Expert Systems - 2022 - Abugabah - Health Care Intelligent System A Neural Netw

14680394, 2022, 9, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/exsy.13003 by <Shibboleth>-member@uel.ac.uk, Wiley Online Library on [23/02/2023].
See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Received: 22 December 2021 Revised: 26 February 2022 Accepted: 10 March 2022
DOI: 10.1111/exsy.13003
ORIGINAL ARTICLE
Health care intelligent system: A neural network based method

for early diagnosis of Alzheimer's disease using MRI images
Ahed Abugabah1 | Atif Mehmood2 | Sultan Almotairi3 | Ahmad A. L. Smadi2
1
College of Technological Innovation, Zayed
University, Abu Dhabi, UAE Abstract
2
School of Artificial Intelligence, Xidian Alzheimer's disease (AD) is a neurodegenerative disease that causes memory loss and
University, Xidian, China
is considered the most common type of dementia. In many countries, AD is com-
3
Faculty of Community College, Majmaah
University, Al Majma'ah, Saudi Arabia monly affecting senior citizens having an aged more than 65 years. Machine learning-
based approaches have some limitations due to data pre-processing issues. We pro-
Correspondence
Ahed Abugabah, College of Technological pose a health care intelligent system based on a deep convolutional neural network
Innovation, Zayed University, Abu Dhabi (DCNN) in this research work. It classifies normal control (NC), mild cognitive impair-
Campus, UAE.
Email: ahed.abugabah@zu.ac.ae, ment (MCI), and AD. The proposed model is employed on white matter (WM), and
ahed_abuga0905@yahoo.com grey matter (GM) tissues with more cognitive decline features. In the experimental
Funding information process, we used 375 Magnetic Resonance Image (MRI) subjects collected from
This work was supported by Zayed University, Alzheimer's disease neuroimaging initiative (ADNI), including 130 NC people,
Under Cluster Fund No R17089,
Abu Dhabi, U.A.E. 120 MCI patients, and 125 AD patients. We extract three major regions during pre-
processing, that is, WM, GM and cerebrospinal fluid (CSF). This study shows promis-
ing classification results for NC versus AD 97.94%, MCI versus AD 92.84%, and NC
versus MCI 88.15% on GM images. Furthermore, our proposed model attained
95.97%, 90.82%, and 86.87% on the same three binary classes on WM tissue,
respectively. When comparing existing studies in terms of accuracy and other evalua-
tion parameters, we found that our proposed approach shows better results than
those approaches based on the CNN method.
KEYWORDS
Alzheimer's disease, intelligent system, neural networks, neuroimaging
1 | I N T RO DU CT I O N
Alzheimer's disease (AD) is a neurodegenerative disease and one of the most famous types of dementia. In this situation, the patient's mental
health progressively decreases, and they can do nothing in everyday routine work. Affected patients are more dependent on family members for
survival. It has also been reported that 116 million peoples will suffer from AD disease in 2050 (Gao et al., 2015). In this disease, the human brain
cerebral cortex shrinks and the ventricles in the brain are expanded. Researchers find out the disease's progress on this evidence that it directly
affects thinking, decision making, and memory loss (Bastin et al., 2021), There is no cure for AD, while available treatments only reduce the pro-
gression of AD. Early treatment's fundamental significance at the initial stage of AD may help reduce disease progression (Ramzan et al., 2020),
Mild cognitive impairment (MCI) is an intermediate stage, mostly recognized as a preclinical AD stage. Mild cognitive impairment (MCI) is a phase
in the aging technique that appears among ordinary intellectual deterioration as well as memory loss. Remembering, communication, reasoning,
and judgement issues are all common symptoms. Research studies show that 15% to 20% of MCI patients progress to AD, but it takes years. If
the MCI stage is examined in-depth, it can decrease AD's more high-risk population (Wang et al., 2020). Early diagnosis of AD produced promising
outcomes. Furthermore, to find out the stable MCI (sMCI) and progressive MCI (pMCI) is a challenging task for researchers (Beheshti et al., 2019).
Expert Systems. 2022;39:e13003. wileyonlinelibrary.com/journal/exsy © 2022 John Wiley & Sons Ltd. 1 of 13
https://doi.org/10.1111/exsy.13003
14680394, 2022, 9, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/exsy.13003 by <Shibboleth>-member@uel.ac.uk, Wiley Online Library on [23/02/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
2 of 13 ABUGABAH ET AL.
The ADNI healthcare database consists of data on every individual's enlistment, demography, physical exams, and critical concentration. As
comma-separated values (CSV) documents, the whole collection of medical studies may be obtained in quantity. The ADNI therapeutic database
contains data on every patient's recruiting, demography, fitness exams, as well as tested hypotheses.
A field of investigation that utilizes imaging technologies to analyse the anatomy as well as the functioning of the skull, with pictures of the
brain produced in a non-invasive manner is termed neuroimaging. It evaluates a number of systems in the diseased limitation, such as cognitive,
executive function, as well as neurological alterations. Neuroimaging modalities also play a pivotal role to diagnose normal control (NC) and AD
patients. These modalities such as magnetic resonance imaging (MRI), positron emission tomography (PET), computed tomography (CT), functional
magnetic resonance imaging (fMRI), single-photon emission computed tomography (SPECT) and structural magnetic resonance imaging (sMRI)
(AlZubi, 2020). These modalities hold information regarding the cerebral cortex, hypothalamus, amygdala, and frontal lobes. All the subcortical
structure is involved in cognitive and behavioural operations (Zhang et al., 2011). Recently, MRI is considered as a primary modality for the diag-
nosis of AD. It has been producing more satisfactory outcomes in terms of NC, MCI, and AD classification. There are several brain images features
that help distinguish between different stages of AD (Kim et al., 2019). Three significant features, such as GM, WM, and CSF help researchers
classify AD stages (Ge et al., 2019).
Machine learning has presented promising results in different application fields such as computer vision (Bharati & Pramanik, 2020), internet
of things, natural language processing, object recognition, and image fusion (Alsmadi et al., 2018), (Al Smadi & Abugabah, 2018). Some researchers
have used machine learning for disease diagnostic systems. In medical image analysis, machine learning is still suffering from two major issues. 1)
high labor intensity required to label data and 2) in different modalities such as MRI and PET, other regions and ages formed during data distribu-
tion (Dua et al., 2020). Intelligent systems are high-tech devices that can interpret besides react to their circumstances. It comes in a diversity of
shapes as well as sizes, from Roombas to face-scanning technology to Amazon's tailored purchasing instructions.
It is directly impacting the classification performance. Recently, deep learning replaces machine learning due to the medical sector's high pre-
diction rate (Mehmood et al., 2020), (Munadhil et al., 2020). A deep learning technique appears to be an improvement in computer-aided diagnosis
(CAD) (Zhu et al., 2019). Convolutional neural network (CNN) is a widely known deep learning method that achieves the highest success in medi-
cal fields, such as neuroimaging, natural language processing, and speech recognition (Alarifi et al., 2019). Furthermore, many researchers devel-
oped several models based on CNN mostly useful for AD diagnosis. Recently, researchers introduced several machine learning methods for AD
classification (Alsharif et al., 2020). These models have some limitations during training the model from scratch because its required a considerable
amount of data for training the model. In medical imaging, the availability of annotated data samples is a big problem (Janghel & Rathore, 2021),
