Received February 5, 2021, accepted February 12, 2021, date of publication February 23, 2021, date of current version

March 3, 2021.
Digital Object Identifier 10.1109/ACCESS.2021.3061487

An Efficient Classification of MRI Brain Images

MUHAMMAD ASSAM 1 , HIRA KANWAL 2 , UMAR FAROOQ3 , SAID KHALID SHAH3 ,
ARIF MEHMOOD 2 , AND GYU SANG CHOI 4
1 Collegeof Computer Science and Technology, Zhejiang University, Hangzhou 310027, China
2 Department of Computer Science and Information Technology, The Islamia University of Bahawalpur, Bahawalpur 63100, Pakistan
3 Department of Computer Science, University of Science and Technology (UST) at Bannu, Bannu 28100, Pakistan
4 Department of Information and Communication Engineering, Yeungnam University, Gyeongsan 38542, South Korea

Corresponding authors: Muhammad Assam ([email protected]), Hira Kanwal ([email protected]), Arif Mehmood
([email protected]), and Gyu Sang Choi ([email protected])
This work was supported in part by the Basic Science Research Program through the National Research Foundation of Korea (NRF)
funded by the Ministry of Education under Grant NRF-2019R1A2C1006159, and in part by the Ministry of Science and ICT (MSIT),
South Korea, under the Information Technology Research Center (ITRC) support program, supervised by the Institute for Information and
Communications Technology Promotion (IITP), under Grant IITP-2020-2016-0-00313.

ABSTRACT The unprecedented improvements in computing capabilities and the introduction of advanced
techniques for the analysis, interpretation, processing, and visualization of images have greatly diversified
the domain of medical sciences and resulted in the field of medical imaging. The Magnetic Resonance
Imaging (MRI), an advanced imaging technique, is capable of producing high quality images of the human
body including the brain for diagnosis purposes. This paper proposes a simple but efficient solution for the
classification of MRI brain images into normal, and abnormal images containing disorders and injuries.
It uses images with brain tumor, acute stroke and alzheimer, besides normal images, from the public dataset
developed by harvard medical school, for evaluation purposes. The proposed model is a four step process,
in which the steps are named: 1). Pre-processing, 2). Features Extraction, 3). Features Reduction, and 4).
Classification. Median filter, being one of the best algorithms, is used for the removal of noise such as salt
and pepper, and unwanted components such as scalp and skull, in the pre-processing step. During this stage,
the images are converted from gray scale to colored images for further processing. In second step, it uses
Discrete Wavelet Transform (DWT) technique to extract different features from the images. In third stage,
Color Moments (CMs) are used to reduce the number of features and get an optimal set of characteristics.
Images with the optimal set of features are passed to different classifiers for the classification of images.
The Feed Forward - ANN (FF-ANN), an individual classifier, which was given a 65% to 35% split ratio for
training and testing, and hybrid classifiers called: Random Subspace with Random Forest (RSwithRF) and
Random Subspace with Bayesian Network (RSwithBN), which used 10-Fold cross validation technique,
resulted in 95.83%, 97.14% and 95.71% accurate classification, in corresponding order. These promising
results show that the proposed method is robust and efficient, in comparison with, existing classification
methods in terms of accuracy with smaller number of optimal features.

INDEX TERMS Color moments (CMs), feed forward artificial neural network (FF-ANN), random subspace,
random forest, bays.net, principle component analysis (PCA), discrete wavelet transforms (DWT).

I. INTRODUCTION properties of cerebrum MRI get the consideration of scientists

T he robotized classification of images, obtained from Mag- and clinicians.
netic Resonance Imaging (MRI), is a critical procedure and Brain MRI provides better results, when it is contrasted
for this reason, a number of classification strategies are devel- with other imaging modalities such as Computed Tomog-
oped in the most recent decades. It plays a vital role in raphy (CT), Positron Emanation Tomography (PET), where
analyzing and examining human mind. Brain MRI has sig- delicate tissue outline is important. Manual review of cere-
nificantly enhanced the findings and treatments of cerebrum brum MRI is a hectic job due to huge amount of information,
pathology due to rich data, it produces, about the delicate it contains. To overcome this issue, automatic methods are
tissue life structures. The non-obtrusive and torment free introduced for the examination of brain MRI images [1]–[4].
The cerebrum MRI is an imperative route for recogniz-
ing sound brains, and brains having distinctive mind sick-
The associate editor coordinating the review of this manuscript and nesses such as cerebrum tumor, Alzheimer, and stroke. The
approving it for publication was Sotirios Goudos . standard classification model contains four phases which

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
VOLUME 9, 2021 33313
 M. Assam et al.: Efficient Classification of MRI Brain Images

