428 IEEE TRANSACTIONS ON NANOBIOSCIENCE, VOL. 16, NO.
6, SEPTEMBER 2017
Alzheimer’s Disease Classification Based on
Individual Hierarchical Networks Constructed
With 3-D Texture Features
Jin Liu, Jianxin Wang, Senior Member, IEEE , Bin Hu, Fang-Xiang Wu, Senior Member, IEEE , and Yi Pan
Abstract — Brain network plays an important role in repre-
senting abnormalities in Alzheimers disease (AD) and mild
cognitive impairment (MCI), which includes MCIc (MCI con-
verted to AD) and MCInc (MCI not converted to AD). In our
previous study, we proposed an AD classification approach
based on individual hierarchical networks constructed with
3D texture features of brain images. However, we only used
edge features of the networks without node features of
the networks. In this paper, we propose a framework of
the combination of multiple kernels to combine edge fea-
tures and node features for AD classification. An evaluation
of the proposed approach has been conducted with MRI
images of 710 subjects (230 health controls (HC), 280 MCI
(including 120 MCIc and 160 MCInc), and 200 AD) from
the Alzheimer’s disease neuroimaging initiative database
by using ten-fold cross validation. Experimental results
show that the proposed method is not only superior to
the existing AD classification methods, but also efficient
and promising for clinical applications for the diagnosis of
AD via MRI images. Furthermore, the results also indicate
that 3D texture could detect the subtle texture differences
between tissues in AD, MCI, and HC, and texture features of
MRI images might be related to the severity of AD cognitive
impairment. These results suggest that 3D texture is a useful
aid in AD diagnosis.
Index Terms — Alzheimer’s disease, 3D texture, individual
hierarchical network, multiple kernel learning, classifica-
tion.
I. I NTRODUCTION
A LZHEIMER’S Disease (AD) is a progressive neurode-
generative disease which is characterized by memory
loss, poor judgment, language deterioration and so forth. Mild
cognitive impairment (MCI), commonly defined as a subtle
Manuscript received December 13, 2016; revised March 15, 2017;
accepted May 16, 2017. Date of publication May 23, 2017; date of current
version September 20, 2017. This work was supported in part by the
National Natural Science Foundation of China under Grant 61232001,
Grant 61420106009, and Grant 61622213 and in part by the Doctoral
Student Independent Exploration Innovative under Project 2015zzts051.
(Corresponding author: Jianxin Wang.)
J. Liu and J. Wang are with the School of Information Science and
Engineering, Central South University, Changsha 410083, China (e-mail:
jxwang@mail.csu.edu.cn).
B. Hu is with the School of Information Science and Engineering,
Lanzhou University, Lanzhou 730000, China.
F.-X. Wu is with the Division of Biomedical Engineering and the
Department of Mechanical Engineering, University of Saskatchewan,
Saskatoon, SK S7N5A9, Canada.
Y. Pan is with the Department of Computer Science, Georgia State
University, Atlanta, GA 30302 USA.
Digital Object Identifier 10.1109/TNB.2017.2707139
1536-1241 © 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
but measurable memory loss, is considered as the transition
stage between health controls (HC) and dementia. Patients
with MCI have been reported more prone to progress to AD
than HC [1], [2]. The symptoms of dementia seriously affect
patient’s daily life. As the major form of dementia, AD affects
26.6 million of people worldwide in 2006 and it is predicted to
affect 1 in 85 people by 2050 [3]. Therefore, precise prediction
and diagnosis of AD, especially at its early warning stage such
as MCI, have become a crucial step to retard or even avoid
dementia.
With the rapid development of brain neuroimaging tech-
niques, such as magnetic resonance imaging (MRI), functional
MRI (fMRI), diffusion tensor imaging (DTI), etc., brain struc-
tural and functional changes can be examined with various
acquired images [4]–[7]. Owning to its easy access in clinical
settings, MRI receives the most attention compared with other
imaging techniques. Studies using MRI have showed that
AD patients are closely related to brain atrophy [8]–[13].
For example, hippocampal atrophy is considered to be the sig-
nature of AD and it is therefore used as a clinical marker [4];
medial temporal lobe (MTL) atrophy as assessed by using MRI
has proven to be an effective clinical aid in the early diagnosis
of AD [8], [9]; the location and degree of gray matter atrophy
can be accurately visualized with MRI, which are used for the
clinical diagnosis of AD [12].
Texture features play an important role in image analysis.
Many approaches have been developed for extracting tex-
ture features from digital images, such as transform-based,
structural-based and statistical-based methods [14]. In fact,
the application of the texture features to medical images
is not new. It was applied to a large variety of patholo-
gies, by using many different approaches. For example,
Georgiadis et al. [15] applied co-occurrence and run-length
matrices to the characterization of different types of brain
tumors; Theocharakis et al. [16] studied multiple sclerosis by
using histogram, co-occurrence and run-length matrix-based
features extracted from fluid-attenuated inversion recovery
MRI images; Zhang et al. [17] applied textural features based
on the polar Stockwell Transform to gadolinium-enhanced
T2 MRI images, also for the study of multiple sclerosis.
In particular, some studies also applied textural features to AD.
Sayeed et al. [18] used a trace transform-based approach on
brain positron emission tomography sinograms to distinguish
patients with AD from HC. Kodama et al. [19] used both
LIU et al.: AD CLASSIFICATION BASED ON INDIVIDUAL HIERARCHICAL NETWORKS CONSTRUCTED WITH 3-D TEXTURE FEATURES
co-occurrence and run-length matrix parameters to differenti-
ate among patients with AD, those with Lewy bodies, and HC.
Zhang et al. [20] applied 3D texture features to improve the
accuracy of AD classification.