1) making annotated data samples requires maximum time and is more expensive, 2) during training the model with less number of data samples
arises an overfitting issue, which directly impacts on classification performance, 3) in medical data acquisition ethical and privacy issues arise.
This paper develops a CNN model for binary classification of NC, MCI, and AD. More significantly, the proposed model takes input in GM
and WM images extracted from MRI by using the SPM12 tool. We overcome the shortcoming data complication by using an augmentation
approach that also helps to reduce the overfitting issue. We acquired the data samples from Alzheimer's Disease Neurological initiative (ADNI)
database.
The paper is structured as follows.
The Section one Describe about the Materials and Methods, the Section two gives the detail about the Experimental Results, the
Section three describe about the, Discussion and Finally Conclusion.
1.1 | Machine learning and deep learning based approaches
Alzheimer's classification has been a domain of active interest for scientists and researchers around the world. Alzheimer's disease impairs trans-
mission, metabolic, as well as repair mechanisms in synapses and their connections. Alzheimer's disease damages neuronal and their interconnec-
tions in memory-related areas of the brain, such as the forebrain lobes as well as initially hippocampus (Varatharajan et al., 2018). Many machine
learning-based approaches have been developed to extract useful features from the MRI images to classify AD stages. The authors in (Asl
et al., 2018) introduced a CAD system by using a support vector machine (SVM) to classify NC and AD patients. Reference (Neffati et al., 2019),
introduced a feature extraction method by using SVM and downsized kernel principal component analysis (DKPCA). During training the model on
an open-access series of imaging studies (OASIS) dataset, they achieved a 92.5% accuracy rate by using a multi support vector machine (MSVM).
The Open Access Series of Imaging Studies (OASIS) is a gathering of freely accessible neuroscience data sources for data and testing. Approxi-
mately 400 demented and non-demented people spanning the adulthood lifetime were doing in the first OASIS collection of cross-sectional MRI
data. The OASIS longitudinal collection of information is being released publicly to stimulate more research into aging and illness mechanisms, as
well as the development of better tools for researching them. MSVM's goal is to create all conceivable supervised learning given a sample number
of classes, with every classification merely researching 2 of the classrooms. MSVM is used to categorize the various traffic volume constraint. The
authors in (AlZubi et al., 2020), proposed a machine learning-based technique that can extract multivariate biomarkers from MRI images. These
biomarkers are more helpful for early diagnosis of MCI stages such as stable MCI (sMCI) and covetable MCI (cMCI). They achieved 76% classifica-
tion accuracy on AD versus CN and 66% between sMCI versus cMCI.
ABUGABAH ET AL. 3 of 13
Deep learning approaches have recently introduced more advancement to diagnose AD patients through automatic feature detection on MRI
images (Rajkumar et al., 2014). The authors in designed three models based on CNN, recurrent neural network (RNN), and long short-term mem-
ory (LSTM). In the second phase, they combined these models with the help of a weighted average approach. In the experimental processes, they
used the OASIS dataset and achieved 89.75% accuracy. The authors in (Al-Shoukry et al., 2020) presented common deep learning techniques that
are more suitable for MCI and AD classification. They also provide recommendations related to deep learning models that can help researchers
for the early detection of AD. Iman et al. introduced an automatic CAD system for the prediction of MCI conversion. That used the whole brain
voxel-wise analysis that was used for classification the MCI stages. They obtained 91.3% classification accuracy for healthy controls and AD
patients. The authors in (Li et al., 2020) designed a method where 4D deep learning algorithms were used to identify AD. Here they can apply
time-varying and spatial information consecutively. In, the authors presented a study that used a deep neural network (DNN) for early diagnosis
of dementia. A metric for how accurate a forecast is on a training sample is named F-score and it is generally referred as F1-score. It is utilized to
evaluate classifiers processes, which characterize illustrations as optimistic else negligible. On the above, DNN system obtained F1-score for MCI
0.88 and the severe dementia score is 0.94.
The authors in introduced a study for early diagnosis of AD based on a deep learning approach. They used two modalities such as fMRI and
PET images. Useful features extracted from these modalities by using CNN produced the final classification results of 99.95% on fMRI modality
and an average accuracy of 73.46% on PET images. They collected all data samples from the ADNI database. The aberrant build-up of molecules
in and among brain cells is considered towards be the source of Alzheimer's disease. Amyloid is among the genes associated, besides accumulation
of it creating plaques encompassing brain matter. The major molecule is tau, which makes tangles inside brain cells as deposition. The authors in
(El-Sappagh et al., 2020) developed a strong ensemble deep learning model based on CNN and bidirectional LSTM. The authors in (Tufail
et al., 2020), employed a 2D deep CNN model to extract the features from full brain images. Inception3 and Xception are also used in this study
through transfer learning. On the Imagenet database, Xception somewhat beats Inception v3, but massively exceeds it on a bigger object recogni-
tion raw data with 17,000 classifications. It also has the identical amount of hyperparameters as Inception, meaning higher computing perfor-
mance. In (Taheri Gorji & Kaabouch, 2019), researchers developed a CNN model for NC and MCI stage classification by using GM features that
extract from the MRI subjects, they found that the CNN model achieves accuracy rates 94.54%, 93.96%, and 93% for CN/LMCI, EMCI versus
LMCI and CN/EMCI, respectively. However, existing deep learning studies produced improved results in terms of text and medical image classifi-
cation through automatic feature detection on a given task.
2 | MATERIALS AND METHODS
2.1 | Database
All cases used in this study were obtained from the ADNI database (http://adni.loni.usc.edu/data-samples/access-data/). In the ADNI, the data-
base stored at least two years' details of patients, including different modalities such as MRI, PET and fMRI. It also has complete demographic
information of each patient. The ADNI was launched in 2004 under Dr Michael W. Weiner's supervision from the University of California. ADNI
(Jack Jr et al., 2008) project was funded as a public-private partnership with the help of 20 companies and two major foundations such as the
National Institute of Aging (NIA), National Institute of Health (NIH), and National Institute of Biomedical Imaging and Bioengineering (NIBIB). The
major objective of ADNI is to provide annotated data samples for researchers to measure the progression of MCI and AD patients. These data
samples are available in different modalities such as MRI, PET, and sMRI. All demographic details are shown in Table 1.
2.2 | Image pre-processing
We used the Z840 workstation Intel Xeon (R) E-5-2630v3 for data pre-processing and model execution. After the acquisition, MRI data from the
ADNI database were segmented using the statistical parameter mapping (SPM12) https://www.fil.ion.ucl.ac.uk/spm/. During the segmentation pro-
cess of Neuroimaging Informatics Technology Initiative format (NIFTI) subjects, the brain tissue is segmented into three major parts such as GM,
TABLE 1 Demographic details of 375 subjects used for experimental processes
Classes Subjects Age Gender male/female

