Distributed and Parallel Databases

https://doi.org/10.1007/s10619-021-07345-y

Detection of Alzheimer’s disease using grey wolf

optimization based clustering algorithm and deep neural
network from magnetic resonance images

Halebeedu Subbaraya Suresha1 ·

Srirangapatna Sampathkumaran Parthasarathy2

Accepted: 11 June 2021

© The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature
2021

Abstract
The automated magnetic resonance imaging (MRI) processing techniques are gain-
ing more importance in Alzheimer disease (AD) recognition, because it effectively
diagnosis the pathology of the brain. Currently, computer aided diagnosis based on
image analysis is an emerging tool to support AD diagnosis. In this research study,
a new system is developed for enhancing the performance of AD recognition. Ini-
tially, the brain images were acquired from three online datasets and one real-time
dataset such as AD Neuroimaging Initiative (ADNI), Minimal Interval Resonance
Imaging in AD (MIRIAD), and Open Access Series of Imaging Studies (OASIS)
and National Institute of Mental Health and Neuro Sciences (NIMHANS). Then,
adaptive histogram equalization (AHE) and grey wolf optimization based cluster-
ing algorithm (GWOCA) were applied for denoising and segmenting the brain tis-
sues; grey matter (GM), cerebro-spinal fluid (CSF), and white matter (WM) from
the acquired images. After segmentation, the feature extraction was performed by
utilizing dual tree complex wavelet transform (DTCWT), local ternary pattern (LTP)
and Tamura features to extract the feature vectors from the segmented brain tissues.
Then, ReliefF methodology was used to select the active features from the extracted
feature vectors. Finally, the selected active feature values were classified into three
classes [AD, normal and mild cognitive impairment (MCI)] utilizing deep neural
network (DNN) classifier. From the simulation result, it is clear that the proposed
framework achieved good performance in disease classification and almost showed
2.2–6% enhancement in accuracy of all four datasets.

Keywords Alzheimer disease recognition and classification · Deep neural network ·

Grey wolf optimization based clustering algorithm · Histogram equalization ·
ReliefF algorithm

* Halebeedu Subbaraya Suresha

srisuri75@gmail.com; suresha.phd2018@gmail.com
Extended author information available on the last page of the article

13
Vol.:(0123456789)
 Distributed and Parallel Databases

Abbreviations
AHE Adaptive histogram equalization
AD Alzheimer disease
ADNI Alzheimer Disease Neuroimaging Initiative
CSF Cerebro-spinal fluid
CNN Convolutional neural network
DM2L Deep multitask multichannel learning
DNN Deep neural network
GLCM Gray level co-occurrence matrix
GM Grey matter
GWOCA Grey wolf optimization based clustering algorithm
HMM Hidden Markov models
LBP Local binary pattern
LTP Local ternary pattern
LSTM Long short-term memory
LR Lucy Richardson
MRI Magnetic resonance imaging
MCI Mild cognitive impairment
MIRIAD Minimal Interval Resonance Imaging in AD
NIMHANS National Institute of Mental Health and Neuro Sciences
NN Neural network
OASIS Open Access Series of Imaging Studies
PDF Probability distribution function
DTCWT​ Tree complex wavelet transform
WM White matter

List of symbols
A Partitioned instances
A and C Coefficient vectors
c Centre pixel
C Class of cluster
C∗ Optimal clusters
D Distance
f (X, C) Statistical function
g(Z j ) Pooling
g0 (n) and g1 (n) Low and high pass filter for second wavelet tree
h0 (n) and h1 (n) Low and high pass filter for first wavelet tree
Hj Nearest hit instances
k Cluster center
m Hidden nodes
Mj Nearest miss instances
N Number of population
Of Optimized value
p Neighbouring pixel
̂pj Sparsity penalty
⃗r1 and ⃗r2 Random values

13
Distributed and Parallel Databases

ri Randomly picked the instances

t Current iteration
Tc Threshold constant
wij Model parameter
W[A] Quality estimation
x1 , x2 , … , xn Input data instances
xi Auto encoder hidden unit
̂
Xprey Position of prey
Xwolf Grey wolf position

Greek letters
𝛼 Alpha solution
𝛽 Beta solution
ΔH Horizontal convolved grey-scale images
ΔV Vertical convolved grey-scale images
𝛿 Delta solution
𝜃 Directionality
𝜆 Weight delay
𝜇4 Fourth moment of mean
𝜎 Variance
Ψh (t) First wavelet tree
Ψg (t) Second wavelet tree
Ω Omega solution

1 Introduction

In recent decades, AD is a serious neuro-degenerative disorder that leads to abstract

thinking, disorientation, memory loss, etc. [1]. Hence, the only way to control AD
is the early detection that helps to slow down the neuro-degeneration [2, 3]. The
biomarkers based on neuro-imaging modalities such as ultrasound, diffusion ten-
sor imaging, magnetic resonance imaging (MRI), functional MRI, and computed
tomography measures the metabolic burden with dissimilar radioactive tracers that
discriminates AD with promising results [4, 5]. Presently, machine and deep learn-
ing methods have gained great interest among the researchers to analyse neuro-
imaging techniques [6]. The machine and deep learning methods lessen the issues
of conventional approaches by using available data information [7–9]. In recent peri-
ods, numerous methodologies are developed by the researchers for AD recognition
such as support vector machine [10–12], neural network [13], sparse representation
[14], Random forest [15], linear discriminant analysis [16], artificial neural networks
[17], etc. Due to the nonlinear nature of extracted features, the prior approaches are
not effective in classifying the brain images. In this study, a new framework is pro-
posed for improving the performance of AD recognition and classification.
In this work, the input brain scans were collected from four databases; ADNI,
OASIS, MIRIAD, and NIMHANS. After collecting the brain scans, AHE was uti-
lized to enhance the contrast of the image by redistributing the lightness or pixel

13
 Distributed and Parallel Databases

value of the image. Then, region growing was utilized to perform skull stripping
that significantly avoids the misclassification of brain tissues and also reduces the
system complexity. After skull removal, segmentation was done utilizing GWOCA
for segmenting the brain tissues such as GM, CSF, and WM. The major two benefits
of GWOCA are, (i) easy implementation, and (ii) high speed performance. After
segmenting the brain tissues, feature extraction was performed by utilizing LTP,
DTCWT and Tamura features to extract the feature vectors. The descriptor level
features significantly diminish the semantic space between the local and global fea-
tures. Then, ReliefF method was used to reduce the dimension of extracted features
that helps in better classification. Finally, the obtained optimal features were fed to
DNN to classify AD, MCI, and normal subjects. The DNN classifier was the best
choice for classification, if the undertaken data was unstructured in nature. Finally,
the proposed framework performance was related with prior research works in light
of f score, Matthews correlation coefficient (MCC), false omission rate (FOR), miss
rate, accuracy and error rate for evaluating the efficiency of the proposed model. The
contribution of the proposed framework is described below,

• Developed an effective pre-processing technique for enhancing the brain image

quality to make the detection of AD easier.
• Developed a new optimization based clustering algorithm (GWOCA) for
improving the segmentation performance.
• Developed a deep learning based classification methodology for assisting the cli-
nicians in the early identification of AD.
• The ReliefF feature selection methods selects 30% of features from the total fea-
tures and apply to the DNN for classification. This helps to reduce the computa-
tion time of the classification in the proposed method.

This research paper is prearranged as follows. Some existing research papers in

AD identification and classification are reviewed in the Sect. 2. In Sect. 3, detailed
explanation of the proposed framework is given. The quantitative and comparative
study of proposed framework is denoted in the Sect. 4. Section 5 explains the con-
clusion of the proposed work.