Recently, studies focused on constructing an individual
network to accurately identify patients and HC. For example,
Raj et al. [21] constructed a single-subject correlation matrix
with regional cortical thickness, volume, and curvature to
distinguish temporal lobe epilepsy from HC. Wee et al. [22]
and Dai et al. [23] had expressed similar views, which
construct a network based on cortical thickness via exponential
function to make predictions of AD and HC. Our early
work [24] constructed a hierarchical network and extracted
their edge features for AD classification. However, this work
did not use the node features of the hierarchical network to
AD classification.
Based on the above analysis, in this study, we propose
a novel feature representation method based on individual
hierarchical networks for AD classification. To construct a
hierarchical network, firstly, three novel atlases are generated
by three different rules, which combine the existing regions
of Automated Anatomical Labeling (AAL) atlas [25] and then
all regions of four atlases (including AAL atlas and three
novel atlases) are linked between two regions. Subsequently,
we register each experimental subject to the hierarchical net-
work to generate an individual hierarchical network. For each
individual hierarchical network, we extract both node features
and edge features, which are calculated by 3D textural features
including energy (ENE), contrast (CON), inverse difference
moment (IDM), entropy (ENT), difference variance (DVA) and
difference entropy (DEN). To obtain the most discriminative
features for AD classification, we use F-score to rank all
features in descending order. Afterwards, we select the top
ranked features of each feature set to obtain the best clas-
sification accuracy and the corresponding classifier for each
classification by using MKBoost algorithm. Subsequently,
we use a simple weighted combination framework of multiple
kernels to combine six classifiers of node features to generate
the final classifier for node features. Finally, we combine the
classifier of node features with the classifier of edge features
to generate the final classifier based on individual hierarchial
network features.
II. M ATERIALS AND M ETHODS
A. Data Description
A subset of the Alzheimers Disease Neuroimaging Initia-
tive (ADNI) dataset [26] is used to evaluate our proposed
method. This subset includes 710 subjects with T1 -weighted
MRI brain images, which are composed of 200 subjects with
Alzheimer’s Disease (AD), 280 subjects with mild cognitive
impairment (MCI), out of which 120 subjects with MCI
converted to AD within 18 months (MCIc) while 160 sub-
jects with MCI did not convert (MCInc)) and 230 healthy
controls (HC). The demographic information used on the
710 subjects is shown in Table I, where MMSE stands for
Mini Mental State Examination. For more details with the
ADNI database, please see http://adni.loni.usc.edu/.
429
TABLE I
D EMOGRAPHIC I NFORMATION OF 710 S UBJECTS
(F ROM ADNI D ATABASE )
B. Data Preprocessing
A standard pre-processing procedure is used to the
710 T1 -weighted MRI brain images. First, a non-parametric
non-uniform bias correction (N3) algorithm [27] is used
to correct intensity inhomogeneity. Then, skull stripping
is performed by using BET [28]. Afterwards, each brain
image is segmented into three tissues (gray matter (GM),
white matter (WM), and cerebrospinal fluid (CSF)) by using
FAST [29], [30]. Since the GM is most affected by AD and
thus widely investigated in the literature [23], [31], [32], in this
study the GM images are used for feature extraction and
classification. Finally, all GM images are affine aligned by
FLIRT [29], [30].
C. Individual Hierarchical Network Construction
Heterogeneous network construction plays an important
role in biomedical research [24], [33]–[35]. In this article,
we construct a heterogeneous network based on different
texture properties, called individual hierarchical network. The
process of the individual network construction is the same as
our early work [24]. The more details are as follows.
1) The Definition of ROIs: The GM-based Automated
Anatomical Labeling (AAL) atlas previously proposed by
Tzourio-Mazoyer et al. [25] has been widely used in neuro-
science research [31], [36]–[38]. This atlas includes 90 cere-
bral regions (45 in each hemisphere) and 26 cerebellar regions
(9 in each cerebellar hemisphere and 8 in the vermis).
Since cerebellar regions are not related to AD according
to some studies such as [31] and [39] , in this study, the
90 cerebral regions are used for AD classification as shown
in Table II. For more details with AAL atlas, please see
www.cyceron.fr/index.php/en/plateforme-en/freeware.
As some ROIs of the original Brodmann atlas have been
further subdivided into smaller ROIs, e.g., 23a and 23b,
we believe that the integration of ROIs could provide insightful
information. For example, Cingulum, as a ROI, includes
Cingulum_Ant, Cingulum_Mid and Cingulum_Post; limbic
system, as a ROI, includes Cingulum_Ant, Cingulum_Mid,
Cingulum_Post, Hippocampus, ParaHippocampal and Amyg-
dala; and cerebrum, as a ROI, includes 90 cerebral regions of
AAL atlas. Each bigger ROI is generated by a combination of
several existing ROIs with a certain structure and function of
similarity and has a specific function. Therefore, we use AAL
atlas to generate three new atlas with a different number of
ROIs by different rules. More details are as follows:
• The AAL atlas including 90 cerebral ROIs is denoted
as S1.
430 IEEE TRANSACTIONS ON NANOBIOSCIENCE, VOL. 16, NO. 6, SEPTEMBER 2017

TABLE II
C EREBELLAR ROIs B ASED ON AAL ATLAS IN THE P RESENT S TUDY

• Several ROIs in S1 are combined into a ROI if the
first two digits and the last digit of their No. are the
same. For example, Cingulum_Ant_L (4001), Cingu-
lum_Mid_L (4011) and Cingulum_Post_L (4021) are
considered as a ROI. Based on this rule, the atlas includes
54 ROIs and is denoted as S2.
• Similar to the rule of the generation of S2. Some
ROIs in S1 are combined into a ROI if the first and
the last digit of their No. are the same. For example,
Cingulum_Ant_L (4001), Cingulum_Mid_L (4011),
Cingulum_Post_L (4021), Hippocampus_L (4101),
ParaHippocampal_L (4111) and Amygdala_L (4201)
are considered as a ROI. Based on this rule, the atlas
includes 14 ROIs and is denoted as S3.