NC 130 70 ± 8.5 70/60
MCI 120 71 ± 6.5 72/48
AD 125 72 ± 5.5 70/55
WM, and CSF. In our proposed study, we used WM and GM images that are closely associated with MCI progress. In the second step, we applied
the normalization operation by using the Montreal Neurological Institute space (MNI), where we used (2 2 2) mm voxel size. Finally, we applied the
smoothing operation by using full width at half maximum (FWHM). Normalization steps able to make the all images in accurate shape those able to
extract the features. After pre-processing, we convert the 3D images through MATLAB into 2D slices and resize them 224 224 using the pro-
posed model. The advantages of normalization technique is 1) Because normalization removes redundant information, a shorter dataset may be pre-
served. As a consequence, the application's total area is lowered. 2) Higher efficiency is assured, which is related to the previous statement. After
completing all pre-processing steps, we have three views of brain images, such as coronal, axial, and sagittal, as shown in Figure 1.
2.3 | Convolution neural network architecture
Convolution neural networks process information like the human brain. In CNN, artificial neurons are prearranged in layers, and each neuron is
fully connected to other neurons with the help of weight connection (Yaqub et al., 2020). Major four steps occur in deep learning models that are
based on CNN, 1)learn from data samples, 2) propagate formation, 3) calculate the error value 4) to narrow down the error value by adjusting the
weights (Kaur & Gandhi, 2020). In CNN, the first layer extracts the features from the input data, and these features are then stored in the feature
map. After each convolutional layer, one activation layer has used that changes the negative value into zero. They use the learning object's original
number of pixels as data and identify characteristics from it. Through acquiring visual attributes from tiny squares of input information, this level
assures the physical link among images. The importance of activation functions in the construction of a human brain cannot be overstated. Every
neuron has an activation function, which is a linear formula. The perceptron informs us which transistors in every level will be stimulated when it
is used. Therefore, the well-known activation function Rectified Linear Unit (ReLU) is mostly used for training a model based on CNN due to the
fast performance. The rectified linear activation functional (ReLU) is a least-square scalar quantity which, if the source is affirmative, outputs the
control signal; else, it outputs zero. The disappearing gradients issue is fixed with the rectified linear activation functional, enabling systems to
train quicker as well as increase performance.
Here, we present the general formulation of the convolutional layer for our CNN proposed model as follows:
Ih Kh þ 2Pd
Height ¼ þ 1, ð1Þ
S
where Ih, Kh denote the image height and kernel height, respectively. Pd refers to padding, and S denotes the stride value.
Iw Kw þ 2Pd
Width ¼ þ 1, ð2Þ
S
where Iw, Kw denote the image width and kernel width, respectively.
HW Fs þ 2Pd
Cn ¼ þ 1, ð3Þ
S
where the convolution layer HW refers to image height and width, Fs denotes the filter size.
FIGURE 1 The proposed model structure for the diagnosis of AD