are: preprocessing, features extraction, features reduction and HG and LG GBMs. The average accuracy achieved, for this
classification of brain MRI images [1], [5]. The preprocessing work, was 90%. In short, the results produced using both
is the simplest phase among all phases of the classification techniques were good but in term of accuracy, the perfor-
model. In preprocessing stage, a noise removal algorithm is mance of supervised techniques was better than unsupervised
used for the removal of salt-and-pepper noise, and unwanted techniques.
components such as scalp and skull from the images. Due to This research work proposes quite a useful technique to
the removal of noise from images, their quality is improved. automate the time-consuming manual procedure of physi-
In this stage, the images are also converted from gray scale to cians, and provides a clear and effective methodology for the
color (RGB) images and, thus, the utilisation of rich informa- classification of MRI images into sound images and abnormal
tion in colored images increase classification accuracy. For images containing disturbances and injuries. The proposed
the removal of noise from images, a number of algorithms technique is using a four step process, in which, an image
are used [6], in which the median filter has performed better passes through four stages termed as: Preprocessing, Features
for the removal of salt and pepper noise from images. It is Extraction, Features Reduction and Classification. Median
also better because of not distressing the edges of images [7]. filter is used in preprocessing stage for the removal of noise
Feature extraction stage, which is followed by the pre- such as salt-and-pepper, and unwanted components such as
processing, is not only important but also a difficult task scalp and skull. These images are, then, converted from gray
[8], [9], in which the format of an image is changed to a scale to color (RGB) images for further processing. Discrete
set of features. Images contain a lot of features but most of Wavelet Transform (DWT) is applied for features extraction
them are redundant, which are not useful for classification. in second stage. Color Moments (CMs) are used to reduce and
The most monotonous task is the selection of an optimal set select the optimal set of features. These optimal features are
of features. A number of techniques are available to extract sent to FF-ANN using percentage split and hybrid classifiers
features from images, in which Discrete Wavelet Transform named: Random Subspace with Random Forest (RSwithRF)
(DWT) [1], [10]–[12], Principle Component Analysis (PCA), and Random Subspace with Bayesian Network (RSwithBN)
Independent Component Analysis (ICA), gabor features, and using 10-Fold cross validation for recognizing sound brains
minimum noise fraction transform [13]–[16] are the most and brains with distinctive sicknesses.
widely used. When features are extracted, irrelevant features The existing methods use a large set of parameters for
need to be reduced as they increase compilation time and classification purposes, which greatly increase their complex-
memory usage. In this stage, those features are selected that ity in terms of space and time. The current studies, espe-
are optimal and useful. Several algorithms such as Genetic cially those using hybrid classifiers, are carried out using T1
Algorithm (GA) [17], PCA [1], [18], [19], ICA [20] and weighted images. The basic aim of the this research is to
Linear Discriminant Analysis (LDA) [21] are used for dimen- develop a fast and efficient method that determines and uses
sionality reduction. This reduced set of features is used in the a small set of optimal parameters. This work has used hybrid
last stage for classification. There are two broad categories classifiers for an improved accuracy. Furthermore, this work
for the classification of MRI Brain images called: supervised is based on T2 weighted images. The main contribution of
and unsupervised techniques. Supervised techniques include this work is as follows:
K Nearest Neighbors (KNN) and Support Vector Machine
(SVM) [11], [18], [22], [23] while unsupervised techniques 1) classification of images with an accuracy almost the
include Fuzzy C mean, and s-means [3], [23]. same or even slightly better than other classifiers and
Gupta et al. [24] suggested in their research work, a non- that too with a set of only nine parameters.
invasive system for the detection of brain glioma. The tex- 2) reduction the complexity of the proposed method com-
ture and morphological features with ensemble learning were pared with other techniques as it processes each image
used for detection purpose. Promising results were achieved against a limited set of features.
which are 97.37% on JMCD and 98.38% on BraTS. Arasi et
al. [25] proposed a clinical support system for improving the
accuracy of detection and classification of brain tumor from II. RELATED WORK
the BraTS dataset using images. The GLCM extraction tech- Zhang et al. in [1] proposed a new method for the classifica-
nique was used for collecting features of tumor region and tion of T2-weighted brain MRI images that consists of three
LOBSVM was used for classification purpose. It achieved an stages. In the first stage (feature extraction), it extracts 1024
average accuracy of 97.69%. Ullah et al. [26] proposed an features from each image by using DWT. These features are
enhanced technique for classification of brain MRI images reduced to 19 features, in stage 2 using PCA. The reduced
into normal and abnormal images using color features and set of 19 features is, then, feed to an ANN classifier in
Artificial Neural Network (ANN). It achieved an accuracy third stage, for the classification. It achieved good results in
of 100% and 90% during training and testing phases, corre- terms of accuracy. Rajini et al. in [23] proposed an auto-
spondingly. Jeyong et al. [27] used machine learning method mated approach to classify brain MRI images into normal
on DSC-MR images which are based on delta-radiomic fea- and abnormal images, in a two stage process. The first stage
tures. The proposed algorithm was used for classification of extracts features using DWT, which are, then, reduced to an