• The whole brain is regarded as a ROI. The atlas only
includes 1 ROI and is denoted as S4.
2) Construction of the Individual Hierarchical Network: A net-
work is typically defined as G = (V, E), where V is the set
of nodes and E is the set of edges. In this study, we assume
that a node is a cerebral region as defined by the above-
mentioned four atlas (S1, S2, S3 and S4) and an edge is a link
between two ROIs from the above-mentioned four atlas. Based
on this assumption, we can construct an individual hierarchical
network for each subject as shown in Fig.1 [24].
For each node (ROI) of the individual hierarchical net-
work, we extract six 3D texture properties to represent it,
including energy (ENE), contrast (CON), inverse difference
moment (IDM), entropy (ENT), difference variance (DVA) and
difference entropy (DEN). For each node of each 3D texture
property, we firstly calculate the 3D texture property of the
ROI in 0°, 45°, 90° and 135° directions [40] by using
gray level co-occurrence matrices (GLCM) method and then
LIU et al.: AD CLASSIFICATION BASED ON INDIVIDUAL HIERARCHICAL NETWORKS CONSTRUCTED WITH 3-D TEXTURE FEATURES 431

The dimension of edge features is 12, 561((90 × 89 + 54 ×

53 + 14 × 13)/2 + 0 + (90 × 54 + 90 × 14 + 90 × 1) + (54 ×
14 + 54 × 1) + (14 × 1)). These two kinds of features are used
in the subsequent analyses.

D. Feature Selection
In general, some of the features are noisy or redundant,
and even negatively affect classification. Therefore, feature
selection is an essential step before performing classification.
The feature ranking method [41]–[44] has been widely applied
in feature selection. F-score [45] is a simple and quite
effective method which evaluates the discrimination between
two different datasets. In this study, the F-score method is
used to calculate the discrimination of each feature. The
F-score of the i -th feature is calculated as following:
(+) 2 (−) 2
(x i − x i ) + (x i − xi )
F(i ) =
Fig. 1. The individual hierarchical network. 
n+
(+) (+) 2 
n−
(−) (−) 2
1
n + −1 (x k,i − x i ) + 1
n − −1 (x k,i − x i )
k=1 k=1
compute the average texture property of the ROI in four (8)
directions. The six 3D texture properties are calculated as where n + and n − are denoted as the number of positive and
following: (+) (−)
 negative samples respectively; x i , x i and x i are the aver-
ENE = p(i, j )2 (1) age of the i -th feature of the all, positive and negative samples
(+)
i=1 j =1 respectively; x k,i is the i -th feature of the k-th positive sample,
(−)
 and x k,i is the i -th feature of the k-th negative sample. The
CON = (i − j )2 p(i, j ) (2) numerator indicates the discrimination between the positive
i=1 j =1 and negative samples, and the denominator indicates the one
 1 within each of the two samples. The larger the F-score
I DM = p(i, j ) (3) is, the more likely this feature is discriminative. Therefore,
i=1 j =1
1 + (i − j )2 by using this score, the features with good discriminative
 power are selected to AD classification.
E NT = − p(i, j ) log( p(i, j )) (4)
i=1 j =1
E. Classification and Evaluation

2
DV A = (l − u x−y ) px−y (l) (5) Multiple Kernel Learning (MKL) is a promising family
l=0 of machine learning algorithms using multiple kernel func-
 tions for various challenging data mining tasks [46]–[49].
DE N = − px−y (l) log( p x−y (l)) (6) In this study, we firstly use MKBoost algorithm proposed by
l=0
Xia and Hoi [49] to obtain the best classification accuracy and
  p(i, j ) is the value of GLCM, px−y (l) =
where the corresponding classifier for each feature set, such as ENE
p(i, j ), u x−y is the average value of px−y (l), where features, CON features, IDM features, ENT features, DVA
i=1 j =1
l = |i − j|. For more details with GLCM, please see [40]. features, DEN features and edge features. For each feature
Based on the above analysis, a 6-dimensional vector is set, we create a set of 13 base kernels, i.e.,
• Gaussian kernels with 10 different
generated to represent each node of the individual hierarchical  
network. In this study, to obtain the weight of each edge widths ( 2−4 , 2−3 , . . . , 24 , 25 ) on all features.
of the individual hierarchical network, we calculate Pearson • Polynomial kernels of degree 1 to 3 on all features.
correlation coefficients (PCC) between two nodes below: In our study, we set the total number of boosting trials T
E(ab) − E(a)E(b) to 100, the boosting sampling ratio to 0.2, and the sam-
PCC(a, b) =   (7) pling decay factor to 2−5 for MKBoost-S2. Support vector
E(a2 ) − E 2 (a) E(b2 ) − E 2 (b) machine (SVM) is adopted to perform classification and
where a, b are node vectors, E(·) is the average value of a implemented by the LIBSVM library [50], with a default value
vector. for the parameter C (= 1).