In this research work, we used seven convolutional layers with five max-pooling layers. These pooling layers are inserted in between con-
volutional layers that help to reduce the spatial dimension. Although with only few more levels, the dimensionality of the sources in picture cate-
gorization are so large that the quantity of multiplied operations is in the billions. The source dimensionality for the subsequent stage is reduced
by aggregating the convolutional layers, which saves computations. We used 224 224 input image size in the proposed approach. We used dif-
ferent filters in convolutional layers such as 32, 64, 192, 128, 64, 256, and 128. The complete details of kernel size and pool size are shown in
Table 2. Finally, five fully connected layers are applied with a 0.2 dropout rate in the proposed model using the ReLU activation function.
A software system studies to implement categorization occupational stress from pictures, language, else audio in reinforcement learning.
Deep learning methods can obtain updated precision, although surpassing individual intelligence in certain situations. Deep learning-based models
require a large amount of annotated data samples for attaining outstanding results. To overcome this issue, we used the augmentation approach
to increase data samples (Aderghal et al., 2018). Because of less number of annotated data samples, CNN approaches to create an overfitting
issue. Image augmentation is a technology of finding alternative images in order to develop more information for the modelling training procedure.
Across other terms, it is the methodology of vastly boosting the information accessible for deep learning model development. We used the aug-
mentation technique with four different parameters, such as width shift range 0.1 degree, shear range 0.15 degree, height shift range 0.1 degree,
and channel shift 150.
3 | EXPERIMENTAL RESULTS
Our approach was implemented by using the Keras library and tensor flow on the backend. Classification of AD stages through full brain MRI
images is a difficult task. Keras is the most user-friendly software library for novices. Its simplicity aids in bringing machine learning from the realm
of fantasy to actuality. It creates a basis that may be taught in a short amount of time. You'll be capable to stack elements like a pro with Keras.
TensorFlow is a Google-developed open-source platform for running computer vision, learning techniques, and certain other analytical as well as
predicted analytical applications. Several high-level and low-level APIs are included in the architecture. In addition to the segmentation of MRI
images, the distinction between NC, MCI, and AD images is also more helpful. When we applied the segmentation on MRI images, we obtained
GM, WM, and CFG images as an output of segmentation. In the experimental process, we used GM and WM images for classification because
WM and GM are close to cognitive decline.
Our proposed diagnosis system is based on CNN because CNN plays a key role in capturing the more useful features during the training
model. In this research work, we used 375 subjects (NC = 130, MCI = 120, AD =125) for the segmentation process. There are 5 stages in the
market segmentation methodology such as segmenting prospective consumers; categorizing goods; creating a real economy matrix and calculating
TABLE 2 Detail parameters of proposed CNN approach
Layer no Layers Kernel size Stride No. of filters

1 CN1 + Re 3 1 32
Max-Pooling1 2 2
2 CN2 + Re + BN 3 1 64
3 CN3 + Re 5 1 192
Max-Pooling2 2 2
4 CN4 + Re 3 1 128
Max-Pooling3 2 2
5 CN5 + Re + BN 7 1 64
6 CN6 + Re 3 1 256
Max-Pooling4 2 2
7 CN7 + Re 5 1 128
Max-Pooling5 2 2
FC1 + Re+128 - - -
FC1 + Re+256 - - -
FC1 + Re+1024 - - -
FC1 + Re+32 - - -
FC1 + Re+1024 - - -
Softmax
Abbreviations: BN, batch normalization, CN, convolutional layer, FC, fully connected, re, ReLU.
industry proportions; selecting appropriate customers; implement business strategies for reaching customer groups. After all the pre-processing
steps, each class has 4000 scans of GM and the same as WM images. These data samples were randomly divided into five groups for training and
testing, and the details are shown in Table 3.
3.1 | Experiment 1
In this section of this study, we present the three binary class results on GM images. These classes are based on five different groups of training
and testing data samples, as shown in Table 4. In theory, a decent division to begin with is throwing 80% of the collected data group, 10% in the
classification algorithm, and 10% in the training images. The best alternative of the testing, verification, and training sets is determined by criteria
like the application scenario, empirical models, information dimensionality, and so on. Training and testing ratios introduced to check model actual
performance on any stage. These ratios also helpful to acquire the different learning rata of the model. The results are shown in three binary clas-
ses in terms of accuracy. The performance of the proposed model indicates that “NC versus AD” attained the highest accuracy of 97.94% on G4
and 96.80% on G1. The proposed technique showed relatively high accuracy on “MCI versus AD” that is, 92.84%, and the low accuracy is
91.30%. This shows that the CNN model is still rather robust. In experiment 1, the last pair of binary classification “NC versus MCI” attained
88.15% accuracy on G1 and the second highest accuracy is 88.12% on G4.
To examine the performance of binary class “NC versus AD” on five groups with different training and testing ratios, as shown in Figure 2.
The proposed technique results are promising because it provides results with accuracy values ranging from 96.76% to 97.94%. A confusion
matrix was developed that shows how well a recognition system (else multilayer perceptron) performs on a testing dataset whereby the real data
have been collected. It is straightforward, but the associated nomenclature might be perplexing. The confusion matrix is shown the complete
results of the proposed technique.
In Figure 3, we can notice that each group has shown the mismatch value to other classes. In each group, classify the true label and predicted
labels of images. According to the confusion matrix, these five groups obtained the MCI versus AD classification accuracy 92.56%, 92.25%,
91.55%, 92.84%, and 91.30%. Furthermore, in Figure 4, the confusion matrix shows the prediction images of NC versus MCI class on GM images
and obtained the classification accuracy 88.15%, 87.70%, 85.67%, 88.12% and 87.39%, respectively. If we check the percentage value according
to the confusion matrix, the G1 group produced the maximum prediction rate in terms of classification.
3.2 | Experiment 2
In this section, we apply the second experiment on WM images extract from the MRI data. We applied our proposed model on the same binary
classes that were used in experiment 1. In Table 5, it is visible that our proposed approach achieved the highest accuracy of 95.97% on “NC ver-
sus AD” by using the G4 splitting data ratio. Furthermore, 90.82% accuracy on “MCI versus AD” and 86.87% on NC versus MCI. If we check the
proposed approach, the results lie between on “NC versus AD” is 93.07% to 95.97%, “MCI versus AD” is 89.62% to 90.82%, and “NC versus
MCI” lies between 84.87% to 86.87% accuracy rate.
TABLE 3 Data splitting ratio for training and testing the model
Data splitting Group1 Group2 Group3 Group4 Group5