2 Literature survey

In recent decades, many research works were carried out in AD detection and classi-
fication. In this segment, some existing research papers were investigated by means
of database, methodology, advantages, and limitations.
Sampath and Indumathi [18] developed a new automated framework to recog-
nize AD for lessening the mortality rate. Initially, the scans were collected from
ADNI database and then Lucy Richardson (LR) method was used to eliminate the
Gaussian noise from the collected brain images. After denoising the images, pro-
long adaptive exclusive analytical atlas method was used for segmenting the affected
region from the healthy region. Then, Gray level co-occurrence matrix (GLCM)
features were applied to extract the feature vectors from the segmented regions. At

13
Distributed and Parallel Databases

last, N-fold cross validation methodology was utilized to recognize the AD related
features effectively. From the simulation result, the developed framework achieved
good recognition performance related to the previous works in terms of specificity,
accuracy and sensitivity. Hence, the LR approach was non-linear in nature, so its
convergence was very slow.
Tong et al. [19] developed a new multimodal classification approach for recog-
nizing the AD utilizing MRI data. Initially, a non-linear graph fusion method was
applied to construct a graph for every modality separately. Then, a unified graph
was obtained by performing iterative cross diffusion procedure that significantly
improves the connection between the graphs. The iterative cross diffusion procedure
was repeated until the unified graph converges. At last, the classification was per-
formed by utilizing the obtained unified graph. In this research work, ADNI dataset
was utilized to verify the efficacy of the developed approach. From the experimen-
tal analysis, the developed classification methodology attained better performance
related to the existing classifiers. Unlike other classification methodologies, the
graph based approaches consumes more memory space.
Beheshti et al. [20] developed a new system to recognize AD based on classifica-
tion error and feature ranking. In this literature, several feature ranking approaches
were utilized to extract the feature vectors like statistical dependency, Gini index,
Fisher’s criterion, Pearson’s correlation coefficient, t-test score, information gain
and mutual information. Then, classification error was accomplished for finding the
optimum size of the selected feature vectors. Finally, SVM classification approach
was used to classify the collected brain images as MCI, AD, and normal. Experi-
mental estimation shows that the proposed work performance was superior related
to the previous research works. Therefore, SVM was a binary classifier that supports
only two class classification problems.
Beheshti and Demirel [21] introduced a new technique for AD recognition and
classification. In this work, probability distribution function (PDF) approach was
used as a feature selection technique that significantly diminishes the high dimen-
sional data into low dimensional data. The PDF technique not only lessens the
dimension of the feature vectors and also considerably extracts the optimal features.
The selected feature vectors were classified by utilizing SVM classifier. The experi-
mental outcome showed that the PDF–SVM was highly reliable in AD recognition
and classification compared to the earlier systems. One of the major drawbacks of
PDF approach was it includes more outliers that may results in misclassification.
Liu et al. [22] presented a new deep learning system for automatic recognition
of AD. At first, the anatomical landmarks from the images were identified and then
convolutional neural network (CNN) was used for feature learning. In this litera-
ture, the efficacy of the developed system was verified on three databases such as
MIRIAD, ADNI-1, and ADNI-2. Experimental consequence demonstrates that the
developed deep learning system attained better performance in both image classi-
fication and retrieval. Usually, the performance of the CNN model depends on the
amount of data, where it performs poorly if the input data was less.
Liu et al. [23] developed deep multitask multichannel learning (DM2L) approach
for AD recognition. Initially, the anatomical landmarks from the MR imaging
were identified and then DM2L approach was utilized for image classification and

13
 Distributed and Parallel Databases

regression. The developed DM2L approach not only extracts the demographic infor-
mation from the data and also automatically learns the discriminative features for
classification. The effectiveness of the developed approach was tested on MIRIAD,
ADNI-1, and ADNI-2 databases. From the experimental simulation, the developed
approach DM2L outperforms the earlier approaches in disease classification and
regression. The DM2L methodology have the problem of over fitting, because it
needs huge dataset for training.
Islam and Zhang [24] utilized deep CNN method for recognizing and classify-
ing the AD. The efficiency of the developed method was tested on an imbalanced
database (OASIS). As mentioned previously, the deep learning models performs
inadequately if the collected input data was less. Chen and Pham [25] introduced a
framework for AD recognition and classification. In this literature, semi-variogram
and regularity dimension were utilized for extracting the structural features from the
brain scans. The extracted features were classified into four classes by utilizing Hid-
den Markov Models (HMM) such as middle-aged, elder, non-demented young, and
demented elder subjects. The HMMs model often have a large number of unstruc-
tured parameters that may leads to misclassification.
Wei et al. [26] has developed a new framework for Alzheimer detection. In this
literature, adaptive histogram was utilized for image enhancement and then the
enhanced image was decomposed into four intrinsic mode functions using bidi-
rectional empirical mode. Then, local binary patterns (LBP) was used for feature
extraction. Finally, random forest and SVM-Poly 1 were applied for binary clas-
sification. Sultan et al. [27] developed an effective model for Alzheimer disease
detection. The developed model includes three modules; medicine, object and face
recognition. In this literature, speeded up robust features and histogram of oriented
gradient features were utilized for feature extraction. Finally, multi-SVM classifier
was applied for disease classification. In addition, Kamathe and Joshi [28] devel-
oped a new tissue segmentation and classification approach for Alzheimer detection
based on SVM, and independent component analysis with band expansion process.
Additionally, Park et al. [29] developed a novel feature selection algorithm; Raman
spectra for automatic Alzheimer detection. The developed algorithm effectively
finds the peak from the preprocessed spectrum as the feature candidates for classi-
fication. However, the developed frameworks were effective in binary classification
and it was not suitable for multi-class classification. Jo et al. [30] applied convolu-
tional neural network (CNN) for the classification of AD. The ADNI dataset was
applied to test the performance of CNN in AD classification. Goceri [31] proposed
CNN and gradient based optimization method for the AD classification. The ADNI
dataset was applied to test the performance of proposed CNN classification.

3 Proposed model

AD is a form of neuro-degenerative disorder that leads to synaptic dysfunction,

structural changes, and progressive loss of cognitive function in the brain [32,
33]. In this research paper, a supervised framework is proposed for early iden-
tification of AD. The proposed framework includes five phases such as image

13
Distributed and Parallel Databases

acquisition, image denoising, segmentation, feature extraction and selection, and

classification. The graphical design of the proposed framework is given in Fig. 1.

Fig. 1  Workflow of the pro-

posed model Image collection
ADNI, OASIS, MIRIAD and
NIMHANS

Image pre-processing
Adaptive histogram equalization
and region growing for skull
removal

Image segmentation
Grey wolf optimization based
clustering

Feature extraction
Local ternary pattern, dual-tree
complex wavelet transform, and
Tamura features

Feature selection
ReliefF algorithm

Classification
Deep neural network based on
sparse auto-encoder

13
 Distributed and Parallel Databases

Fig. 2  Sample image of MIRIAD dataset, a normal and b Alzheimer control

Fig. 3  Sample image of ADNI dataset, a normal, b MCI and c Alzheimer control

3.1 Image collection and pre processing

In this work, ADNI, OASIS, MIRIAD, and NIMHANS (simulated) datasets are
utilized for verifying the efficiency of the proposed model. The MIRIAD dataset
includes 708 MRI scans of 23 healthy subjects and 46 Alzheimer affected subjects
[34]. In this database, the brain scans are captured at the intervals of 2, 6, 14, 26, 38,
and 52 weeks from the baseline. In addition, the MIRIAD dataset contains the infor-
mation about subject’s age, gender and Mini Mental State Examination (MMSE).
Sample brain image of MIRIAD dataset is shown in Fig. 2.
Meanwhile, the ADNI database comprises of 1.5 T and 3.0 T t1w MRI brain
scans for 819 individuals (192 subjects with AD, 229 normal subjects and 398 sub-
jects with MCI). In this dataset, the brain scans are recorded for 12 months by using
some functional and standard cognitive measures [35, 36]. The sample brain scans
of ADNI dataset is indicated in Fig. 3.
Correspondingly, NIMHANS database includes 800 scans of 99 subjects
(39 AD patients, and 60 normal subjects), where the individuals age ranges
from 55 to 87 years. In this dataset, the AD patients are chosen from outpatient