At this point, we can obtain node features and edge fea- Then, we use a simple framework of the combination
tures for each individual hierarchical network. Node features of multiple kernels (CK) to generate a final classifier as
include ENE features, CON features, IDM features, ENT shown in Fig.2. Suppose there are N kernel functions,
features, DVA features and DEN features, which make up a {k1 (·, ·), . . . , k N (·, ·)}, corresponding to N feature sets, X =
159(90 + 54 + 14 + 1) dimensional vector for each subject. {X (1) , . . . , X (N) }, where matrix X (n) represents samples using
432
Fig. 2. A simple framework of the combination of multiple kernels.
the n-th set of features. The combination kernel between
samples x i and x j can be computed by
k(x i , x j ) = f β (k1 (x i(1), x (1) (N) (N)
j ), . . . , k N (x i , x j ))
where f β is either a linear or non-linear function with parame-
ter β. In the simplest type with Equation (9), it is a weighted
linear combination of kernel matrices:

N
k(X, X) = βn kn (X (n) , X (n) )
n=1
where β = [β1 , . . . , β N ] can be determined either prior to
learning or by the learning procedure. In this study, we use
the best classification accuracies for each feature set as the
weights.
The evaluation of our proposed method is conducted on four
types in AD diagnosis:
• T1: AD/HC classification;
• T2: AD/MCI classification;
• T3: MCI/HC classification;
• T4: MCIc/MCInc classification;
T1 is relatively simple, T2 and T3 are more difficult, T4 is the
most difficult and draws relatively less attention in previous
studies. However, to explore the progression of AD, the latter
three types (T2, T3 and T4) are more important than the first
type (T1).
In this study, a 10-fold cross validation strategy is applied
to evaluate the classification performance of our proposed
method. To avoid over-fitting, firstly, all subjects for each
classification are randomly equally partitioned into 10 subsets,
which are denoted as {S1 , S2 , · · · , S10 } and S1 is select as
validation set. Then, {S2 , · · · , S10 } as a whole data set is
further randomly equally partitioned into 10 subsets, and
one subset is randomly selected as testing set and the other
9 subsets are used to train a classifier by using 10-fold cross-
validation. We select different numbers of features to perform
classification and can obtain corresponding mean classifica-
tion accuracies by using 10-fold cross-validation based data
set {S2 , · · · , S10 }. Subsequently, we use the corresponding
feature set of the validation set S1 to evaluate performance of
each classification including classification accuracy as show
in Fig.4. In our work, we select the feature set which can
obtain the best classification accuracy as the next classification
features. For new subjects, a corresponding feature set is used
directly to perform classification of each subject. To avoid any
bias introduced by randomly partitioning the dataset in the
IEEE TRANSACTIONS ON NANOBIOSCIENCE, VOL. 16, NO. 6, SEPTEMBER 2017
cross validation, we repeat this process 100 times. To quan-
tify the classification performance of T1, T2, T3 and T4,
accuracy (ACC), sensitivity (SEN), and specificity (SPE) are
used and calculated as follows:
TP +TN
ACC = (11)
T P + T N + FP + FN
TP
SE N = (12)
T P + FN
TN
SPE = (13)
T N + FP
where TP, FP, TN and FN are the number of true positive
samples, false positive samples, true negative samples and
false negative samples respectively. The AUC (area under
receiver operating characteristic (ROC) curve) value is an
(9) important index to measure the overall performance of the
classification method. The higher the AUC value, the better
the performance of the classification method is. Therefore,
the AUC value is also used in our proposed method.
III. R ESULTS AND D ISCUSSION
(10)
A. Overall Classification Framework
An overview of AD classification framework based on indi-
vidual hierarchical networks is summarized in Fig.3. We firstly
construct an individual hierarchical network for each subject
and then extract node features and edge features. Subsequently,
we use MKBoost algorithm to obtain the best classifica-
tion accuracy and the corresponding classifier (ENE-MK,
CON-MK, IDM-MK, ENT-MK, DVA-MK, DEN-MK and
E-MK) for each of node features and edge features. After-
wards, we use the above mentioned framework of the
combination of multiple kernels to generate a new classi-
fier (N-CK) based on six node features. Finally, we combine
N-CK classifier and E-MK classifier to generate a final clas-
sifier to complete the final AD classification.
B. Comparison of Different Features
Firstly, we only consider node features including ENE
features, CON features, IDM features, ENT features, DVA
features and DEN features. For each node feature set, we use
MKBoost method to obtain the best classification accuracy (as
shown in Fig.4 and Table III) and the corresponding classi-
fier. Based on the best classification accuracy of each node
feature set, we use them as the weights to combine the six
corresponding classifiers as the final classifier (N-CK) of node
features. The results of N-CK in the above four types of in
AD diagnosis are shown in Table IV. As shown in Table IV
and Fig.4, we can easily find that the classification accuracy of
N-CK is superior to any node feature set. The results demon-
strate that the different features possess complementary infor-
mation, and the combination of these features could therefore
improve AD classification accuracy compared to the separate
features by using the combination of multiple kernels for
different features.
Secondly, we only consider edge features. Similar to the
classification of node features, we use MKBoost method to
obtain the best classification accuracy (as shown in Fig.5 and
LIU et al.: AD CLASSIFICATION BASED ON INDIVIDUAL HIERARCHICAL NETWORKS CONSTRUCTED WITH 3-D TEXTURE FEATURES 433

Fig. 3. A flowchart for novel individual hierarchical network-based feature representation for Alzheimer’s disease classification.

Fig. 4. Classification accuracy for four types. The features are ranked according to F -score in descending order.

Table IV) and the corresponding classifier. This work is the
same as [24].
We assume that the combination of node features and
edge features may improve the classification accuracy of AD.
Therefore, the best classification accuracies of node fea-
tures and edge features are used as the weights to com-
bine their classifiers (namely N-CK and E-MK) as the final
classifier (NECK) for the above four types of AD clas-
sification. The results of NECK are shown in Table IV.