33314 VOLUME 9, 2021

M. Assam et al.: Efficient Classification of MRI Brain Images

optimal features’ set using PCA. The second stage performs

classification using two classifiers with an accuracy of 94.5%.
Othman et al. [28] have used SVM to differentiate between
the normal and abnormal MRI images. The technique, they
presented, involves inputting brain MRI dataset, the imple-
mentation of wavelet-based feature extraction, and, then,
classification using SVM. The images in the image set
include T2 flair weighted images having a resolution of
256 × 256 pixels. This work considered a total of 32 images,
in which 22 images were normal while 10 were abnormal.
The images are passed to SVM classifier for classification
after the successful extraction. Saritha et al. [29] proposed
a new technique for the classification of MRI brain images,
which used assimilating wavelet entropy-based spider web
plots and probabilistic neural network (PNN). In the first step,
wavelet entropy-based spider web plots is used for feature
extraction while PNN is used for classification, in the second
stage. For the evaluation purposes, they used 75 T2 weighted
images, each having a resolution of 256 × 256 pixels. PNN is
a good technique for the classification of patterns and, there-
fore, the proposed classifier response in terms of accuracy
was good.
Lahmiri et al. [30] proposed an automated approach to FIGURE 1. The research model adopted for this work.
classify healthy images and those which are affected by dis-
eases such as Glioma, Tumor and Alzheimer. DWT is used
to extract optimal features from sub-bands LH and HL. The Wahid et al. [33] also presented a three stage automated
brain MRI images are decomposed and features are extracted model for classification. In the first stage (pre-processing),
in both horizontal and vertical orders, respectively. Three noise is removed from the images. Two types of features:
classifiers PNN, KNN, Learning Vector Quantization (LVQ), color moments and texture are extracted in the second stage,
combined in a single SVM, were used for the classification which are classified using probabilistic classifier. The clas-
to improve the precision and effectiveness. They verified this sifiers used were based on logistic function, and a total
technique on a dataset taken from Harvard medical college. of 150 images were taken into consideration, in which 66%
This combined approach produced fruitful results. were used to train the model and 34% were used to test the
Nandpuru et al. in [31] proposed a robotized technique model. Overall accuracy achieved was 90.66%.
to differentiate between effected and healthy MRI images. The rest of this paper is organised as follows. Section III
Median filter was used for the removal of salt-and-pepper describes the Materials and Methods. The proposed mech-
noise, and unwanted components such as scalp and skull. The anism is presented in section IV. Section V provides the
images quality was improved by reducing noise. It extracted evaluation results. Conclusion is presented in section VI.
four kind of features, which are: power law transformation,
texture, symmetrical and gray scale features, respectively. III. MATERIALS AND METHODS
PCA is used to reduce these features to an optimal set of One real world dataset is used in proposed work, which was
features, which are, then, classified using SVM in the clas- taken from the Harvard Medical School [34]. The proposed
sification phase. For assessment purposes, they used Linear model (presented in Figure. 1) used in the research have
Kernels (LKs), Quadratic Kernels (QKs) and Polynomial four (4) phases which are pre-processing, features extraction,
Kernels (PKs), whose accuracy was 74%, 84%, and 76%, features reduction and classification. Features are extracted
respectively. in feature extraction stage from brain MRI images for their
Kalbkhani et al. [32] suggested a three stage technique potential use in characterizing them, which are, then, used
to categorize normal and abnormal MRI brain images. 2-D in our research work. Recent studies suggested that ANN
discrete wavelet transform is used, in first stage, for fea- and hybrid classification techniques are most suitable meth-
tures’ extraction. To select optimal and efficient features, the ods for classification due to their high accuracy rates. This
multi-cluster feature selection method is used. It reduced the article presents our investigations for the classification of
initial set of features to 41, which is forwarded to the next MRI brain images on well-known classification techniques
stage for classification. The researchers used multi-cluster namely: the Feed Forward - ANN (FF-ANN), Random Sub-
features and KNN to classify healthy and those images that space with Random Forest (RSwithRF) and Random Sub-
contain injuries and disorders. The classification accuracy space with Bayesian Network (RSwithBN). To the best of our
was found good when compared with state of art techniques. knowledge, no classification technique has been used with

VOLUME 9, 2021 33315

 M. Assam et al.: Efficient Classification of MRI Brain Images

B. FEATURES EXTRACTION
To achieve higher level of accuracy in the classification stage,
it is essential to select an optimal set of features in the feature
extraction stage. DWT is one the most powerful mathematical
tool, which uses dyadic scales and positions, and it imple-
ments wavelet transform [22], [28]. To extract features from
MRI brain images, DWT technique is used, in this work.
DWT not only offers knowledge about time but also the
frequency domain. Basic introduction of DWT is presented
next.
Let suppose the square-integral function will be x(t) the
wavelet transform which is continuous of x(t) relative to the
given wavelet 9c,d (t) is defined in Equation. 1.
R∞
W 9(c,d) = −∞ x(t) × 9c,d (t) dx (1)

where,
√
9c,d (t) = 1/ c9(t − c/d) (2)

Variables ‘c’ and ‘d’ are positive real numbers in

Equation 2. Using translation and dilation, the wavelet
9c,d (t) is computed from mother wavelet 9(t) where ‘c’
represent the dilation factor and ‘d’ represent the transla-
FIGURE 2. Illustrating pre-processing of images. The image in: a). gray tion parameter. The simplest and mostly used wavelet for
scale before processing, and b). Color (RBG) form after processing.
image processing is haar [35]. To decompose an image into
sub-hands with their relative DWT co-efficient, the two cas-
cading low and high pass filters of DWT technique, generally
an optimal set of only nine features. For the classification used to satisfy specific constraints, were used.
of MRI brain images into normal and abnormal groups, the
Caj,k (n) = DS[ x(n)g ×j(n-2j k)]
P
proposed model and the classification techniques that it uses, (3)
have proved best in terms of accuracy compared with similar P
Cdj,k (n) = DS[ x(n)h ×j(n-2 k)] j (4)
techniques found in the Literature.
Caj,k and Cdj,k in Equation 3 and 4 are approximation
A. PREPROCESSING and detailed component coefficients while high and low pass
Though, MRI produces high quality images, however, the filters are denoted by h(n) and g(n), correspondingly. The
images include unwanted components such as scalp and variables ‘j’ and ‘k’ denote wavelet scale and transition factor
skull and they might have noise due to the negligence of in these equations. The operator DS (↓) shows down sam-
operators. To improve the accuracy of the proposed tech- pling. The overall process is called wavelet decomposition
nique, it is necessary that the images should be not only tree, which is illustrated in Figure 3.
sharp but free of unwanted components and noise. In the DWT is applied on every facet of the images separately.
proposed work, median filter is used in the pre-processing Figure 4 produces the results of four sub-bands (namely the
stage for the removal of salt-pepper noise, and scalp and HH, HL, LL, and LH) of images on each scale. The LL
skull without affecting the edges of brain MRI images. In sub-band shows the approximation factor while the rest of
this stage, the given images are also converted from gray the sub-bands indicate the detailed element of an MRI brain
scale to colored (RGB) images that provide rich informa- image. An MRI image can be decomposed up-to several
tion. This work used a 3 × 3 mask for condensing com- levels to get more compact approximation factor. It should
putation time because of a small size window [17]. After be kept in mind that, if, we intend to increase the influence of
successful execution of first phase, the images achieved decomposition levels, we should increase the decomposition
are free of unwanted components and noise, which are levels to an appropriate level.
converted to color (RGB) images. The reason behind this The proposed work used haar wavelet to decompose the
conversion is the fact that images in color (RGB) form images up to three levels to extract features from the MRI
contains rich information as compared to images in gray images. Figure 5 illustrates the layout of the complete set of
scale. Figure 2 illustrates the pre-processing stage of images. sub bands up-to three levels.
Figure 2 (a) presents the image in gray scale before process- The left hand side of Figure 6 shows a color (RGB) image,
ing while Figure 2 (b) is in converted color (RBG) form after which is converted from a gray scale image. The size of
processing. this colored (RGB) image is 256 × 256×3 which is too