13
Distributed and Parallel Databases

Fig. 4  Sample image of NIMHANS dataset a normal and b Alzheimer control

Table 1  MRI acquisition details

Flip angle (°) 10
Orientation Sagittal
Slice number 128
Sequence MP-RAGE
Resolution (pixels) 256 × 256 (1 × 1 mm)
Thickness, gap (mm) 1.25, 0
Echo time (TE) 4 ms
Repetition time (TR) 9.7 ms
TI 20 ms
TD 200 ms

department, geriatric clinic, NIMHANS, Bangalore [37]. The sample brain scans
of NIMHANS data base is specified in Fig. 4.
Similarly, OASIS database comprises of 434 MR sessions and 416 subjects,
whose age ranges from 18 to 96. Every individuals includes three or four MRI
scans, which are obtained in single scan sessions. In addition, OASIS dataset
contains information about the individual’s demographics like clinical dementia
rating, MMSE score, socioeconomic status, age, number of patients, education,
and gender [38]. Table 1 describes about the MRI acquisition details. The sample
brain scans of OASIS database is indicated in Fig. 5.
After collecting the image, AHE is applied for enhancing the quality of brain
scans by transforming the pixel intensity values [39]. The undertaken approach
overcomes the problems in the global linear minimum–maximum windowing
and also delivers the desired information about the image. Then, region growing
is applied for removing skull region from the enhanced image. It is also named
as pixel based image segmentation methodology, because it includes seed point
selection [40]. Region growing determines the neighbourhood image pixels of
seed points and also examines whether the neighbourhood pixels should be added

13
 Distributed and Parallel Databases

Fig. 5  Sample image of OASIS dataset a normal control, b MCI, and c AD

to the selected region or not. After performing 250 iterations, the skull region
is appropriately removed from the enhanced brain scans. Sample enhanced and
skull removed image is graphically denoted in Fig. 6.

3.2 Image segmentation

After denoising the image, segmentation is accomplished by utilizing GWOCA in

order to segment the brain tissues. Generally, clustering is the procedure of dividing
the data points n into clusters k on the basis of Euclidean distance in the d dimen-
sional space Rd . Let, the data is considered as X = {x1 , x2 , … , xn }, and the k clusters
are stated as C = {C1 , … , CK }. The clusters should maintain three basic characteris-
tics, which are mathematically indicated in the Eqs. (1)–(3). Hence, each and every
cluster need to have at least one data point as mentioned in Eq. (1).
Ci ≠ ∅, ∀i ∈ {1, 2, … , k}. (1)
Two dissimilar clusters should not have data points in common that is mathemati-
cally expressed in Eq. (2).

Fig. 6  a Input image, b enhanced image, and c skull removed image

13
Distributed and Parallel Databases

Ci ∩ Cj = ∅, ∀i ≠ j and i, j ∈ {1, 2, … , k. (2)

Every data point need to be combined with a cluster as denoted in Eq. (3).

⋃
k
Ci = X. (3)
i=1

In clustering, it is very hard to identify the optimum clusters C∗ on the basis of

feasible solutions C = {C1 , C2 , … ., CN(n,k) , where, N(n, k) is stated as number of
feasible clusters that is mathematically denoted in the Eqs. (4) and (5).
( )
1 ∑
k
k
N(n, k) = (−1)k−i (i)n , (4)
k! i=1 i

Of = Optimize f (X, C), (5)

c

where Of is optimized value of f (X, C), f (X, C) is indicated as statistical function

that validates the quality of cluster partition. Then, the undertaken GWO algorithm
imitates the social hierarchy and hunting behaviour of grey wolves, which usually
includes three basic steps such as encircling, hunting and attacking prey. Consider
the best solution as 𝛼 , and the remaining solutions are named as 𝛽 , 𝛿 and Ω for mod-
elling the hierarchy of wolves. Initially, the grey wolves encircle the prey during the
hunt. Encircling performance of grey wolves are simulated by using the Eqs. (6) and
(7).
| |
D = |C ⋅ Xprey (t) − Xwolf (t)|, (6)
| |

Xwolf (t + 1) = Xprey (t) − A ⋅ D, (7)

where t is indicated as current iteration, Xprey is represented as position of prey,
A and C are represented as coefficient vectors, and Xwolf is indicated as grey wolf
position. The vectors A and C are estimated by utilizing the Eqs. (8) and ().9
A = 2�a⃗ ⋅ ⃗r1 − a�⃗, (8)

C = 2⃗r2 , (9)
where a�⃗ is minimized from two to zero for every iteration and ⃗r1 and ⃗r2 are denoted
as random values that ranges from [0,1]. Regularly, the hunting is guided by 𝛼 value,
occasionally 𝛽 and 𝛿 values are utilized. For imitating the hunting behaviour of grey
wolves, three best solutions of 𝛼 , 𝛽 and 𝛿 values are obtained [41]. Other search
agents like Ω updates the grey wolf position by utilizing the Eqs. (10), (11), and
(12).

13
 Distributed and Parallel Databases

| |
D𝛼 = ||C1 ⋅ X𝛼 − X ||, D𝛽 = |C2 ⋅ X𝛽 − X |, D𝛿 = ||C2 ⋅ X𝛿 − X ||, (10)
| |

X1 = X𝛼 − A1 ⋅ D𝛼 , X2 = X𝛽 − A2 ⋅ D𝛽 , X3 = X𝛿 − A3 ⋅ D𝛿 , (11)

X�⃗1 + X�⃗2 + X�⃗3

X(t + 1) = . (12)
3
Correspondingly, the basic steps of GWO are followed in the GWOCA algo-
rithm are detailed below,
Agent representation each agent represents k cluster centers and the agent
length is indicated as k × d , where d is indicated as position of the cluster center.
Population initialization in this segment, the cluster centers k are initially
encoded with each agents. Then, the agents are initialized to selected data points
k in the given database. This procedure is repeated for every agents in the total
population N .
Fitness function computation generally, the computation of fitness function
includes two main phases. In first phase, the clusters k are generated on the basis
of clusters centers, which are encoded with the agents. Then, replace the respec-
tive agents with the mean data points of the cluster. In second phase, Euclidean
distance between the data points are calculated by using the Eq. (13).
�
n
� �
f (X, C) min ‖xi − Cl ‖2 � l = 1, 2, … , k}. (13)
�
i=1

Position updation in this segment, the search agent position is updated based
on the location of 𝛼, 𝛽, and Δ.The parameter C is utilized for reducing the prema-
ture convergence, and the parameter A is utilized for exploitation and exploration.
Termination criterion the parameter setting, position updation and fitness com-
putation are executed for 50 iterations. The 𝛼 agent delivers optimal solution to
the clustering issues at the 50th iteration. The sample segmented image is graphi-
cally stated in Fig. 7.

Fig. 7  a Enhanced image, b WM, c GM, and d CSF

13
Distributed and Parallel Databases

3.3 Feature extraction

After segmenting the brain tissues, feature extraction is performed using LTP,
DTCWT, and Tamura features. The hybrid feature extraction includes the advantage
like improved data visualization, speed up the training procedure, increase in explain-
ablity of classification model and over fitting risk reduction. The detailed explanation
about the individual feature is specified in following section.

3.3.1 Local ternary pattern

LTP is an extension of local binary pattern (LBP) that utilizes a thresholding con-
stant for thresholding the pixels into three values from segmented images. Let us
consider Tc as a thresholding constant, p as a neighboring pixels, and c as a center
pixel value. The result of the thresholding is determined in Eq. (14).