As shown in Table IV, we can easily find that the classification
performance (AUC) of NECK is superior to both N-CK
and E-MK.
Based on the above analysis, we can demonstrate that our
proposed feature representation based on individual hierar-
chical networks is efficient for improving AD classification
performance.
C. Comparison With Existing Classification Methods
In order to validate the superiority of our proposed method,
we compare the six existing methods for AD classification as
follows:
• Chupin et al. [51] proposed a segmentation method to
segment hippocampus and used the obtained hippocampal
volumes for AD classification.
434
TABLE III
T HE B EST C LASSIFICATION ACCURACY FOR N ODE F EATURES
Fig. 5. Classification accuracy for four types. The features are ranked
according to F -score in descending order.
• Ahmed et al. [52] proposed to use hippocampus visual
features and cerebrospinal fluid (CSF) volume for AD
classification.
• Dai et al. [23] proposed to construct an individual net-
work based on mean cortical thickness and used the
network edge features for AD classification.
• Suk et al. [53] proposed to use DBM (Deep Boltzmann
Machine) to find a latent hierarchical feature representa-
tion from a 3D patch for AD classification.
• Khedher et al. [54] proposed to use GM and WM for
AD classification.
IEEE TRANSACTIONS ON NANOBIOSCIENCE, VOL. 16, NO. 6, SEPTEMBER 2017
TABLE IV
T HE C LASSIFICATION P ERFORMANCE FOR D IFFERENT
N ETWORK F EATURES
TABLE V
C OMPARISON TO E XISTING W ORKS FOR T1
TABLE VI
C OMPARISON TO E XISTING W ORKS FOR T2
TABLE VII
C OMPARISON TO E XISTING W ORKS FOR T3
• Liu et al. [24] proposed to use edge features of the
hierarchial networks for AD classification.
For fair comparisons, the experimental images of these meth-
ods are the same as those of our proposed method. Although
Suk et al. [53] used two modal images: MRI and positron
emission tomography (PET), only the results based on MRI
images are reported in this study. The comparative results for
T1, T2, T3 and T4 are shown in Tables V, VI, VII and VIII
respectively.
As can be seen from Tables V to VIII, our proposed method
substantially outperforms the other six existing methods in
LIU et al.: AD CLASSIFICATION BASED ON INDIVIDUAL HIERARCHICAL NETWORKS CONSTRUCTED WITH 3-D TEXTURE FEATURES
TABLE VIII
C OMPARISON TO E XISTING W ORKS FOR T4
terms of ACC, SEN and SPE. For T1, although the SEN
of Liu et al. [24] method (95.03%) is better than our pro-
posed method (92.49%), it is only about 2 percentage lower.
For T2, although the SPE of Suk et al. [53] method (90.65%)
is better than our proposed method (88.82%), it is only
about 1 percentage lower. For T3, although only the ACC
of our proposed method is the best, the SEN of our pro-
posed method (90.74%) is about 12 percentage higher than
Dai et al. [23] method (78.51%) and the SPE of our pro-
posed method (84.83%) is about 27 percentage higher than
Suk et al. [53] method (57.23%). For T4, although the SPE
of our proposed method (72.24%) is about 16 percentage
lower than Suk et al. [53] method (88.49%), the SEN of our
proposed method (76.13%) is about 36 percentage higher than
Suk et al. [53] method (40.55%).
As shown in Tables V, VI, VII and VIII, the AUC values
of our proposed method are obviously superior to the other
six existing methods for the above four classifications. The
results demonstrate that our proposed method is efficient and
has strong robustness.
IV. C ONCLUSION
We have proposed a novel feature representation method
based on individual hierarchial networks for AD classification.
For each individual hierarchial network, we extract both node
features and edge features based on 3D textural features.
To obtain the most discriminative features for each classi-
fication, we use F-score to rank all features in descend-
ing order. The top ranked features from each feature set
(ENE features, CON features, IDM features, ENT features,
DVA features, DEN features and edge features) and MKBoost
method are used to obtain the best classification accuracy
and the corresponding classifier for each classification. Subse-
quently, we use a simple weighted combination framework of
multiple kernels to combine six classifiers of node features
to generate the final classifier for node features. Finally,
we combine the classifier of node features with the classifier of
edge features to generate the final classifier based on individual
hierarchial network features. Experimental results on the MRI
data from ADNI database have shown that our proposed
method is efficient and promising for clinical applications for
the diagnosis of AD.
In the current article, we only use a simple method (i.e.,
F-score) to select the classification features. Other feature
selection methods, such as considering strong correlation
between features with the F-score, can also be incorporated
into our framework, which will be our future work.
435
R EFERENCES
[1] R. C. Petersen et al., “Current concepts in mild cognitive impairment,”
Arch. Neurol., vol. 58, no. 12, pp. 1985–1992, 2001.
[2] M. Grundman et al., “Mild cognitive impairment can be distinguished
from Alzheimer disease and normal aging for clinical trials,” Arch. Neu-
rol., vol. 61, no. 1, pp. 59–66, 2004.
[3] R. Brookmeyer, E. Johnson, K. Ziegler-Graham, and H. M. Arrighi,
“Forecasting the global burden of Alzheimer’s disease,” Alzheimer’s
Dementia, vol. 3, no. 3, pp. 186–191, Jul. 2007.
[4] N. Schuff et al., “MRI of hippocampal volume loss in early Alzheimer’s
disease in relation to ApoE genotype and biomarkers,” Brain, vol. 132,
no. 4, pp. 1067–1077, 2009.