Training 64% 70% 76% 82% 88%
Testing 36% 30% 24% 18% 12%
TABLE 4 Performance evaluation of three binary classes by using GM images
Groups
Binary classes G1 G2 G3 G4 G5
NC versus AD 96.80% 97.16% 96.76% 97.94% 97.29%
MCI versus AD 92.56% 92.25% 91.55% 92.84% 91.30%
NC versus MCI 88.15% 87.70% 85.67% 88.12% 87.39%
Note: G1, G2, G3, G4, G5 = group number.

F I G U R E 2 Confusion matrix for binary class NC versus AD on testing GM images. (a) G1 (training 64% testing 36%); (b) G2 (training 70%
testing 30%); (c) G3 (training 76% testing 24%); (d) G5 (training 82% testing 18%); (e) G5 (training 88% testing 12%)
F I G U R E 3 Confusion matrix for binary class MCI versus AD on testing GM images. (a) G1 (training 64% testing 36%); (b) G2 (training 70%
F I G U R E 4 Confusion matrix for binary class NC versus AD on testing WM images. (a) G1 (training 64% testing 36%); (b) G2 (training 70%
TABLE 5 Performance evaluation on three binary classes by using WM images
Groups
Binary classes G1 G2 G3 G4 G5
NC versus AD 95.48% 94.87 93.07 95.97 94.58
MCI versus AD 90.69 89.62 88.89 90.76 90.82
NC versus MCI 86.69 84.87 85.04 86.24 86.87
In Figures 4, 5, and 6, we can notice that each group has shown the detailed result of testing images that classified our proposed model. If we
saw that Figure 4, it has demonstrated the NC versus AD classification confusion matrix on five groups and attained an average accuracy of
95.48%, 94.87%, 93.07%, 95.97% and 94.58%. Furthermore, we can observe in Figure 5, the confusion matrix that shows the prediction images
of MCI versus AD class on GM images and achieved the average accuracy rate on five groups 90.69%, 89.62%, 88.89%, 90.76%, and 90.82%,
respectively. If we check the percentage value according to the confusion matrix, the G5 group produced the maximum prediction rate in terms
of classification accuracy. In the last, Figure 6, show the confusion matrix of “NC versus MCI” on WM images and obtained 86.69%, 84.87%,
85.04%, 86.24% and 86.87% accuracy rate. Finally, confusion matrix are able to shown the true positive and true negative values according to the
problems.
4 | DISCUSSION
Recently, many researchers developed a CNN model for the early diagnosis of AD (Malekzadeh et al., 2021). MCI is the risk point that has a
chance for patients to convert into the AD stage within a short time. A precise and reliable diagnosis of MCI can reduce the risk of dementia
(Battineni et al., 2020). Thus, developing an accurate and consistent automatic system for NC, MCI, and AD classification plays a key role in the
aging research field. In deep learning, especially, CNNs can produce promising results in computer vision, video processing, and image processing.
The reason behind CNN's performance is that it can extract automatic features from images without any manual pre-processing (Shahzad
F I G U R E 5 Confusion matrix for binary class MCI versus AD on testing WM images. (a) G1 (training 64% testing 36%); (b) G2 (training 70%
F I G U R E 6 Confusion matrix for binary class NC versus MCI on testing WM images. (a) G1 (training 64% testing 36%); (b) G2 (training 70%
TABLE 6 Proposed model comparison for NC versus AD classification (%)
Studies Modality Model Accuracy

Liu [6] PET CNN 91.20%
Liu [36] MRI + PET CNN 93.26%
Feng [37] MRI + PET CNN 94.29%
Proposed MRI CNN 97.94% GM images
95.97%WM images
TABLE 7 Proposed model comparison for MCI versus AD classification (%)

Li [18] PET+MRI CNN 70.10%
Li [24] MRI C3d-LSTM 92.11%
90.82%WM images
TABLE 8 Proposed model comparison for NC versus MCI classification (%)