⎧ 1 if p > c + k
⎪
LTP = ⎨ 0 if p > c − k and p < c + k . (14)
⎪ −1 if p < c − k
⎩

In LTP, every threshold image pixels includes any one of these three values and
the neighbourhood pixels are combined after thresholding with the ternary pattern
[42]. The ternary pattern is divided into two binary patterns, because the ternary val-
ues are higher in range. The histograms are concatenated for generating a descriptor,
which is very successful in the application of AD identification. The simple idea of
LTP method is to transform the image from intensity space to order space, where the
order of neighbouring pixel is used for creating a monotonic illumination invariant
code for every brain scan.

3.3.2 Dual tree complex wavelet transform

DTCWT is utilized to overcome the drawback of shift variance in the two dimen-
sional image. Initially, DTCWT calculates complex transform of an image by utiliz-
ing four important properties such as shift invariance, computational efficiency high
directionality, and perfect reconstruction from segmented images [43]. The DTCWT
provides four sub images in that fourth level image is utilized for feature extraction.
In this transformation approach, the sub-bands are strongly oriented in the direc-
tions of ±15 ,±45 , and ±75 . Hence, this strong orientation separates the positive
◦ ◦ ◦

and negative frequencies. In DTCWT, shift invariance is an important property that

is mathematically stated in Eq. (15).
Ψg (t) ≈ H{Ψh (t)}, (15)
where Ψh (t) is stated as first wavelet tree and Ψg (t) is denoted as second wavelet tree.
These two wavelet trees are used for satisfying the condition of perfect reconstruc-
tion. Let h0 (n) and h1 (n) are indicated as low and high pass filters for the first wavelet
Ψh (t) tree and g0 (n) and g1 (n) are considered as low and high pass filters for the

13
 Distributed and Parallel Databases

second wavelet Ψg (t) tree. These two wavelet tress are constructed together in the
DTCWT approach that is mathematically denoted in Eq. (16).
Ψ(t) = Ψh (t) + iΨg (t). (16)
The DTCWT approach utilizes filter for all real values, so there is no com-
putation on complex numbers that makes the implementation computationally
effective.

3.3.3 Tamura features

Tamura is an effective texture feature that corresponds to human visual percep-

tion [44]. Usually, texture features are the most extensively utilized visual features
in image recognition and classification, because of its dissimilar orientation of the
objects in the image. Tamura contains six features like directionality, roughness,
coarseness, contrast, regularity and line-likeness and features are measured from
segmented images. In that, contrast, directionality and coarseness are utilized in this
research study for extracting the feature vectors.
Coarseness it utilizes repetition rates and scale for finding the segmented image
texture size. Coarseness takes averages at each point over neighbourhood image pix-
els, where the size of the power is two. The average of neighbourhood pixel size
2k × 2k at the point (x, y) is mathematically denoted in Eq. (17).

∑ ∑
x+2k−1 −1 y+2k−1 −1
Ak (x, y) = f (i, j)∕22k . (17)
i=x−2k−1 j=y−2k−1

Then, find the variations in non-overlapped neighbourhood image pixels in both

vertical and horizontal orientations. Though, the horizontal orientation is mathemat-
ically represented in Eq. (18). Finally, choose the best size every one point that gives
the maximum output value, where k increases E in either direction.
| ( ) ( )|
Ek,h (x, y) = |Ak x + 2k−1 , y − Ak x − 2k−1 , y |. (18)
| |
Contrast it captures the dynamic range of grey levels in the segmented brain
images by distributing the pixel values. The contrast measure is mathematically
indicated in Eq. (19).
𝜎
Contrast = , where 𝛼4 = 𝜇4 ∕𝜎 4 , (19)
(𝛼4 )n

where 𝜇4 is denoted as fourth moment of mean and 𝜎 is indicated as variance.

Directionality it is used to estimate the total degree of directionality by determin-
ing the pixel angles and magnitudes of segmented images. In this scenario, a histo-
gram of edge probability is built by quantising the edge angle that helps in sharp-
ening the edges of the brain images. The formula for directionality 𝜃 is given in
Eq. (20).

13
Distributed and Parallel Databases

ΔV 𝜋
𝜃 = tan−1 + , (20)
ΔH 2
where ΔH and ΔV are horizontal and vertical convolved grey-scale images. Fur-
ther, ReliefF algorithm is utilized for choosing the active features from the extracted
features.

3.4 Feature selection

Usually, feature selection identifies the relevant subsets of data on the basis of a specific
criterion. In imagining analysis, feature selection helps in improving the learning per-
formance and also provides better understand which pixels or features are important for
classification. In this work, ReliefF approach is used to pick the active features from the
extracted feature values that effectively reduces the dimension of the extracted features.
This action significantly reduce the system complexity and provides better classifica-
tion. The ReliefF approach is very effective while dealing with noisy and real time data.
Initially, the ReliefF approach randomly picks the instances ri and then searches for
the k nearest neighbour in the same class is called as nearest hit instances Hj and in the
different class is named as nearest miss instances Mj. In ReliefF approach, Manhattan
distance measure is used for determining the nearest miss instances Mj and nearest hit
instances Hj. The key advantage of Manhattan distance is requiring only limited fea-
tures for representing the data that is sufficient to achieve precise neighbourhood selec-
tion for better disease recognition [45].
Then, the values ri , Hj , and Mj are utilized to update the quality estimation W[A] for
all the attributes A. If Hj and ri have same values, A is partitioned into two instances
with the same classes that is essential for minimizing the quality estimation W[A]. If
Hj and ri have different values, A is partitioned into two instances with the different
classes that is essential to maximize the quality estimation W[A]. The whole procedure
is repeated for m times, where m is denotedas user-determined parameter. In this work,
the user-determined parameter is fixed as 25. In ReliefF approach, the quality estima-
tion W[A] is updated by utilizing the Eqs. (21)–(23). The distance between the attrib-
utes and the nearest hit is calculated in Eq. (22), and the distance between the attrib-
utes and the nearest miss is calculated in Eq. (23). The features are selected in Eq. (21)
based on Eqs. (22) and (23).
− −
W[A] = W[A] + (H + M )∕25, (21)

where

− ∑
k
H= − D(A, ri , Hj )∕k, (22)
j=1

[( ) ]
− ∑ P(C) ∑
k
M= ( ( )) D(A, ri , Mj (C)) ∕k, (23)
C≠cl(ri ) 1 − P cl ri j=1

13
 Distributed and Parallel Databases

where W[A] is indicated as quality estimation, ri is represented as instances, A is

indicated as attributes, Hj and Mj are denoted as nearest hit and nearest miss values,
P(C) is represented as earlier class, D is indicated as distance, C is stated as number
of classes, and cl(ri ) is indicated as ith sample class.

3.5 Deep neural network classification

After feature selection, DNN is used for classifying the stages of AD. At first,
the developed classifier performs training and testing process on the segmented
images, whereas the classification accuracy mainly depends on training. The major
advantage of DNN method is that once the network is trained, it has the capacity
to deliver accurate transformation parameters even if the original image is altered
[46]. In addition, the input pattern is delivered to get an extreme fast output. The
term network indicates inter-connection between the neurons in the diverse system.
The DNN classifier includes three layers; input layer comprises of input neurons that
transfer data by means of neurotransmitters to the second layer (hidden layer) and
the third layer is the output layer.
Usually, DNN is a feed forward network with greedy layer wise training. In DNN,
without any looping functions the data transmits from the input layer to the output
layer. In classification, the possibilities of missing value is very low that is consid-
ered as one of the major advantages of DNN. During prediction time, DNN assigns
a classification score f (x) for every input data sample x = [x1 , … , xN ]. Normally, f
is represented as a function that involves in a sequence of layers for computation that
is mathematically denoted in Eq. (24).
∑
Zij= xi wij ; Zj = Zij + bj ; Xj = g(Zj ), (24)
i

where xi is stated as input layer, xj is denoted as output layer, g(Z j ) is represented as

pooling or mapping function, and wij is represented as model parameter. Here, layer-
wise relevance propagation is used for decomposing the output of classifier f (x) by
means of relevance ri for every input component. This process contributes to the
classification decision that is given in the Eq. (25).
∑
f (x) = ri . (25)
i