[5] X. Wu et al., “The receiver operational characteristic for binary clas-
sification with multiple indices and its application to the neuroimaging
study of Alzheimer’s disease,” IEEE/ACM Trans. Comput. Biol. Bioinf.,
vol. 10, no. 1, pp. 173–180, Jan./Feb. 2013.
[6] J. Liu, M. Li, J. Wang, F. Wu, T. Liu, and Y. Pan, “A survey of
MRI-based brain tumor segmentation methods,” Tsinghua Sci. Technol.,
vol. 19, no. 6, pp. 578–595, Dec. 2014.
[7] Y. Xiao et al., “Patch-based label fusion segmentation of brainstem struc-
tures with dual-contrast MRI for Parkinson’s disease,” Int. J. Comput.
Assist. Radiol. Surgery, vol. 10, no. 7, pp. 1029–1041, Jul. 2015.
[8] C. R. Jack et al., “Medial temporal atrophy on MRI in normal aging and
very mild Alzheimer’s disease,” Neurology, vol. 49, no. 3, pp. 786–794,
1997.
[9] P. J. Visser, F. R. J. Verhey, P. A. M. Hofman, P. Scheltens, and J. Jolles,
“Medial temporal lobe atrophy predicts Alzheimer’s disease in patients
with minor cognitive impairment,” J. Neurol., Neurosurgery Psychiatry,
vol. 72, no. 4, pp. 491–497, 2002.
[10] C. DeCarli et al., “Qualitative estimates of medial temporal atrophy as a
predictor of progression from mild cognitive impairment to dementia,”
Arch. Neurol., vol. 64, no. 1, pp. 108–115, 2007.
[11] C. Jack et al., “Atrophy rates accelerate in amnestic mild cognitive
impairment,” Neurology, vol. 70, no. 19, pp. 1740–1752, 2008.
[12] G. B. Frisoni, N. C. Fox, C. R. Jack, P. Scheltens, and P. M. Thompson,
“The clinical use of structural MRI in Alzheimer disease,” Nature Rev.
Neurol., vol. 6, no. 2, pp. 67–77, 2010.
[13] L. Clerx et al., “Measurements of medial temporal lobe atrophy for
prediction of Alzheimer’s disease in subjects with mild cognitive impair-
ment,” Neurobiol. Aging, vol. 34, no. 8, pp. 2003–2013, Aug. 2013.
[14] M. H. Bharati, J. J. Liu, and J. F. MacGregor, “Image texture analysis:
Methods and comparisons,” Chemometrics Intell. Lab. Syst., vol. 72,
no. 1, pp. 57–71, Jun. 2004.
[15] P. Georgiadis et al., “Enhancing the discrimination accuracy between
metastases, gliomas and meningiomas on brain MRI by volumetric
textural features and ensemble pattern recognition methods,” Magn.
Reson. Imag., vol. 27, no. 1, pp. 120–130, Jan. 2009.
[16] P. Theocharakis et al., “Pattern recognition system for the discrimination
of multiple sclerosis from cerebral microangiopathy lesions based on
texture analysis of magnetic resonance images,” Magn. Reson. Imag.,
vol. 27, no. 3, pp. 417–422, Apr. 2009.
[17] J. Zhang, L. Tong, L. Wang, and N. Li, “Texture analysis of multiple
sclerosis: A comparative study,” Magn. Reson. Imag., vol. 26, no. 8,
pp. 1160–1166, Oct. 2008.
[18] A. Sayeed, M. Petrou, N. Spyrou, A. Kadyrov, and T. Spinks, “Diag-
nostic features of Alzheimer’s disease extracted from PET sinograms,”
Phys. Med. Biol., vol. 47, no. 1, p. 137, 2001.
[19] N. Kodama, Y. Kawase, and K. Okamoto, “Application of texture analy-
sis to differentiation of dementia with lewy bodies from Alzheimer’s
disease on magnetic resonance images,” in World Congress on Medical
Physics and Biomedical Engineering. Berlin, Germany: Springer, 2007,
pp. 1444–1446.
[20] J. Zhang, C. Yu, G. Jiang, W. Liu, and L. Tong, “3D texture analysis
on MRI images of Alzheimer’s disease,” Brain Imag. Behavior, vol. 6,
no. 1, pp. 61–69, Mar. 2012.
[21] A. Raj, S. G. Müeller, K. Young, K. D. Laxer, and M. Weiner, “Network-
level analysis of cortical thickness of the epileptic brain,” NeuroImage,
vol. 52, no. 4, pp. 1302–1313, Oct. 2010.
[22] C.-Y. Wee, P.-T. Yap, and D. Shen, “Prediction of Alzheimer’s disease
and mild cognitive impairment using cortical morphological patterns,”
Hum. Brain Mapp., vol. 34, no. 12, pp. 3411–3425, Dec. 2013.
[23] D. Dai, H. He, J. T. Vogelstein, and Z. Hou, “Accurate prediction of ad
patients using cortical thickness networks,” Mach. Vis. Appl., vol. 24,
no. 7, pp. 1445–1457, 2013.
436
[24] J. Liu, M. Li, W. Lan, F.-X. Wu, Y. Pan, and J. Wang, “Classi-
fication of Alzheimer’s disease using whole brain hierarchical net-
work,” IEEE/ACM Trans. Comput. Biol. Bioinf., to be published, doi:
10.1109/TCBB.2016.2635144.
[25] N. Tzourio-Mazoyer et al., “Automated anatomical labeling of activa-
tions in SPM using a macroscopic anatomical parcellation of the MNI
MRI single-subject brain,” NeuroImage, vol. 15, no. 1, pp. 273–289,
2002.
[26] M. C. Carrillo, L. J. Bain, G. B. Frisoni, and M. W. Weiner, “Worldwide
Alzheimer’s disease neuroimaging initiative,” Alzheimer’s Dementia,
vol. 8, no. 4, pp. 337–342, Jul. 2012.