Li [18] PET+MRI CNN 77.40%
Li [24] MRI C3d-LSTM 88.12%
Liu [6] FDG + PET CNN 78.90%
86.87%WM images
et al., 2020). In previous research studies, machine learning-based approaches needed manual feature extraction, and this process required
experts. However, deep learning-based approaches can extract useful features, and some researchers only used the GM images or WM matter
images for the classification purpose (Cheng et al., 2018).
We also developed a deep CNN based automated classification system for NC, MCI, and AD. The proposed approach can successfully dis-
criminate between these classes by using GM and WM images. This model used different kernel sizes and other classifier parameters, as men-
tioned earlier in Table 2. After performance analysis of the proposed model on different binary classes by using various data splitting groups,
(Zeng et al., 2019), our proposed model produced state-of-the-art results in terms of testing accuracy. It was possible because we reduce the
overfitting issue that directly impacts on model performance. The overfitting can be reduced by ensembling, cross validation, take a break early,
increase the amount of data in the training results, regularization, and get rid of features. We achieved the best results to classify the following
pairs: NC versus AD on GM images (accuracy 97.94%), MCI versus AD on GM images (accuracy 92.84%), and NC versus MCI on GM images
(accuracy 88.15%). Furthermore, the same binary groups achieved the accuracy of WM images (95.97%, 90.82%, and 86.87%).
From the results in Table 6, we can observe that three studies investigated AD patients' and normal control. In these studies, they applied dif-
ferent modalities with the CNN approach and other feature extraction techniques. Using the image retrieval technique, we may form novel char-
acteristics that are a linear mixture of accessible functionalities. Whenever opposed towards the earlier extracted features, the newer number of
characteristics will have positive variations. The main target is to utilize limited functionality to collect the identical amount of information. Fur-
thermore, they also have a different splitting ratio for training and validation subjects that are acquired from the ADNI database. For NC versus
AD classification, Feng et al. achieved the highest accuracy of 94.29%, but our proposed model attained 97.94%. Table 7 has demonstrated the
comparison of mild cognitive impairment and AD patients. Li et al. reached the 92.11% highest accuracy in this comparison, and our proposed
n et al., 2018), At last, we discussed a comparison of normal control and mild
model achieved 92.84% accuracy, as shown in Table 7. (Cabrera-Leo
cognitive impairment, as shown in Table 8. We noticed that our proposed model produced promising results and achieved the highest accuracy
rate, 88.15%. Our CNN technique is more reliable and robust for these three binary classes than that used in this research work.
5 | C O N CL U S I O N
This research work proposed a unique CNN method for the accurate classification of NC, MCI, and AD images. In the experimental process, we
used two major brain tissues (GM and WM) as inputs obtained after the segmentation of MRI subjects. In the proposed approach, we used
350 subjects for training and testing; these subjects were obtained from the ADNI database. The experimental results are attained from the five
different splitting ratio groups on GM and WM images. Our proposed model achieved the best classification results on NC versus AD groups in
the GM and WM images. The pair of MCI versus AD has attained better accuracy than NC versus MCI concerning both MRI tissues. Our CNN
approach yielded 97.94% classification accuracy on GM images for NC/AD, 92.84% for MCI/AD, and 88.15% classification accuracy for pairs of
NC/MCI. These all results mentioned above are achieved from the GM images. All experimental results of three class pairs on five data splitting
groups demonstrated that our proposed model achieved promising results in terms of classification.
ACKNOWLEDGEMEN TS
Data collection and sharing for this project was funded by the Alzheimer's Disease Neuroimaging Initiative (ADNI) (National Institutes of Health
Grant U01 AG024904) and DOD ADNI (Department of Defence award number W81XWH-12-2-0012). ADNI is funded by the National Institute
on Aging, the National Institute of Biomedical Imaging and Bioengineering, and through generous contributions from the following: AbbVie,
Alzheimer's Association; Alzheimer's Drug Discovery Foundation; Araclon Biotech; BioClinica, Inc.; Biogen; Bristol-Myers Squibb Company; Cer-
eSpir, Inc.; Cogstate; Eisai Inc.; Elan Pharmaceuticals, Inc.; Eli Lilly and Company; EuroImmun; F. Hoffmann-La Roche Ltd and its affiliated company
Genentech, Inc.; Fujirebio; GE Healthcare; IXICO Ltd.; Janssen Alzheimer Immunotherapy Research & Development, LLC.; Johnson & Johnson
Pharmaceutical Research & Development LLC.; Lumosity; Lundbeck; Merck & Co., Inc.; Meso Scale Diagnostics, LLC.; NeuroRx Research; Neu-
rotrack Technologies; Novartis Pharmaceuticals Corporation; Pfizer Inc.; Piramal Imaging; Servier; Takeda Pharmaceutical Company; and Transi-
tion Therapeutics. The Canadian Institutes of Health Research is providing funds to support ADNI clinical sites in Canada. Private sector
contributions are facilitated by the Foundation for the National Institutes of Health (www.fnih.org). The grantee organization is the Northern Cali-
fornia Institute for Research and Education, and the study is coordinated by the Alzheimer's Therapeutic Research Institute at the University of
Southern California. ADNI data are disseminated by the Laboratory for Neuro Imaging at the University of Southern California.
CONF LICT OF IN TE RE ST
The authors declare that they have no conflicts of interest to report regarding the present study.
INFORMED CONSENT
This research does not involve humans and/or animals and informed consent is not applicable.
E TH I CS S T A TE M E N T
No study-specific approval by any authority for this research.
DATA AVAI LAB ILITY S TATEMENT