33316 VOLUME 9, 2021

M. Assam et al.: Efficient Classification of MRI Brain Images

Color Moments (CMs) are applied to an approximate coeffi-

cient at level 3.
C. FEATURES REDUCTION
The proposed model, at this stage, proposed using quite a
useful technique for features reduction. Since, 3072 features
extracted in features extraction stage were quite large and
intense for computation and classification, we used CMs
to greatly reduce the features set without the information.
This reduction was important as a set of large number of
features not only increases computation overhead but also
memory usage. The proposed method extracted Red, Blue
and Green channels from the converted color (RGB) images.
For each channel, the values for Mean, Skewness and Stan-
dard deviation (variance) are calculated, which are very use-
ful in classification. An optimal set of only nine features is
FIGURE 3. A 3-Level decomposition tree. S represents the root image obtained, in this stage, which represents a complete image
while ca and cd having 1, 2, and 3 represents the approximation and used for an increased classification accuracy while reducing
detailed component coefficients at level 1, 2, and 3.
computation overhead and complexity [33]. The equations,
given below, mathematically illustrate the Mean, Variance,
and Skewness for the Red, Green and Blue channels, respec-
tively. Equation 5, 6, and 7 represent the Mean, Variance and
Skewness for the Red, Equation 8, 9, and 10 for the Green,
while Equation 11, 12, and 13 represent them for the Blue
channel.

M1,1 = 1/N N
P
j=1 Ij (5)
√ PN 2
M1,2 = 2 j=1 (tj - M1,1 ) (6)
FIGURE 4. Diagram of DWT in 2D.
PN 3
M1,3 = 1/N j=1 (Ij - M1,1 ) (7)
PN
M2,1 = 1/N j=1 Ij (8)
√ PN 2
M2,2 = 2 j=1 (tj - M2,1 ) (9)
PN
M2,3 = 1/N j=1 (Ij - M2,1 )3 (10)
PN
M3,1 = 1/N j=1 Ij (11)
√ PN 2
M3,2 = 2 j=1 (tj - M3,1 ) (12)
PN 3
M3,3 = 1/N j=1 (Ij - M3,1 ) (13)

The variables ‘I’ and ‘N’, in the above equations, denote

the intensity and total number of pixels, respectively. These
equations compute three color channels (Red, Green and
Blue) from the color (RGB) images at level 3 using approxi-
mate coefficients presented in Figure 6. The Mean, Variance
FIGURE 5. Layout of the wavelet sub-bands.
and Skewness are, then, calculated for the three channels.
The ultimate set of these nine features is stored in a one
huge for computation, so it should be reduced and com- dimensional array, which are used by the classifiers such as
pressed without loss of optimal information. The proposed FF-ANN and hybrid classifiers RSwithRF and RSwithBN,
method decomposed this image up-to 3 levels, as we used for classification purposes.
wavelet decomposition, which reduces and trims the size of
an image using 3 level decomposition. The right hand side D. CLASSIFICATIONS
of Figure 6 shows this decomposed image. The approxi- The proposed technique used an individual classifier called:
mate coefficient (LL3) is our interesting part and its size is Feed Forward - ANN (FF-ANN), and two hybrid clas-
32×32 × 3.3072, which is still very large as an input for sifiers namely: Random Subspace with Random Forest
classifiers, in terms of computation. Therefore, it should be (RSwithRF) and Random Subspace with Bayesian Net-
optimised to reduce computation overhead. For this reason, work (RSwithBN) for classification purposes, in the last

VOLUME 9, 2021 33317

 M. Assam et al.: Efficient Classification of MRI Brain Images

FIGURE 7. The ANN used in the proposed model.