[27] J. G. Sled, A. P. Zijdenbos, and A. C. Evans, “A nonparamet-
ric method for automatic correction of intensity nonuniformity in
MRI data,” IEEE Trans. Med. Imag., vol. 17, no. 1, pp. 87–97,
Feb. 1998.
[28] S. M. Smith, “Fast robust automated brain extraction,” Human Brain
Mapping, vol. 17, no. 3, pp. 143–155, 2002.
[29] M. Jenkinson and S. Smith, “A global optimisation method for robust
affine registration of brain images,” Med. Image Anal., vol. 5, no. 2,
pp. 143–156, Jun. 2001.
[30] M. Jenkinson, P. Bannister, M. Brady, and S. Smith, “Improved opti-
mization for the robust and accurate linear registration and motion
correction of brain images,” NeuroImage, vol. 17, no. 2, pp. 825–841,
Oct. 2002.
[31] B. Magnin et al., “Support vector machine-based classification of
Alzheimer’s disease from whole-brain anatomical MRI,” Neuroradiol-
ogy, vol. 51, no. 2, pp. 73–83, Feb. 2009.
[32] D. Zhang, Y. Wang, L. Zhou, H. Yuan, and D. Shen, “Multimodal
classification of Alzheimer’s disease and mild cognitive impairment,”
NeuroImage, vol. 55, no. 3, pp. 856–867, Apr. 2011.
[33] X. Peng, J. Wang, W. Peng, F.-X. Wu, and Y. Pan, “Protein-
protein interactions: Detection, reliability assessment and applica-
tions,” Briefings Bioinform., p. bbw066, 2016. [Online]. Available:
https://doi.org/10.1093/bib/bbw066
[34] H. Luo et al., “Drug repositioning based on comprehensive similarity
measures and bi-random walk algorithm,” Bioinformatics, vol. 32,
no. 17, pp. 2664–2671, 2016.
[35] M. Li, P. Ni, X. Chen, J. Wang, F. Wu, and Y. Pan, “Construction
of refined protein interaction network for predicting essential pro-
teins,” IEEE/ACM Trans. Comput. Biol. Bioinf., to be published, doi:
10.1109/TCBB.2017.2665482.
[36] Y. Liu et al., “Disrupted small-world networks in schizophrenia,” Brain,
vol. 131, no. 4, pp. 945–961, 2008.
[37] R. Vandenberghe et al., “18 F-flutemetamol amyloid imaging in
Alzheimer disease and mild cognitive impairment: A phase 2 trial,” Ann.
Neurol., vol. 68, no. 3, pp. 319–329, Sep. 2010.
[38] J. Wang et al., “Disrupted functional brain connectome in individuals
at risk for Alzheimer’s disease,” Biol. Psychiatry, vol. 73, no. 5,
pp. 472–481, Mar. 2013.
[39] J. Zhang, L. Zhou, L. Wang, and W. Li, “Functional brain network
classification with compact representation of SICE matrices,” IEEE
Trans. Biomed. Eng., vol. 62, no. 6, pp. 1623–1634, Jun. 2015.
[40] R. M. Haralick, K. Shanmugam, and I. H. Dinstein, “Textural fea-
tures for image classification,” IEEE Trans. Syst., Man, Cybern., Syst.,
vol. SMC-3, no. 6, pp. 610–621, 1973.
[41] C. Chu, A.-L. Hsu, K.-H. Chou, P. Bandettini, and C. Lin, “Does feature
selection improve classification accuracy? Impact of sample size and
feature selection on classification using anatomical magnetic resonance
images,” NeuroImage, vol. 60, no. 1, pp. 59–70, Mar. 2012.
[42] F. Liu, H.-I. Suk, C.-Y. Wee, H. Chen, and D. Shen, “High-order graph
matching based feature selection for Alzheimer’s disease identifica-
tion,” in Medical Image Computing and Computer-Assisted Intervention.
Springer, 2013, pp. 311–318.
[43] J. Gao et al., “A novel approach for lie detection based on f-score and
extreme learning machine,” PLoS ONE, vol. 8, no. 6, p. e64704, 2013.
[44] F. Liu et al., “Multivariate classification of social anxiety disorder using
whole brain functional connectivity,” Brain Struct. Funct., vol. 220,
no. 1, pp. 101–115, Jan. 2015.
[45] Y.-W. Chen and C.-J. Lin, “Combining svms with various feature
selection strategies,” in Feature Extraction. Berlin, Germany: Springer,
2006, pp. 315–324.
[46] S. Sonnenburg, G. Rätsch, C. Schäfer, and B. Schölkopf, “Large scale
multiple kernel learning,” J. Mach. Learn. Res., vol. 7, pp. 1531–1565,
Jul. 2006.
[47] A. Rakotomamonjy, F. Bach, S. Canu, and Y. Grandvalet, “SimpleMKL,”
J. Mach. Learn. Res., vol. 9, pp. 2491–2521, Nov. 2008.
IEEE TRANSACTIONS ON NANOBIOSCIENCE, VOL. 16, NO. 6, SEPTEMBER 2017
[48] Z. Xu, R. Jin, I. King, and M. Lyu, “An extended level method for
efficient multiple kernel learning,” in Proc. Adv. Neural Inf. Process.
Syst., 2009, pp. 1825–1832.
[49] H. Xia and S. C. H. Hoi, “MKBoost: A framework of multiple kernel
boosting,” IEEE Trans. Knowl. Data Eng., vol. 25, no. 7, pp. 1574–1586,
Jul. 2013.