If ri < 0, it is indicated as neutral evidence, otherwise it called as negative evi-

dence of the classification. If ri > 0, it is represented as positive evidence that highly
supports the classification decision. Though, the relevance attribute ri is calculated
by using Eq. (26).
� zij
ri = ∑ . (26)
j i zij

DNN is a hierarchical feature learning classification method that helps to inves-

tigate about the unknown feature coherences of input. In DNN, a greedy layer wise

13
Distributed and Parallel Databases

unsupervised pre-training is used to derive the low level features from the high level
features. Therefore, the key objective of DNN is to manage the complex functions
that represents the high level abstraction. Usually, the DNN classifier includes mul-
tiple layers of sparse auto encoders, and the output of sparse auto encoders supports
the successive input layers. Auto encoder learn an approximation to the identity
function that is mathematically shown in Eq. (27).
̂
x = hw,b (x) ≈ x. (27)
DNN classifier exploits unsupervised pre-training with greedy layer wise train-
ing. The first sparse auto encoder (1st hidden layer) is trained on a raw input image
(x) in order to learn the primary features h(1). All the bias and weight parameters
have learned the lesson of the cost function during pre-training. In the next hidden
layer, the auto encoder determines the features by using the similar technique from
the preceding hidden layers. Equation (28) describes the cost function of sparse auto
encoder with softmax classifier.

1 ∑( ) ∑ ( ) 𝜆 ∑∑ 2
n m n m
cost =
2n i=1
̂
xi − xi 2 + 𝛽 KL p|p̂J + 𝜃 ,
2 i=1 j=1 ij (28)
j=1

where m is represented as hidden nodes,n is denoted as input nodes,𝛽 is indicated

as weight of sparsity penalty, ̂
pj is stated as probability of firing activity,𝜌 is rep-
resented as sparsity parameter, 𝜆 is represented as weight delay, KL is expressed
as Kullback–Leibler divergence function, ̂ xi is denoted as auto encoder hidden unit,
when the network is given a specific input x and 𝜃 is indicated as weight of hidden
nodes.

4 Result and discussion

This section explained about the result and discussion of the proposed framework. In
this research study, MATLAB (2018A) was used for experimental evaluation. In this
scenario, the proposed work performance was compared with dissimilar classifiers
and some prior research works on the databases; ADNI, MIRIAD, OASIS and NIM-
HANS for estimating the proposed work effectiveness. In this study, the proposed
framework performance was validated in light of accuracy, f-score, MCC, FOR,
error rate and miss rate. Generally, the performance measure is determined as the
process of collecting and analysing the information on the basis of system perfor-
mance. The GWO number of population is set as 100, and number of iteration is set
as 50. The parameter settings of DNN classifier is as follows; hidden layers = 300,
sparsity proportion = 0.10, maximum iteration; SAE learning = 50, L2 weight regu-
larization = 0.005, maximum iteration; softmax learning = 50, scale data = false, and
sparsity regularization = 5. The total number of features considered in this method
are 10,761 and ReliefF feature selects 30% of features is 3228 features for classifi-
cation. The feature selection percentage less than 30% affects the performance and
more than 30% has constant performance. The mathematical equations of accuracy,

13
 Distributed and Parallel Databases

f-score, MCC, FOR, error rate and miss rate for multiclass classification is signified
in the Eqs. (29) and (30)–(34).
TP + TN
Accuracy = × 100, (29)
FN + TP + FP + TN

2TP
F-score = × 100, (30)
2TP + FP + FN

TP × TN − FP × FN
MCC = √ × 100, (31)
(TN + FN)(TP + FP)(TN + FP)(TP + FN)

FN
FOR = × 100, (32)
FN + TN

Error rate = 100 − accuracy, (33)

FN
Miss rate = × 100, (34)
FN + TP
where TP is signified as true positive,FP is indicated as false positive, TN is stated as
true negative, and FN is indicated as false negative.

4.1 Quantitative investigation on ADNI database

In this segment, ADNI database is utilized to investigate the performance of the

proposed framework. In Table 2, the proposed framework performance is evalu-
ated in terms of accuracy, f-score and MCC. In this scenario, the quantitative
investigation is done for 150 scans (50 scans from AD class, 50 scans from nor-
mal class and the remaining 50 scans from MCI class) with 80% training and
20% testing of brain images. Additionally, the performance investigation is done
with different classification approaches such as neural network (NN), long short-
term memory (LSTM) and DNN. In the experimental study, the accuracy of DNN

Table 2  Performance investigation in light off-score, accuracy and MCC on ADNI database
Feature selection Classifier Accuracy (%) f-score (%) MCC (%)

Without ReliefF NN 69.47 73.70 72.82

LSTM 96.54 95.81 96.22
DNN 97.02 98.34 96.91
With ReliefF NN 72.91 74.22 73.76
LSTM 97.06 98.15 97.18
DNN 98.6 99.57 99.40

Bold represent the value received by applying proposed method compare to the other existing method

13
Distributed and Parallel Databases

Fig. 8  Graphical illustration of the proposed framework by means off-score, accuracy and MCC on
ADNI database

Table 3  Performance investigation in light of FOR, miss rate, and error rate on ADNI database
Feature selection Classifier FOR (%) Miss rate (%) Error rate (%)

Without ReliefF NN 31.66 30.93 30.53

LSTM 4.29 3.94 3.46
DNN 3.85 4.07 2.98
With ReliefF NN 28.46 28.14 27.09
LSTM 4.06 4.47 2.94
DNN 0.38 1.88 1.40

Bold represent the value received by applying proposed method compare to the other existing method

with ReliefF algorithm is 98.6% that is higher compared to other classifiers. In

this scenario, the proposed work almost showed 1.5% to 25% improvement in
accuracy compared to other classifiers. In addition, the f-score and MCC of the
proposed work is superior related to other comparative classification methods.
Graphical illustration of the proposed framework by means of accuracy, f-score
and MCC on ADNI database is represented in Fig. 8.
In Table 3, the proposed framework performance is evaluated by means of
FOR, miss rate, and error rate. The GWO method has the advantage of strong
exploring capacity and reduce the probability of fall into local optima. The GWO
method hunting process and hierarchical process helps to improve the conver-
gent rate of the algorithm to find optimal solution. These characteristics helps to
provide efficient clustering performance in the proposed method. In with ReliefF
algorithm, the FOR value of DNN classifier is 0.38% and the approaches (NN
and LSTM) achieved 28.46% and 4.06% of FOR value. Likewise, the miss rate
of DNN is 1.88% and the comparative classification methods (NN and LSTM)
attained 28.14% and 4.47% of miss rate. Additionally, the error rate of DNN
method is 1.40% and other classifiers attained 27.09% and 2.94% of error rate.

13
 Distributed and Parallel Databases

Fig. 9  Graphical illustration of the proposed framework in light of FOR, miss rate, and error rate on
ADNI database

Table 4  Performance Method Classifier Accuracy (%) f-score (%) MCC (%)
investigation of GWO in light of
Accuracy, f-score, and MCC on Without GWO NN 68.23 69.11 67.12
ADNI database
LSTM 95.56 96.42 95.41
DNN 97.01 98.01 98.21
With GWO NN 72.91 74.22 73.76
LSTM 97.06 98.15 97.18
DNN 98.6 99.57 99.40

Bold represent the value received by applying proposed method

compare to the other existing method

The graphical illustration of the proposed framework by means of FOR, miss

rate, and error rate on ADNI database is specified in Fig. 9.
The proposed method is evaluated with GWO and without GWO in the ADNI
dataset to analysis the performance, as shown in Table 4. The GWO increases the
performance of the proposed method due to the GWO has the advantages of strong
exploring capacity. The GWO method has better convergence performance and
reduce fall into local optima.