2) CLASSIFICATION USING HYBRID CLASSIFIERS

The proposed work also utilized hybrid techniques to classify
MRI brain images into normal and abnormal images. While
studying the Literature, it was found that the performance
FIGURE 6. blue image, on the left, is an original colored (RGB) image, and
of hybrid classifiers is better than individual classifiers. The
the right side black color image shows the decomposition of an main reason is that these ensembled techniques balance the
color (RGB) image up to level 3, where LL3 is an approximate coefficient performance of individual classifiers before the ultimate deci-
at level 3.
sion is taken. They support individual classifiers against fail-
stage of the model. These classifiers are discussed in this ure and, thus, enable them to perform better [37].
section. In the proposed technique, we hired Random Subspace
(RS) classifier that is used in combination with base learner
1) CLASSIFICATION USING FEED FORWARD-ANN classifiers such as Random Forest (RF) and Bayesian Net-
This work used the FF-ANN classifier for the classification work (BN) in ensemble fashion. These are selected because
of MRI brain images into normal and abnormal classes. they give promising results when combined with weak clas-
The main reason behind using a neural network is, its most sifiers. The results showed that it is a good approach for
extensive and prevalent use, for the classification of patterns, classification of healthy images which are free from diseases
because an ANN not only requires information about prob- and abnormal images that contain injuries and diseases. The
ability distribution but also priori probabilities. The struc- detailed explanations of these classifiers are given below:
ture and working of Artificial Neural Networks (ANNs) are 1) Random Subspace (RS): The best thing about RS is
like human brain as they perform storage, interpretation and that it attempts to increase and preserve its accuracy,
cognitive activities in a similar fashion to human brain. The whenever, it becomes more complex. The RS consists
training of a neural network for a particular activity takes a of several classifiers. It modifies the features space
huge amount of computation time, however, once it is trained, without modifying example space in training set.
then, it identifies the unknown objects in excellent manner. 2) Random Forest (RF): can be used not only for regres-
The mathematical model of an ANN comprises of artificial sion but it can also be used for classification. It contains
and non linear neurons, which run in parallel. An ANN could a number of decision trees just like classifiers. Each
have a single layered or multilayered architecture. Generally, decision tree gives vote to a class and votes of every
a three layered architecture use an input, a hidden and an out- tree in the forest are counted. The tree or class with
put layer. The hidden layer behaves like an interface between maximum votes in the forest is selected [38].
the input and output layers. The intermediate layer imple- 3) Bayesian Network (BN): A BN is a network struc-
ments the main function. There are many types of ANNs, ture that is shortly called Bays Net or Bays model.
in which, the FF-ANN is the most common and simplest It belongs to the probabilistic graphical models, and
among all. In the proposed model, an FF-ANN with two it is represented as a DAG (directed acyclic graph)
layered architecture is used. The two layers are hidden layer defined mathematically in terms of a set of variables
and output layer. The hidden and output layers, used in the (U = {X1 , . . . , Xn } where n ≥ 1) and probability tables
proposed methodology, have ten (10) and one (1) neurons, (Bp = {p(u | pr(u)), where u ∈ U}). The pr(u) represents
as illustrated in Figure 7. Sigmoid function is used in the u’s parent set in the BN. The conditional Q probability
hidden layer while linear function is used in the output layer. distribution of BN is calculated as P(U) = u∈U p(u |
The optimal set of nine features, obtained in last stage, are pr(u)).
sent to hidden layer for execution while the output layer
classifies an image as a normal or abnormal image. A normal IV. THE PROPOSED MECHANISM
image is represented by a 1 while an abnormal by a 0. The The algorithm used in the proposed model (presented in
FF-ANN was trained using BP- Algorithm [36] of Leven algorithm 1) takes MRI brain images as input, which are
berg-Marquardt, which automatically adjust the weights until passed through the four step process, for their classification
it reaches its objective. as normal or abnormal images. This algorithm uses certain

33318 VOLUME 9, 2021

M. Assam et al.: Efficient Classification of MRI Brain Images

terminologies such as assignment operator and functions such TABLE 1. Details of MRI images.
as Median_Filter (K) and RGB(), where the former apply
median filter on images while the latter convert an MRI brain
image from gray scale to a colored (RGB) image. The proce-
dures Decomposed_3L () and Approx_3L() use haar wavelet
for the decomposition and approximation of images to three
(3) levels, respectively. Channel (1), Channel (2), Channel (3)
subroutines are used to extract the Red, Green, Blue channels
of the converted color (RGB) images. The procedures named:
Mean_Image (), Stand_Dev_Image (), and Skewness_Image
() are used to compute the Mean, Standard Deviation and
Skewness of the colored images.

Algorithm 1 The Proposed Classification Algorithm

Require: MRI Brain images, Number of Images;
//Initialisations
1: int n = Number of Iages;
2: float FeaturesDate [n,9] = 0;
3: MRIImage = Nil;
FIGURE 8. MRI sample images A. Normal MRI B. MRI with acute stroke C.
4: for(int i = 1; i <= n; i++) MRI with Alzheimer disease D. MRI with Tumor.
5: MImage = Get the ith image;
//The following are Phase 1 (pre-processing stage)
steps RAM. The system was running 64-bit window 8 operating
6: Median_Filter (MImage ); system. The tools used for experiments were Matlab R2010a
7: L = RGB (MImage ); and Weka having version 7.10.0 and 3.6, respectively.
//The following are Phase 2 (feature extraction A standard dataset containing 70 T2 weighted images was
stage) steps used to evaluate the proposed methodology. The images in
8: N = Decomposed_3L(L); this database had 256 × 256 resolution and this was adopted
9: C = Approx_3L (N); from [34] like other researchers. Among the total 70 images
//The following are Phase 3 (feature reduction considered in this work, 45 images were abnormal and they
stage) steps were effected by three different kind of diseases namely:
10: Red = Channel (1); brain tumor, acute stroke, and alzheimer. From every disease,
11: Green = Channel (2); only 15 images were considered for experimental purpose.
12: Blue = Channel (3); The remaining 25 images were normal and they were not
//Get the nine features for each image and store affected by any kind of injuries. Figure 8 illustrates normal
them in array and abnormal images. This work used a percentage split of
13: FeaturesDate[i,1] = Mean(Red); 65% and 35% for training and testing purpose, when FF-ANN
14: FeaturesDate[i,2] = Standard_Deviation(Red); classifier was used. However, to test hybrid classifiers, it used
15: FeaturesDate[i,3] = Skewness(Red); 10 Fold cross validation technique.
16: FeaturesDate[i,4] = Mean(Green);
17: FeaturesDate[i,5] = Standard_Deviation(Green); A. ALGORITHM ACCURACY
18: FeaturesDate[i,6] = Skewness(Green); This work examined the proposed technique using differ-
19: FeaturesDate[i,7] = Mean(Blue); ent statistical techniques and results are compared with the
20: FeaturesDate[i,8] = Standard_Deviation(Blue); existing work. From the Literature, it was learnt that most of
21: FeaturesDate[i,9] = Skewness(Blue); the researchers used accuracy to measure the performance.
22: end for //The following are Phase 4 (classification Table 2 and 3 present the performance and accuracy achieved
stage) steps by the proposed algorithm for hybrid classifiers and FF-ANN,
23: Apply various classifiers with different arrangements respectively. The accuracy recorded for the RSwithRF and
on array FeaturesDate[,]; RSwithBN classifiers was 97.14% and 95.71%, respectively.
24: Record the accuracy of each classifier against the The classification accuracy of the proposed model for the
given settings; FF-ANN was 100% during the training while it was 91.66%
during the testing stage. Overall, 95.83% accuracy, on aver-
age, was observed based on both training and testing.
V. EVALUATION AND RESULTS The comparative analysis of hybrid and individual classi-
To implement and evaluate the proposed algorithm, this work fiers revealed that hybrid classifiers are more beneficial and
used a Core i5 system having 2.4GHz processor and 3GB sophisticated than individual classifiers. Hybrid classifiers