Data used in the preparation of this article were obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database (adni.loni.usc.
edu). As such, the investigators within the ADNI contributed to the design and implementation of ADNI and/or provided data but did not partici-
pate in the analysis or writing of this report. A complete listing of ADNI investigators can be found at: http://adni.loni.usc.edu/wpcontent/
uploads/how_to_apply/ADNI_Acknowledgement_List.pdf.
ORCID
Ahed Abugabah https://orcid.org/0000-0003-3141-193X
RE FE R ENC E S
Aderghal, K., Khvostikov, A., Krylov, A., Benois-Pineau, J., Afdel, K., & Catheline, G. (2018). Classification of Alzheimer disease on imaging modalities with
deep CNNs using cross-modal transfer learning. Paper presented at 2018 IEEE 31st international symposium on computer-based medical systems
(CBMS). (pp. 345–350). IEEE.
Al Smadi, A., & Abugabah, A. (2018). Intelligent information systems and image processing: A novel pan-sharpening technique based on multiscale decom-
position. Paper presented at Proceedings of the 2018 the 2nd International Conference on Video and Image Processing. (pp. 208–212).
Alarifi, A., AlZubi, A. A., Al-Maitah, M., & Al-Kasasbeh, B. (2019). An optimal sensor placement algorithm (O-SPA) for improving tracking precision of human
activity in real-world healthcare systems. Computer Communications, 148, 9–16.
Alsharif, H., Mohammed, K. H., Anabi, Y. K., & Shehzad, C. A. (2020). Machine learning algorithms for smart data analysis in internet of things environment:
Taxonomies and research trends. Symmetry, 12(1), 88.
Al-Shoukry, S., Rassem, T. H., & Makbol, N. M. (2020). Alzheimer's diseases detection by using deep learning algorithms: A mini-review. IEEE Access, 8,
77131–77141.
Alsmadi, A., Yang, S., & Zhang, K. (2018). Pansharpening via deep guided filtering network. International Journal of Image Processing and Visual Communica-
tion, 5, 1–8.
AlZubi, A. A. (2020). Big data analytic diabetics using map reduce and classification techniques. The Journal of Supercomputing, 76(6), 4328–4337.
AlZubi, A. A., Alarifi, A., & Al-Maitah, M. (2020). Deep brain simulation wearable IoT sensor device based Parkinson brain disorder detection using heuristic
tubu optimized sequence modular neural network. Measurement, 161, 107887.
Asl, E. H., Ghazal, M., Mahmoud, A., Aslantas, A., Shalaby, A. M., et al. (2018). Alzheimer's disease diagnosicts by a 3D deeply supervised adaptable con-
volutional network. Frontiers of Biosciences – Landmark, 23(3), 584–596.
Bastin, C., Giacomelli, F., Miévis, F., Lemaire, C., Guillaume, B., & Salmon, E. (2021). Anosognosia in mild cognitive impairment: Lack of awareness of mem-
ory difficulties characterizes prodromal Alzheimer's disease. Frontiers in Psychiatry, 12, 404.
Battineni, G., Sagaro, G. G., Chinatalapudi, N., & Amenta, F. (2020). Applications of machine learning predictive models in the chronic disease diagnosis.
Journal of Personalized Medicine, 10(2), 21.
Beheshti, I., Mahdipour Hossein-Abad, H., & Matsuda, H. (2019). Identification of Alzheimer's disease on the basis of a voxel-wise approach. Applied Sci-
ences, 9(15), 3063.
Bharati, P., & Pramanik, A. (2020). Deep learning techniques—R-CNN to mask R-CNN: A survey. In Computational Intelligence in Pattern Recognition
(pp. 657–668). Springer.
Cabrera-Leo n, Y., Báez, P. G., Ruiz-Alzola, J., & Suárez-Araujo, C. P. (2018). Classification of mild cognitive impairment stages using machine learning
methods. Paper presented at 2018 IEEE 22nd International Conference on Intelligent Engineering Systems (INES). (pp. 67–72). IEEE.
Cheng, C. K., Tsao, Y. C., Su, Y. C., Sung, F. C., Tai, H. C., & Kung, W. M. (2018). Metabolic risk factors of Alzheimer's disease, dementia with Lewy bodies,
and normal elderly: A population-based study. Behavioural Neurology, 2018, (pp. 1-8).
Dua, M., Makhija, D., Manasa, P. Y. L., & Mishra, P. (2020). A CNN–RNN–LSTM based amalgamation for Alzheimer's disease detection. Journal of Medical
and Biological Engineering, 40(5), 688–706.
El-Sappagh, S., Abuhmed, T., Islam, S. R., & Kwak, K. S. (2020). Multimodal multitask deep learning model for Alzheimer's disease progression detection
based on time series data. Neurocomputing, 412, 197–215.
Gao, Y., Huang, C., Zhao, K., Ma, L., Qiu, X., et al. (2015). Depression as a risk factor for dementia and mild cognitive impairment: A metal – Analysis of lon-
gitudinal studies. International Journal of Geriatric Psychiatry, 28(5), 441–449.
Ge, C., Qu, Q., Gu, I. Y. H., & Jakola, A. S. (2019). Multi-stream multi-scale deep convolutional networks for Alzheimer's disease detection using MR images.
Neurocomputing, 350, 60–69.
Jack, C. R., Jr., Bernstein, M. A., Fox, N. C., Thompson, P., Alexander, G., Harvey, D.,Weiner, M. W. (2008). The Alzheimer's disease neuroimaging initiative
(ADNI): MRI methods. Journal of Magnetic Resonance Imaging: An Official Journal of the International Society for Magnetic Resonance in Medicine, 27(4),
685–691.
Janghel, R. R., & Rathore, Y. K. (2021). Deep convolution neural network based system for early diagnosis of Alzheimer's disease. Irbm, 42(4), 258–267.
Kaur, T., & Gandhi, T. K. (2020). Deep convolutional neural networks with transfer learning for automated brain image classification. Machine Vision and
Applications, 31(3), 1–16.
Kim, J. P., Kim, J., Park, Y. H., Park, S. B., San Lee, J., Yoo, S., … Seong, J. K. (2019). Machine learning based hierarchical classification of frontotemporal
dementia and Alzheimer's disease. NeuroImage: Clinical, 23, 101811.
Li, W., Lin, X., & Chen, X. (2020). Detecting Alzheimer's disease based on 4D fMRI: An exploration under deep learning framework. Neurocomputing, 388,
280–287.
Malekzadeh, A., Zare, A., Yaghoobi, M., Kobravi, H. R., & Alizadehsani, R. (2021). Epileptic seizures detection in EEG signals using fusion handcrafted and
deep learning features. Sensors, 21(22), 7710.
Mehmood, A., Maqsood, M., Bashir, M., & Shuyuan, Y. (2020). A deep siamese convolution neural network for multi-class classification of alzheimer dis-
ease. Brain Sciences, 10(2), 84.
Munadhil, Z., Sadik, K. G., Ammar, M. H., Ali, A., et al. (2020). Neural network-based Alzheimer's patient localization for wireless sensor network in an
indoor environment. IEEE Access, 8, 150527–150538.
Neffati, S., Abdellafou, K. B., Jaffel, I., Taouali, O., & Bouzrara, K. (2019). An improved machine learning technique based on downsized KPCA for
alzheimer's disease classification. International Journal of Imaging Systems and Technology, 29(2), 121–131.
Rajkumar, S., Bardhan, P., Akkireddy, S. K., & Munshi, C. (2014). CT and MRI image fusion based on wavelet transform and neuro-fuzzy concepts with
quantitative analysis. Paper presented at 2014 International conference on electronics and communication systems (ICECS). (pp. 1–6). IEEE.
Ramzan, F., Khan, M. U. G., Rehmat, A., Iqbal, S., Saba, T., Rehman, A., & Mehmood, Z. (2020). A deep learning approach for automated diagnosis and multi-
class classification of Alzheimer's disease stages using resting-state fMRI and residual neural networks. Journal of Medical Systems, 44(2), 1–16.
Shahzad, M. N., Suleman, M., Ahmed, M. A., Riaz, A., & Fatima, K. (2020). Identifying the symptom severity in obsessive-compulsive disorder for classifica-
tion and prediction: An artificial neural network approach. Behavioural Neurology, 22, (pp 1-7).
Taheri Gorji, H., & Kaabouch, N. (2019). A deep learning approach for diagnosis of mild cognitive impairment based on MRI images. Brain Sciences,
9(9), 217.
Tufail, A. B., Ma, Y. K., & Zhang, Q. N. (2020). Binary classification of Alzheimer's disease using sMRI imaging modality and deep learning. Journal of Digital
Imaging, 33(5), 1073–1090.
Varatharajan, R., Manogaran, G., Priyan, M. K., & Sundarasekar, R. (2018). Wearable sensor devices for early detection of Alzheimer disease using dynamic
time warping algorithm. Cluster Computing, 21(1), 681–690.
Wang, M., Yan, Z., Xiao, S. Y., Zuo, C., & Jiang, J. (2020). A novel metabolic connectome method to predict progression to mild cognitive impairment. Behav-
ioural Neurology, 2020, (pp 1-9).
Yaqub, M., Feng, J., Zia, M. S., Arshid, K., Jia, K., Rehman, Z. U., & Mehmood, A. (2020). State-of-the-art CNN optimizer for brain tumor segmentation in
magnetic resonance images. Brain Sciences, 10(7), 427.
Zeng, K., Yang, Y., Xiao, G., & Chen, Z. (2019). A very deep densely connected network for compressed sensing MRI. IEEE Access, 7, 85430–85439.
Zhang, D., Wang, Y., Zhou, L., Yuan, H., Shen, D., & Initiative, A.'s. D. N. (2011). Multimodal classification of Alzheimer's disease and mild cognitive impair-
ment. NeuroImage, 55(3), 856–867.
Zhu, F., Li, X., Mcgonigle, D., Tang, H., He, Z., Zhang, C., … Zhou, W. (2019). Analyze informant-based questionnaire for the early diagnosis of senile demen-
tia using deep learning. IEEE Journal of Translational Engineering in Health and Medicine, 8, 1–6.
AUTHOR BIOGRAPHI ES
Ahed Abugabah is a Professor in Information Systems. He currently works at the College of Technological Innovation at Zayed University. He
worked in higher education in Australia where he received his degrees in information systems. His research interests include Information Sys-
tems, Machine Learning & Data Mining in Health care, and RFID in Health care.
Atif Mehmood received his Ph.D. degree in computer science and technology from the School of Artificial Intelligence, Xidian University,
China 2021. He currently works as Assistant Professor atNUML University Islamabad, His research interests include machine learning and
medical image processing.
Sultan Almotairi currently is an assistant professor in Community College of Majmaah University received his degrees in computer science
from Florida Institute of Technology, USA, His research interests include neural networks, deep learning, pattern recognition, machine learn-
ing, image processing, and computer vision.
Ahmad Al Smadi received the B.S. degree in computer engineering from Yarmouk University, Irbid, Jordan, in 2014, and the M.S. degrees in
intelligent information processing from Xidian University, Xi'an, China, in 2018. He is currently pursuing the Ph.D. degree in computer science
and technology with the School of Artificial Intelligence, Xidian University, Xi'an, China. His research interests include in machine learning and
image processing.
How to cite this article: Abugabah, A., Mehmood, A., Almotairi, S., & Smadi, A. A. L. (2022). Health care intelligent system: A neural
network based method for early diagnosis of Alzheimer's disease using MRI images. Expert Systems, 39(9), e13003. https://doi.org/10.
1111/exsy.13003