4.2 Quantitative investigation on OASIS database

In the Tables 5 and 6, the performance of the proposed work is evaluated with dis-
similar classification approaches in light of accuracy, f-score, MCC, FOR, error rate
and miss rate. In addition, the proposed work is validated for all the three slices
(Axial, Coronal, and Sagittal) in OASIS database. Tables 5 and 6 indicates the per-
formance of the proposed framework with ReliefF and without ReliefF methodol-
ogy. Here, the efficacy of feature selection methodology is evaluated with differ-
ent classification approaches. In with ReliefF methodology, the DNN classifier

13
Distributed and Parallel Databases

Table 5  Performance validation in light off-score, accuracy and MCC on OASIS database
Slices Feature selection Classifier Accuracy (%) f-score (%) MCC (%)

Axial Without ReliefF NN 66.32 69.14 70.72

LSTM 91.28 84.75 86.10
With ReliefF NN 67.10 70.56 72.18
DNN 93.91 86.38 83.14
LSTM 91.67 86.09 88.24
DNN 95.73 88.81 84.34
Coronal Without ReliefF NN 70.71 70.18 64.95
LSTM 86.70 88.33 87.80
DNN 94.30 92.10 91.36
With ReliefF NN 71.99 74.09 65.97
LSTM 90.34 91.01 89.11
DNN 96.27 94.13 93.57
Sagittal Without ReliefF NN 77.82 63.26 83.34
LSTM 90.85 87.82 89.30
DNN 92.29 90.98 88.47
With ReliefF NN 79.47 65.24 85.11
LSTM 92.56 90.99 91.92
DNN 94.27 93.01 90.96

Bold represent the value received by applying proposed method compare to the other existing method

Table 6  Performance investigation in light of FOR, miss rate, and error rate on OASIS database
Slices Feature selection Classifier FOR (%) Miss rate (%) Error rate (%)

Axial Without ReliefF NN 32.41 32.68 33.68

LSTM 6.09 8.13 8.72
DNN 3.45 5.32 6.09
With ReliefF NN 30.07 31.40 32.90
LSTM 6.57 7.70 8.33
DNN 1.98 3.48 4.27
Coronal Without ReliefF NN 27.36 28.57 29.29
LSTM 10.63 12.64 13.30
DNN 3.42 4.80 5.70
With ReliefF NN 26.05 27.82 28.01
LSTM 6.89 8.23 9.66
DNN 0.88 2.80 3.73
Sagittal Without ReliefF NN 20.84 21.89 22.18
LSTM 7.06 8.48 9.15
DNN 6.35 6.37 7.71
With ReliefF NN 19.29 19.53 20.53
LSTM 5.73 5.81 7.44
DNN 3.24 5.06 5.73

Bold represent the value received by applying proposed method compare to the other existing method

13
 Distributed and Parallel Databases

averagely enhanced the accuracy in AD identification up to 2–10% related to with-

out ReliefF methodology. In this work, the hybrid features significantly finds the
non-linear and linear structure of the brain scans that helps in lessening the semantic
space between the high and low level features. The performance measures show that
the proposed work performs effectively well in AD identification related to the exist-
ing works.

4.3 Quantitative investigation on MIRIAD database

In this section, MIRIAD database is utilized to investigate the performance of the

proposed framework for classifying the normal and Alzheimer controls. In this sce-
nario, 80% of the brain scans are utilized for training and 20% of the brain scans
are utilized for testing. Table 7 states the performance of the proposed framework
by means off-score, accuracy, MCC, FOR, error rate and miss rate. In Table 7, the
performance of the proposed framework is evaluated with ReliefF algorithm and
some existing deep learning classification approaches. In ReliefF algorithm, the
DNN classifier averagely enhanced the accuracy up to 13% related to other deep
learning classifiers. In this research study, ReliefF algorithm utilizes only limited
features to represent the data that is adequate to achieve better disease identification
and classification. By choosing the limited features, the proposed framework supe-
riorly diminishes the “curse of dimensionality” issue. Tables 7 and 8 shows that the
ReliefF algorithm performs significantly in AD recognition related to the existing
algorithms. The graphical illustration of the proposed framework by means of accu-
racy, f-score, MCC, FOR, error rate and miss rate on MIRIAD dataset is indicated in
the Figs. 10 and 11.

4.4 Quantitative investigation on NIMHANS database

In this section, NIMHANS dataset is utilized for evaluating the performance of

the proposed framework for binary classification. In this research, AD classifica-
tion is implemented on the basis of digital image-processing platform to classify
two Alzheimer classes: normal, and AD. In this section, Tables 9 and 10 shows
the performance investigation of existing and proposed works for binary classifica-
tion on NIMHANS dataset. The classification accuracy of proposed work (ReliefF

Table 7  Performance investigation in light off-score, accuracy and MCC on MIRIAD database
Feature selection Classifier Accuracy (%) f-score (%) MCC (%)

Without ReliefF NN 81.51 81.37 83.26

LSTM 96.10 95.57 95.41
DNN 96.72 98.75 96.85
With ReliefF NN 85.27 84.64 84.73
LSTM 97.25 96.88 97.26
DNN 99.5 99.79 99.62

Bold represent the value received by applying proposed method compare to the other existing method

13
Distributed and Parallel Databases

Table 8  Performance investigation in light of FOR, error rate and miss rate on MIRIAD database
Feature selection Classifier FOR (%) Miss rate (%) Error rate (%)

Without ReliefF NN 16.97 17.29 18.49

LSTM 3.60 5.23 3.90
DNN 2.64 4.18 3.28
With ReliefF NN 17.66 16.98 14.73
LSTM 3.37 3.89 2.75
DNN 0.36 0.40 0.50

Bold represent the value received by applying proposed method compare to the other existing method

Fig. 10  Graphical illustration of the proposed framework by means off-score, accuracy and MCC on
MIRIAD database

Fig. 11  Graphical evaluation of the proposed framework in light of FOR, error rate and miss rate on
MIRIAD database

13
 Distributed and Parallel Databases

Table 9  Performance investigation in light off-score, accuracy and MCC on NIMHANS database
Feature selection Classifier Accuracy (%) f-score (%) MCC (%)

Without ReliefF NN 75.27 75.13 77.02

LSTM 89.98 89.45 89.29
DNN 91.52 93.55 92.65
With ReliefF NN 74.93 74.30 74.39
LSTM 91.33 92.96 92.34
DNN 93.84 92.13 93.96

Bold represent the value received by applying proposed method compare to the other existing method

Table 10  Performance investigation by means of FOR, error rate and miss rate on NIMHANS database
Feature selection Classifier FOR (%) Miss rate (%) Error rate (%)

Without ReliefF NN 23.98 25.91 24.73

LSTM 10.21 11.11 10.02
DNN 8.82 9.88 8.48
With ReliefF NN 29.56 25.32 25.07
LSTM 11.95 9.47 8.67
DNN 7.70 6.92 6.16

Bold represent the value received by applying proposed method compare to the other existing method

algorithm + DNN) is 93.84% and the existing approaches achieves less classification
accuracy. Correspondingly, the f-score, MCC, FOR, error rate and miss rate value of
the proposed work is really better than the existing approaches. Graphical illustra-
tion of the proposed framework by means of accuracy, f-score, MCC, FOR, error
rate and miss rate on NIMHANS dataset is specified in the Figs. 12 and 13. In this
study, the complexity of the proposed work is estimated using O(N 4 log2 N), where is
N indicated as input size.