VOLUME 9, 2021 33319

 M. Assam et al.: Efficient Classification of MRI Brain Images

TABLE 2. Evaluation results of hybrid classifiers (in terms of accuracy).

TABLE 3. Evaluation results of individual classifiers (in terms of accuracy).

TABLE 4. Comparative analysis based on classification accuracy.

TABLE 5. Features based comparative analysis. VI. CONCLUSION

This paper suggested a new mechanism to differentiate MRI
brain images into normal and abnormal using individual and
hybrid classifiers. The suggested method used median fil-
ter in the pre-processing stage. Discrete Wavelet Transform
was used to extract the features and Color Moments were
introduced to reduce the features to an optimal set of nine
features and, thus, minimise the complexity and memory
usage. These nine features were extracted from all 70 images
considered in this work, in which 25 were normal and the
remaining 45 were abnormal images, which were affected
perform beter than individual classifiers. Detailed compar- by three kind of diseases. These feature sets for all images
ative analysis of proposed method with existing models in were passed to supervised classifiers, in which Feed Forward
terms of accuracy and the number of features being used - ANN, and hybrid classifiers: Random Subspace with Ran-
for classification purposes are presented in Table 4 and 5, dom Forest and Random Subspace with Bayesian Network
respectively. had a promising response in terms of classification accuracy,
in differentiating normal and abnormal images. The accuracy
B. LIMITATIONS OF CURRENT STUDY for the above mentioned individual and hybrid classifiers
Though, the proposed mechanism suggested improvements was 95.83%, 97.14% and 95.71% respectively. Experimental
over the existing methods, however the current study has results further proved that the proposed mechanism is far
a number of limitations. It tested the proposed mech- better than different existing techniques in terms of accuracy
anism against only a single individual and two hybrid and features being used.
classifiers. Similarly, it is tested against 70 images of In future, we intend to extend the proposed model to inves-
only one dataset. Furthermore, this work considered only tigate more individual and hybrid classifiers. This research
three (3) statistical features. Moreover, this work is not would also be extended to evaluate the proposed model with
compared with the most recent studies based on deep more parameters. This work could be also extended to use
learning. different features reduction methodologies while keeping the