Expert Systems - 2022 - Abugabah - Health Care Intelligent System A Neural Netw

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Expert Systems - 2022 - Abugabah - Health Care Intelligent System A Neural Netw

Uploaded by

Copyright:

Available Formats

14680394, 2022, 9, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/exsy.13003 by <Shibboleth>-member@uel.ac.uk, Wiley Online Library on [23/02/2023].

Health care intelligent system: A neural network based method

Ahed Abugabah1 | Atif Mehmood2 | Sultan Almotairi3 | Ahmad A. L. Smadi2

1.1 | Machine learning and deep learning based approaches

2 | MATERIALS AND METHODS

2.2 | Image pre-processing

TABLE 1 Demographic details of 375 subjects used for experimental processes

Classes Subjects Age Gender male/female

2.3 | Convolution neural network architecture

FIGURE 1 The proposed model structure for the diagnosis of AD

TABLE 2 Detail parameters of proposed CNN approach

Layer no Layers Kernel size Stride No. of filters

Data splitting Group1 Group2 Group3 Group4 Group5

TABLE 4 Performance evaluation of three binary classes by using GM images

Note: G1, G2, G3, G4, G5 = group number.

TABLE 5 Performance evaluation on three binary classes by using WM images

TABLE 6 Proposed model comparison for NC versus AD classification (%)

Studies Modality Model Accuracy

TABLE 7 Proposed model comparison for MCI versus AD classification (%)

Studies Modality Model Accuracy

TABLE 8 Proposed model comparison for NC versus MCI classification (%)

Studies Modality Model Accuracy

DATA AVAI LAB ILITY S TATEMENT

You might also like