4.5 Comparative analysis

In this subsection, the comparative study of proposed and existing works is depicted
in Table 11. Beheshti et al. [20] developed a new system for recognizing AD on the
basis of classification error and feature ranking. In this literature, the developed sys-
tem performance was tested on ADNI dataset and achieved 92.48% of accuracy. In
addition, Beheshti and Demirel [21] developed PDF and SVM for AD recognition
and classification. Experimental results showed that the PDF–SVM was highly reli-
able in AD recognition and classification related to the earlier systems, which almost
achieved 89.65% of classification accuracy on ADNI dataset.
Liu et al. [23] developed DM2L approach for AD recognition. The developed
approach extracts the demographic information from the data and also automatically
learns the discriminative features for classification. The efficiency of the developed

13
Distributed and Parallel Databases

Fig. 12  Graphical illustration of the proposed framework by means of f-score, accuracy and MCC on
NIMHANS database

Fig. 13  Graphical illustration of the proposed framework by means of FOR, error rate and miss rate on
NIMHANS dataset

Table 11  Comparative study References Dataset Recognition

accuracy (%)

Beheshti et al. [20] ADNI 92.48

Beheshti and Demirel [21] 89.65
Liu et al. [23] MIRIAD 93.7
Islam and Zhang [24] OASIS 93.18
Chen and Pham [25] 89.3
Jo et al. [30] ADNI 96
Goceri [31] ADNI 98.06
Proposed work ADNI 98.6
MIRIAD 99.5
OASIS 95.42
NIMHANS 93.84

Bold represent the value received by applying proposed method

compare to the other existing method

13
 Distributed and Parallel Databases

approach was tested on MIRIAD database. Here, the developed approach achieved
93.7% of accuracy in AD recognition. Additionally, Islam and Zhang [24] and
Chen and Pham [25] used deep CNN and HMM models to recognize and classify
the AD. The efficiency of the developed method was tested on an imbalanced data-
base (OASIS). The developed systems achieved 93.18% and 89.3% of accuracy in
AD recognition and classification, respectively. Related to the existing studies, the
proposed work attained better outcome by means of accuracy. In this work, fea-
ture selection is an integral part of AD recognition and classification. Each MRI
scans contains numerous features that leads to “curse of dimensionality” problem.
So, feature selection is crucial in selecting the active features that helps in better
classification.

5 Conclusion

In the medical field, MRI based AD recognition is an emerging research topic. The
aim of this research study is to propose an effective segmentation and classifica-
tion methods for classifying the abnormality and normality of AD on NIMHANS,
ADNI, MIRIAD, and OASIS datasets. In this experimental research, hybrid feature
extraction (LTP, DTCWT and Tamura features) is utilized to extract the feature vec-
tors from the pre-processed images. Then, a superior feature selection algorithm:
ReliefF is utilized to eliminate the unwanted or irrelevant features from the extracted
feature vectors. By utilizing the optimal feature values, the abnormality and normal-
ity of AD is classified by using DNN. Related to other existing works, the proposed
framework attained better performance in AD recognition in light of accuracy,
which almost showed 2.2–6% improvement in classification. In future work, the pro-
posed framework is applied on three dimensional data for assisting the clinicians in
early recognition of AD and test in large clinical dataset.

References
1. Ghosh, S., Chandra, A., Mudi, R.K.: A novel fuzzy pixel intensity correlation based segmentation
algorithm for early detection of Alzheimer’s disease. Multimed. Tools Appl. 78(9), 12465–12489
(2019)
2. Counts, S.E., Ikonomovic, M.D., Mercado, N., Vega, I.E., Mufson, E.J.: Biomarkers for the early
detection and progression of Alzheimer’s disease. Neurotherapeutics 14(1), 35–53 (2017)
3. Duraisamy, B., Shanmugam, J.V., Annamalai, J.: Alzheimer disease detection from structural MR
images using FCM based weighted probabilistic neural network. Brain Imaging Behav. 13(1),
87–110 (2019)
4. Baskar, D., Jayanthi, V.S., Jayanthi, A.N.: An efficient classification approach for detection of Alz-
heimer’s disease from biomedical imaging modalities. Multimed. Tools Appl. 78(10), 12883–12915
(2019)
5. Tan, X., Liu, Y., Li, Y., Wang, P., Zeng, X., Yan, F., Li, X.: Localized instance fusion of MRI data
of Alzheimer’s disease for classification based on instance transfer ensemble learning. Biomed. Eng.
Online 17(1), 49 (2018)
6. Magalhães, T.N.C., Weiler, M., Teixeira, C.V.L., Hayata, T., Moraes, A.S., Boldrini, V.O., Dos
Santos, L.M., de Campos, B.M., de Rezende, T.J.R., Joaquim, H.P.G., Talib, L.L.: Systemic

13
Distributed and Parallel Databases

inflammation and multimodal biomarkers in amnestic mild cognitive impairment and Alzheimer’s
disease. Mol. Neurobiol. 55(7), 5689–5697 (2018)
7. Bhateja, V., Moin, A., Srivastava, A., Bao, L.N., Lay-Ekuakille, A., Le, D.N.: Multispectral medical
image fusion in Contourlet domain for computer based diagnosis of Alzheimer’s disease. Rev. Sci.
Instrum. 87(7), 074303 (2016)
8. Escudero, J., Ifeachor, E., Zajicek, J.P., Green, C., Shearer, J., Pearson, S., Alzheimer’s Disease
Neuroimaging Initiative: Machine learning-based method for personalized and cost-effective
detection of Alzheimer’s disease. IEEE Trans. Biomed. Eng. 60(1), 164–168 (2012)
9. Nguyen, T.G., Phan, T.V., Hoang, D.T., Nguyen, T.N., So-In, C.: Efficient SDN-based traffic
monitoring in IoT networks with double deep Q-network. In: International Conference on Com-
putational Data and Social Networks, December 2020 (pp. 26–38). Springer, Cham (2020)
10. Zhang, J., Liu, M., An, L., Gao, Y., Shen, D.: Alzheimer’s disease diagnosis using landmark-
based features from longitudinal structural MR images. IEEE J. Biomed. Health Inform. 21(6),
1607–1616 (2017)
11. Trambaiolli, L.R., Spolaôr, N., Lorena, A.C., Anghinah, R., Sato, J.R.: Feature selection before
EEG classification supports the diagnosis of Alzheimer’s disease. Clin. Neurophysiol. 128(10),
2058–2067 (2017)
12. Beheshti, I., Demirel, H., Alzheimer’s Disease Neuroimaging Initiative: Feature-ranking-based
Alzheimer’s disease classification from structural MRI. Magn. Reson. Imaging 34(3), 252–263
(2016)
13. Ge, C., Qu, Q., Gu, I.Y.H., Jakola, A.S.: Multi-stream multi-scale deep convolutional networks
for Alzheimer’s disease detection using MR images. Neurocomputing 350, 60–69 (2019)
14. Xu, L., Wu, X., Chen, K., Yao, L.: Multi-modality sparse representation-based classification
for Alzheimer’s disease and mild cognitive impairment. Comput. Methods Programs Biomed.
122(2), 182–190 (2015)
15. Gray, K.R., Aljabar, P., Heckemann, R.A., Hammers, A., Rueckert, D., Alzheimer’s Disease
Neuroimaging Initiative: Random forest-based similarity measures for multi-modal classification
of Alzheimer’s disease. Neuroimage 65, 167–175 (2013)
16. Liu, X., Tosun, D., Weiner, M.W., Schuff, N., Alzheimer’s Disease Neuroimaging Initiative:
Locally linear embedding (LLE) for MRI based Alzheimer’s disease classification. Neuroimage
83, 148–157 (2013)
17. Aguilar, C., Westman, E., Muehlboeck, J.S., Mecocci, P., Vellas, B., Tsolaki, M., Kloszewska,
I., Soininen, H., Lovestone, S., Spenger, C., Simmons, A.: Different multivariate techniques for
automated classification of MRI data in Alzheimer’s disease and mild cognitive impairment.
Psychiatry Res. Neuroimaging 212(2), 89–98 (2013)
18. Sampath, R., Indumathi, J.: Earlier detection of Alzheimer disease using N-fold cross validation
approach. J. Med. Syst. 42(11), 217 (2018)
19. Tong, T., Gray, K., Gao, Q., Chen, L., Rueckert, D., Alzheimer’s Disease Neuroimaging Ini-
tiative: Multi-modal classification of Alzheimer’s disease using nonlinear graph fusion. Pattern
Recognit. 63, 171–181 (2017)
20. Beheshti, I., Demirel, H., Farokhian, F., Yang, C., Matsuda, H., Alzheimer’s Disease Neuroimag-
ing Initiative: Structural MRI-based detection of Alzheimer’s disease using feature ranking and
classification error. Comput. Methods Programs Biomed. 137, 177–193 (2016)
21. Beheshti, I., Demirel, H., Alzheimer’s Disease Neuroimaging Initiative: Probability distribution
function-based classification of structural MRI for the detection of Alzheimer’s disease. Comput.
Biol. Med. 64, 208–216 (2015)
22. Liu, M., Zhang, J., Nie, D., Yap, P.T., Shen, D.: Anatomical landmark based deep feature rep-
resentation for MR images in brain disease diagnosis. IEEE J. Biomed. Health Inform. 22(5),
1476–1485 (2018)
23. Liu, M., Zhang, J., Adeli, E., Shen, D.: Joint classification and regression via deep multi-task
multi-channel learning for Alzheimer’s disease diagnosis. IEEE Trans. Biomed. Eng. 66(5),
1195–1206 (2018)
24. Islam, J., Zhang, Y.: Brain MRI analysis for Alzheimer’s disease diagnosis using an ensemble
system of deep convolutional neural networks. Brain Inform. 5(2), 2 (2018)
25. Chen, Y., Pham, T.D.: Development of a brain MRI-based hidden Markov model for dementia
recognition. Biomed. Eng. Online 12(1), S2 (2013)
26. Wei, J.K.E., Jahmunah, V., Pham, T.H., Oh, S.L., Ciaccio, E.J., Acharya, U.R., Yeong, C.H.,
Fabell, M.K.M., Rahmat, K., Vijayananthan, A., Ramli, N.: Automated detection of Alzheimer’s