33320 VOLUME 9, 2021

M. Assam et al.: Efficient Classification of MRI Brain Images

execution time at minimum. It would be interesting to com- [16] A. Korai and M. A. Unar, ‘‘Classification of MRI Images Using Neural
pare the results of the proposed technique with the methods Network,’’ Quaid-E-Awam Univ. Res. J. Eng., Sci. Technol., Nawabshah,
vol. 18, no. 1, pp. 66–71, 2020.
based on deep learning. It would be further interesting to [17] A. Ortiz, J. M. Górriz, J. Ramírez, D. Salas-González, and
check the impact of statistical features other than those used J. M. Llamas-Elvira, ‘‘Two fully-unsupervised methods for MR brain
in this work. image segmentation using SOM-based strategies,’’ Appl. Soft Comput.,
vol. 13, no. 5, pp. 2668–2682, May 2013.
[18] Y. Zhang and L. Wu, ‘‘An MR brain images classifier via principal com-
ACKNOWLEDGMENT ponent analysis and kernel support vector machine,’’ Prog. Electromagn.
Res., vol. 130, pp. 369–388, 2012.
This work was supported in part by the Basic Science [19] E.-S.-A. El-Dahshan, T. Hosny, and A.-B.-M. Salem, ‘‘Hybrid intelligent
Research Program through the National Research Founda- techniques for MRI brain images classification,’’ Digit. Signal Process.,
tion of Korea (NRF) funded by the Ministry of Education vol. 20, no. 2, pp. 433–441, Mar. 2010.
[20] A. Hyvärinen, J. Karhunen, and E. Oja, Independent Component Analysis,
under Grant NRF-2019R1A2C1006159, and in part by the vol. 62, no. 3. Hoboken, NJ, USA: Wiley, 2010, pp. 412–416.
Ministry of Science and ICT (MSIT), South Korea, under [21] M. Li and B. Yuan, ‘‘2D-LDA: A statistical linear discriminant analysis
the Information Technology Research Center (ITRC) support for image matrix,’’ Pattern Recognit. Lett., vol. 26, no. 5, pp. 527–532,
Apr. 2005.
program, supervised by the Institute for Information and [22] W. H. Ibrahim, A. A. A. Osman, and Y. I. Mohamed, ‘‘MRI brain image
Communications Technology Promotion (IITP), under Grant classification using neural networks,’’ in Proc. Int. Conf. Comput., Electr.
IITP-2020-2016-0-00313. Electron. Eng. (ICCEEE), Aug. 2013, pp. 253–258.
[23] N. H. Rajini and R. Bhavani, ‘‘Classification of MRI brain images using k-
nearest neighbor and artificial neural network,’’ in Proc. Int. Conf. Recent
REFERENCES Trends Inf. Technol. (ICRTIT), Jun. 2011, pp. 563–568.
[24] N. Gupta, P. Bhatele, and P. Khanna, ‘‘Glioma detection on brain
[1] Y. Zhang, Z. Dong, L. Wu, and S. Wang, ‘‘A hybrid method for MRI brain
MRIs using texture and morphological features with ensemble learning,’’
image classification,’’ Expert Syst. Appl., vol. 38, no. 8, pp. 10049–10053,
Biomed. Signal Process. Control, vol. 47, pp. 115–125, Jan. 2019.
Aug. 2011.
[25] P. R. E. Arasi and M. Suganthi, ‘‘A clinical support system for brain tumor
[2] Y. Zhang, Z. Dong, A. Liu, S. Wang, G. Ji, Z. Zhang, and J. Yang, classification using soft computing techniques,’’ J. Med. Syst., vol. 43,
‘‘Magnetic resonance brain image classification via stationary wavelet no. 5, p. 144, May 2019.
transform and generalized eigenvalue proximal support vector machine,’’ [26] Z. Ullah, S. H. Lee, and M. Fayaz, ‘‘Enhanced feature extraction tech-
J. Med. Imag. Health Informat., vol. 5, no. 7, pp. 1395–1403, Dec. 2015. nique for brain MRI classification based on Haar wavelet and statistical
[3] M. Maitra and A. Chatterjee, ‘‘A slantlet transform based intelligent sys- moments,’’ Int. J. Adv. Appl. Sci., vol. 6, no. 7, pp. 89–98, Jul. 2019.
tem for magnetic resonance brain image classification,’’ Biomed. Signal [27] J. Jeong, L. Wang, B. Ji, Y. Lei, A. Ali, T. Liu, W. J. Curran, H. Mao,
Process. Control, vol. 1, no. 4, pp. 299–306, Oct. 2006. and X. Yang, ‘‘Machine-learning based classification of glioblastoma
[4] L. M. Fletcher-Heath, L. O. Hall, D. B. Goldgof, and F. R. Murtagh, ‘‘Auto- using delta-radiomic features derived from dynamic susceptibility contrast
matic segmentation of non-enhancing brain tumors in magnetic resonance enhanced magnetic resonance images: Introduction,’’ Quant. Imag. Med.
images,’’ Artif. Intell. Med., vol. 21, nos. 1–3, pp. 43–63, Jan. 2001. Surg., vol. 9, no. 7, p. 1201, 2019.
[5] E.-S.-A. El-Dahshan, H. M. Mohsen, K. Revett, and A.-B.-M. Salem, [28] M. F. B. Othman, N. B. Abdullah, and N. F. B. Kamal, ‘‘MRI brain classi-
‘‘Computer-aided diagnosis of human brain tumor through MRI: A survey fication using support vector machine,’’ in Proc. 4th Int. Conf. Modeling,
and a new algorithm,’’ Expert Syst. Appl., vol. 41, no. 11, pp. 5526–5545, Simulation Appl. Optim., Apr. 2011, pp. 1–4.
Sep. 2014. [29] M. Saritha, K. Paul Joseph, and A. T. Mathew, ‘‘Classification of MRI
[6] R. C. Gonzalez and R. E. Woods, Digital Image Processing. Upper Saddle brain images using combined wavelet entropy based spider Web plots
River, NJ, USA: Prentice-Hall, 2002. and probabilistic neural network,’’ Pattern Recognit. Lett., vol. 34, no. 16,
[7] B. Shinde, D. Mhaske, and A. Dani, ‘‘Study of noise detection and noise pp. 2151–2156, Dec. 2013.
removal techniques in medical images,’’ Int. J. Image, Graph. Signal [30] S. Lahmiri and M. Boukadoum, ‘‘Brain MRI classification using an ensem-
Process., vol. 4, no. 2, p. 51, 2012. ble system and LH and HL wavelet sub-bands features,’’ in Proc. IEEE 3rd
[8] H. D. Cheng, J. Shan, W. Ju, Y. Guo, and L. Zhang, ‘‘Automated breast Int. Workshop Comput. Intell. Med. Imag., Apr. 2011, pp. 1–7.
cancer detection and classification using ultrasound images: A survey,’’ [31] H. B. Nandpuru, S. S. Salankar, and V. R. Bora, ‘‘MRI brain cancer
Pattern Recognit., vol. 43, no. 1, pp. 299–317, Jan. 2010. classification using support vector machine,’’ in Proc. IEEE Students’
[9] M. A. Hamid and N. A. Khan, ‘‘Investigation and classification of MRI Conf. Electr., Electron. Comput. Sci., Mar. 2014, pp. 1–6.
brain tumors using feature extraction technique,’’ J. Med. Biol. Eng., [32] H. Kalbkhani, A. Salimi, and M. G. Shayesteh, ‘‘Classification of brain
vol. 40, pp. 307–317, Mar. 2020. MRI using multi-cluster feature selection and KNN classifier,’’ in Proc.
23rd Iranian Conf. Electr. Eng., May 2015, pp. 93–98.
[10] C. Arizmendi, A. Vellido, and E. Romero, ‘‘Classification of human brain
[33] F. Wahid, R. Ghazali, M. Fayaz, and A. S. Shah, ‘‘Using probabilistic
tumours from MRS data using discrete wavelet transform and Bayesian
classification technique and statistical features for brain magnetic reso-
neural networks,’’ Expert Syst. Appl., vol. 39, no. 5, pp. 5223–5232,
nance imaging (MRI) classification: An application of AI technique in bio-
Apr. 2012.
science,’’ Int. J. Bio-Sci. Bio-Technol., vol. 8, no. 6, pp. 93–106, Dec. 2017.
[11] S. Chaplot, L. M. Patnaik, and N. R. Jagannathan, ‘‘Classification of [34] A. K. Johnson and J. A. Becker. (2020). The Whole Brain Atlas
magnetic resonance brain images using wavelets as input to support vector (Harvard Medical School). Accessed: Nov. 2020. [Online]. Available:
machine and neural network,’’ Biomed. Signal Process. Control, vol. 1, http://www.med.harvard.edu/AANLIB/home.html
no. 1, pp. 86–92, Jan. 2006.
[35] M. Maitra and A. Chatterjee, ‘‘Hybrid multiresolution slantlet transform
[12] P. Hiremath, S. Shivashankar, and J. Pujari, ‘‘Wavelet based features for and fuzzy c-means clustering approach for normal-pathological brain
color texture classification with application to CBIR,’’ Int. J. Comput. Sci. MR image segregation,’’ Med. Eng. Phys., vol. 30, no. 5, pp. 615–623,
Netw. Secur., vol. 6, no. 9A, pp. 124–133, 2006. Jun. 2008.
[13] M. H. Fazel Zarandi, M. Zarinbal, and M. Izadi, ‘‘Systematic image [36] A. Toprak and İ. Güler, ‘‘Suppression of impulse noise in medical images
processing for diagnosing brain tumors: A type-II fuzzy expert system with the use of fuzzy adaptive median filter,’’ J. Med. Syst., vol. 30, no. 6,
approach,’’ Appl. Soft Comput., vol. 11, no. 1, pp. 285–294, Jan. 2011. pp. 465–471, Nov. 2006.
[14] M. M. Badža and M. Č. Barjaktarović, ‘‘Classification of brain tumors from [37] T. G. Dietterich, ‘‘Ensemble methods in machine learning,’’ in Proc.
MRI images using a convolutional neural network,’’ Appl. Sci., vol. 10, Int. Workshop Multiple Classifier Syst. Berlin, Germany: Springer, 2000,
no. 6, p. 1999, Mar. 2020. pp. 1–15.
[15] A.-I. Barranco-Gutiérrez, ‘‘Machine learning for brain images classifi- [38] I. Barandiaran, ‘‘The random subspace method for constructing decision
cation of two language speakers,’’ Comput. Intell. Neurosci., vol. 2020, forests,’’ IEEE Trans. Pattern Anal. Mach. Intell., vol. 20, no. 8, pp. 1–22,
pp. 1–7, Jun. 2020. Aug. 1998.