[50] C.-C. Chang and C.-J. Lin, “LIBSVM: A library for support vec-
tor machines,” ACM Trans. Intell. Syst. Technol., vol. 2, no. 3,
pp. 27:1–27:27, 2011. [Online]. Available: http://www.csie.ntu.edu.
tw/~cjlin/libsvm
[51] M. Chupin et al., “Fully automatic hippocampus segmentation and
classification in Alzheimer’s disease and mild cognitive impairment
applied on data from ADNI,” Hippocampus, vol. 19, no. 6, pp. 579–587,
2009.
[52] O. B. Ahmed et al., “Classification of Alzheimer’s disease subjects
from MRI using hippocampal visual features,” Multimedia Tools Appl.,
vol. 74, no. 4, pp. 1249–1266, Feb. 2015.
[53] H.-I. Suk, S.-W. Lee, and D. Shen, “Hierarchical feature representation
and multimodal fusion with deep learning for AD/MCI diagnosis,”
NeuroImage, vol. 101, pp. 569–582, Nov. 2014.
[54] L. Khedher, J. Ramírez, J. M. Górriz, A. Brahim, and F. Segovia, “Early
diagnosis of Alzheimer?s disease based on partial least squares, principal
component analysis and support vector machine using segmented MRI
images,” Neurocomputing, vol. 151, pp. 139–150, Mar. 2015.
Jin Liu is currently pursuing the Ph.D. degree
with the School of Information Science and Engi-
neering, Central South University, Changsha,
China. His current research interests include
medical image analysis, bioinformatics, and
machine learning.
Jianxin Wang (M’01–SM’14) received the
B.Eng. and M.Eng. degrees in computer engi-
neering and the Ph.D. degree in computer sci-
ence from Central South University, Changsha,
China, in 1992, 1996, and 2001, respectively. He
is currently the Chair and a Professor with the
Department of Computer Science, Central South
University. His current research interests include
algorithm analysis and optimization, parameter-
ized algorithm, bioinformatics, and computer net-
work.
Bin Hu received the M.Sc. degree in computer
science from the Beijing University of Technology
and the Ph.D. degree in computer science from
the Institute of Computing Technology, Chinese
Academy of Science. He is currently a Professor
and the Dean of School of Information Science
and Engineering, Lanzhou University. He is the
Leader of the Intelligent Contextual Computing
Group, Pervasive Computing Center, the Reader
with Birmingham City University, and a Visiting
Professor with the Beijing University of Posts and
Telecommunications, China, and ETH Zürich, Switzerland. His research
interests include pervasive computing, psycho-physiological computing,
cooperative work, and semantic Web.
LIU et al.: AD CLASSIFICATION BASED ON INDIVIDUAL HIERARCHICAL NETWORKS CONSTRUCTED WITH 3-D TEXTURE FEATURES
Fang-Xiang Wu (M’06–SM’11) received the
B.Sc. degree and the M.Sc. degree in applied
mathematics, from the Dalian University of Tech-
nology, Dalian, China, in 1990 and 1993, respec-
tively, the first Ph.D. degree in control theory and
its applications from Northwestern Polytechnical
University, Xi’an, China, in 1998, and the second
Ph.D. degree in biomedical engineering from the
University of Saskatchewan (U of S), Saskatoon,
Canada, in 2004. From 2004 to 2005, he was
a Post-Doctoral Fellow with the Laval University
Medical Research Center, Sainte-Foy, QC, Canada. He is currently a Pro-
fessor with the Division of Biomedical Engineering and the Department
of Mechanical Engineering, U of S. He has authored over 260 technical
papers in refereed journals and conference proceedings. His current
research interests include computational and systems biology, genomic
and proteomic data analysis, biological system identification and para-
meter estimation, and the applications of control theory to biological
systems. He is serving as the Editorial Board Member of three inter-
national journals, the Guest Editor of several international journals,
and as the Program Committee Chair or a member of several interna-
tional conferences. He has also reviewed papers for many international
journals.
437
Yi Pan received the B.Eng. and M.Eng. degrees
in computer engineering from Tsinghua Univer-
sity, China, in 1982 and 1984, respectively, and
the Ph.D. degree in computer science from the
University of Pittsburgh, USA, in 1991. His profile
has been featured as a distinguished alumnus in
both Tsinghua Alumni Newsletter and University
of Pittsburgh CS Alumni Newsletter. He joined
Georgia State University, USA, in 2000 and was
promoted to a Full Professor in 2004, named a
Distinguished University Professor in 2013 and
designated a Regents’ Professor (the highest recognition given to a
Faculty Member by the University System of Georgia) in 2015. He served
as the Chair of the Computer Science Department from 2005 to 2013.
He is currently a Regents’ Professor of Computer Science and an Interim
Associate Dean and the Chair of Biology with Georgia State University.
He is also a Visiting Changjiang Chair Professor with Central South Uni-
versity, China. He has authored over 330 papers, including over 180 SCI
journal papers and 60 IEEE/ACM Transactions papers. In addition, he
has edited/authored 40 books. His research interests include parallel and
cloud computing, wireless networks, and bioinformatics. He is a recipient
of many awards, including the IEEE Transactions Best Paper Award,
four other international conference or journal Best Paper Awards, four
IBM Faculty Awards, two JSPS Sen-ior Invitation Fellowships, the IEEE
BIBE Outstanding Achievement Award, the NSF Research Opportunity
Award, and the AFOSR Summer Faculty Research Fellowship. His
work has been cited over 6500 times. He has served as an Editor-in-
Chief or the Editorial Board Member for 15 journals, including seven
IEEE Transactions. He has organized many international conferences
and delivered keynote speeches at over 50 international conferences
around the world.