13
 Distributed and Parallel Databases

disease using Bi-directional Empirical Model Decomposition. Pattern Recognit. Lett. 135, 106–
113 (2020)
27. Sultan, S., Javed, A., Irtaza, A., Dawood, H., Dawood, H., Bashir, A.K.: A hybrid egocentric
video summarization method to improve the healthcare for Alzheimer patients. J. Ambient Intell.
Humaniz. Comput. 10(10), 4197–4206 (2019)
28. Kamathe, R.S., Joshi, K.R.: A novel method based on independent component analysis for brain
MR image tissue classification into CSF, WM and GM for atrophy detection in Alzheimer’s disease.
Biomed. Signal Process. Control 40, 41–48 (2018)
29. Park, A., Baek, S.J., Shen, A., Hu, J.: Detection of Alzheimer’s disease by Raman spectra of rat’s
platelet with a simple feature selection. Chemom. Intell. Lab. Syst. 121, 52–56 (2013)
30. Jo, T., Nho, K., Saykin, A.J.: Deep learning in Alzheimer’s disease: diagnostic classification and
prognostic prediction using neuroimaging data. Front. Aging Neurosci. 11, 220 (2019)
31. Goceri, E.: Diagnosis of Alzheimer’s disease with Sobolev gradient-based optimization and 3D con-
volutional neural network. Int. J. Numer. Methods Biomed. Eng. 35(7), e3225 (2019)
32. Thomas, K.R., Edmonds, E.C., Eppig, J.S., Wong, C.G., Weigand, A.J., Bangen, K.J., Jak, A.J.,
Delano-Wood, L., Galasko, D.R., Salmon, D.P., Edland, S.D.: MCI-to-normal reversion using neu-
ropsychological criteria in the Alzheimer’s Disease Neuroimaging Initiative. Alzheimer’s Dement.
15(10), 1322–1332 (2019)
33. Li, H., Habes, M., Wolk, D.A., Fan, Y., Alzheimer’s Disease Neuroimaging Initiative: A deep learn-
ing model for early prediction of Alzheimer’s disease dementia based on hippocampal magnetic
resonance imaging data. Alzheimer’s Dement. 15(8), 1059–1070 (2019)
34. Malone, I.B., Cash, D., Ridgway, G.R., MacManus, D.G., Ourselin, S., Fox, N.C., Schott, J.M.:
MIRIAD-Public release of a multiple time point Alzheimer’s MR imaging dataset. Neuroimage 70,
33–36 (2013)
35. Basaia, S., Agosta, F., Wagner, L., Canu, E., Magnani, G., Santangelo, R., Filippi, M., Alzheimer’s
Disease Neuroimaging Initiative: Automated classification of Alzheimer’s disease and mild cogni-
tive impairment using a single MRI and deep neural networks. NeuroImage Clin. 21, 101645 (2019)
36. Petersen, R.C., Aisen, P.S., Beckett, L.A., Donohue, M.C., Gamst, A.C., Harvey, D.J., Jack, C.R.,
Jagust, W.J., Shaw, L.M., Toga, A.W., Trojanowski, J.Q.: Alzheimer’s Disease Neuroimaging Initia-
tive (ADNI): clinical characterization. Neurology 74(3), 201–209 (2010)
37. Tripathi, R., Kumar, J.K., Bharath, S., Marimuthu, P., Varghese, M.: Clinical validity of NIMHANS
neuropsychological battery for elderly: a preliminary report. Indian J. Psychiatry 55(3), 279 (2013)
38. Marcus, D.S., Wang, T.H., Parker, J., Csernansky, J.G., Morris, J.C., Buckner, R.L.: Open Access
Series of Imaging Studies (OASIS): cross-sectional MRI data in young, middle aged, nondemented,
and demented older adults. J. Cogn. Neurosci. 19(9), 1498–1507 (2007)
39. Isa, I.S., Sulaiman, S.N., Mustapha, M., Karim, N.K.A.: Automatic contrast enhancement of brain
MR images using Average Intensity Replacement based on Adaptive Histogram Equalization (AIR-
AHE). Biocybern. Biomed. Eng. 37(1), 24–34 (2017)
40. Park, J.G., Lee, C.: Skull stripping based on region growing for magnetic resonance brain images.
Neuroimage 47(4), 1394–1407 (2009)
41. Kapoor, S., Zeya, I., Singhal, C., Nanda, S.J.: A grey wolf optimizer based automatic clustering
algorithm for satellite image segmentation. Procedia Comput. Sci. 115, 415–422 (2017)
42. Athertya, J.S., Kumar, G.S., Govindaraj, J.: Detection of Modic changes in MR images of spine
using local binary patterns. Biocybern. Biomed. Eng. 39(1), 17–29 (2019)
43. Yang, P., Yang, G.: Feature extraction using dual-tree complex wavelet transform and Gray level co-
occurrence matrix. Neurocomputing 197, 212–220 (2016)
44. Howarth, P., Rüger, S.: Evaluation of texture features for content-based image retrieval. In: Interna-
tional Conference on Image and Video Retrieval, pp. 326–334. Springer, Berlin (2004)
45. Reyes, O., Morell, C., Ventura, S.: Scalable extensions of the ReliefF algorithm for weighting and
selecting features on the multi-label learning context. Neurocomputing 161, 168–182 (2015)
46. Saouli, R., Akil, M., Kachouri, R.: Fully automatic brain tumor segmentation using end-to-end
incremental deep neural networks in MRI images. Comput. Methods Programs Biomed. 166, 39–49
(2018)

Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published
maps and institutional affiliations.

13
Distributed and Parallel Databases

Authors and Affiliations

Halebeedu Subbaraya Suresha1 ·

Srirangapatna Sampathkumaran Parthasarathy2
Srirangapatna Sampathkumaran Parthasarathy
vsarathypartha@yahoo.com
1
Department of ECE, PET Research Centre, Mandya, University of Mysore, Mysore, India
2
Department of ECE, P.E.S College of Engineering, Mandya, India

13