72.7%, SPE =

63.0%) in MCI case.

3. Multi-Feature approach gives the classiﬁcation re-

sults improvement by 4.5% compared to the sin-

gle feature case for AD/NC classiﬁcation. The

best result was 88.5% (SEN = 87.9%, SPE =

89.1%) when 3 features were used. In MCI case

using our approach we improved accuracy by

1.5% and is equal to b − ACC = 71.2% (SEN=

72.7%, SPE = 69.8%) when 6 features were used.

In both AD/NC and MCI/NC cases QDA didn’t

show any signiﬁcant advantage over LDA.

REFERENCES

Braak, H. and Braak, E. (1997). Staging of alzheimer-

related cortical destruction. Int Psychogeriatry, 9

Suppl 1.

Chupin, M., Gerardin, E., Cuingnet, R., Boutet, C.,

Lemieux, L., Lehericy, S., Benali, H., Garnero, L.,

and Colliot, O. (2009). Fully automatic hippocampus

segmentation and classiﬁcation in alzheimer’s disease

and mild cognitive impairment applied on data from

adni. Hippocampus, 19(6):579–587.

Davatzikos, C., Fan, Y., Wu, X., Shen, D., and Resnick,

S. M. Detection of prodromal alzheimer’s disease via

IJCCI2012-InternationalJointConferenceonComputationalIntelligence

690

pattern classiﬁcation of magnetic resonance imaging.

Neurobiology of Aging, 29(4):514–523.

Desikan, R. S., Cabral, H. J., Hess, C. P., Dillon, W. P.,

Glastonbury, C. M., Weiner, M. W., Schmansky, N. J.,

Greve, D. N., Salat, D. H., Buckner, R. L., Fischl, B.,

and Initiative, A. D. N. (2009). Automated mri mea-

sures identify individuals with mild cognitive impair-

ment and alzheimer’s disease. Brain, 132(8):2048–

2057.

Du, A. T., Schuff, N., Amend, D., Laakso, M. P., Hsu,

Y. Y., Jagust, W. J., Yaffe, K., Kramer, J. H., Reed, B.,

Norman, D., Chui, H. C., and Weiner, M. W. (2001).

Magnetic resonance imaging of the entorhinal cortex

and hippocampus in mild cognitive impairment and

alzheimer’s disease. Journal of Neurology, Neuro-

surgery & Psychiatry, 71(4):441–447.

Fan, Y., Batmanghelich, N., Clark, C. M., and Davatzikos,

C. (2008). Spatial patterns of brain atrophy in mci

patients, identiﬁed via high-dimensional pattern clas-

siﬁcation, predict subsequent cognitive decline. Neu-

roImage, 39(4):1731–1743.

Fennema-Notestine, C., Hagler, D. J., McEvoy, L. K.,

Fleisher, A. S., Wu, E. H., Karow, D. S., and Dale,

A. M. (2009). Structural mri biomarkers for preclini-

cal and mild alzheimer’s disease. Human Brain Map-

ping, 30(10):3238–3253.

Fischl, B., Salat, D. H., Busa, E., Albert, M., Dieterich,

M., Haselgrove, C., van der Kouwe, A., Killiany, R.,

Kennedy, D., Klaveness, S., Montillo, A., Makris, N.,

Rosen, B., and Dale, A. M. (2002). Whole brain

segmentation: Automated labeling of neuroanatomi-

cal structures in the human brain. Neuron, 33(3):341–

355.

Gerardin, E., Chetelat, G., Chupin, M., Cuingnet, R., Des-

granges, B., Kim, H.-S., Niethammer, M., Dubois, B.,

Lehericy, S., Garnero, L., Eustache, F., and Colliot, O.

(2009). Multidimensional classiﬁcation of hippocam-

pal shape features discriminates alzheimer’s disease

and mild cognitive impairment from normal aging.

NeuroImage, 47(4):1476–1486.

Goldszal, A. F., Davatzikos, C., Pham, D. L., Yan, M. X.,

Bryan, R. N., and Resnick, S. M. (1998). An image-

processing system for qualitative and quantitative vol-

umetric analysis of brain images. Journal Of Com-

puter Assisted Tomography, 22(5):827–837.

Juottonen, K., Laakso, M., Insausti, R., Lehtovirta, M.,

Pitknen, A., Partanen, K., and Soininen, H. (1998).

Volumes of the entorhinal and perirhinal cortices in

alzheimers disease. Neurobiology of Aging, 19(1):15–

22.

Kloppel, S., Stonnington, C. M., Chu, C., Draganski, B.,

Scahill, R. I., Rohrer, J. D., Fox, N. C., Jack, C. R.,

Ashburner, J., and Frackowiak, R. S. J. (2008). Auto-

matic classiﬁcation of mr scans in alzheimer’s disease.

Brain, 131(3):681–689.

Lotjonen, J., Wolz, R., Koikkalainen, J., Julkunen, V., Thur-

fjell, L., Lundqvist, R., Waldemar, G., Soininen, H.,

and Rueckert, D. (2011). Fast and robust extraction

of hippocampus from mr images for diagnostics of

alzheimer’s disease. NeuroImage, 56(1):185–196.

Magnin, B., Mesrob, L., Kinkingnhun, S., Plgrini-Issac,

M., Colliot, O., Sarazin, M., Dubois, B., Lehricy, S.,

and Benali, H. (2009). Support vector machine-based

classiﬁcation of alzheimers disease from whole-brain

anatomical mri. Neuroradiology, 51:73–83.

Pennanen, C., Kivipelto, M., Tuomainen, S., Hartikainen,

P., Hnninen, T., Laakso, M. P., Hallikainen, M., Van-

hanen, M., Nissinen, A., Helkala, E.-L., Vainio, P.,

Vanninen, R., Partanen, K., and Soininen, H. (2004).

Hippocampus and entorhinal cortex in mild cogni-

tive impairment and early ad. Neurobiology of Aging,

25(3):303–310.

Tohka, J., Zijdenbos, A., and Evans, A. (2004). Fast and ro-

bust parameter estimation for statistical partial volume

models in brain mri. NeuroImage, 23(1):84–97.

Vemuri, P., Gunter, J. L., Senjem, M. L., Whitwell, J. L.,

Kantarci, K., Knopman, D. S., Boeve, B. F., Petersen,

R. C., and Jr., C. R. J. (2008). Alzheimer’s disease di-

agnosis in individual subjects using structural mr im-

ages: Validation studies. NeuroImage, 39(3):1186–

1197.

EarlyAlzheimer'sDiseaseProgressionDetectionusingMulti-subnetworksoftheBrain

691