VOLUME 9, 2021 33321

 M. Assam et al.: Efficient Classification of MRI Brain Images

[39] H. Wang and B. Fei, ‘‘A modified fuzzy C-means classification method SAID KHALID SHAH received the Ph.D. degree
using a multiscale diffusion filtering scheme,’’ Med. Image Anal., vol. 13, from the University of East Anglia, U.K. In 2004,
no. 2, pp. 193–202, Apr. 2009. he joined the University of Science and Technol-
ogy, Bannu, Pakistan, as a Lecturer, and promoted
as an Assistant Professor, in 2012. His research
MUHAMMAD ASSAM received the B.Sc. degree interest includes medical image processing and
in computer software engineering from the Uni- analysis, such as segmentation, visualization, and
versity of Engineering and Technology Peshawar, registration.
Pakistan, in 2011, and the M.Sc. degree in software
engineering from the University of Engineering
and Technology, Taxila, Pakistan, in 2018. He is
currently pursuing the Ph.D. degree in computer
science and technology with Zhejiang University,
China. Since November 2011, he has been working
as a Lecturer (on study leave) with the Department ARIF MEHMOOD received the Ph.D. degree
of Software Engineering, University of Science and Technology, Bannu, from the Department of Information and Com-
Pakistan. His research interests include brain machine interface, medical munication Engineering, Yeungnam University,
image processing, machine/deep learning, the Internet of Things(IoT), and South Korea, in November 2017. Since Novem-
computer vision. ber 2017, he has been working as an Assistant
Professor with the Department of Computer Sci-
ence and Information Technology, IUB, Pakistan.
HIRA KANWAL received the master’s degree in His recent research interest includes data min-
computer software engineering from the National ing, mainly working on AI and deep learning
University of Science and Technology, Pakistan. based text mining and data science management
She is currently working as an Associate Lecturer technologies.
with The Islamia University of Bahawalpur. Her
recent research interests include machine learn-
ing, recommender systems, data science, and data
mining.

GYU SANG CHOI received the Ph.D. degree from

the Department of Computer Science and Engi-
UMAR FAROOQ received the Ph.D. degree in neering, Pennsylvania State University, Univer-
computer science from the University of East sity Park, PA, USA, in 2005. He was a Research
Anglia, Norwich, U.K., in 2012. He is currently an Staff Member with the Samsung Advanced Insti-
Assistant Professor with the University of Science tute of Technology (SAIT), Samsung Electronics,
and Technology, Bannu, Pakistan. His research from 2006 to 2009. Since 2009, he has been a Fac-
interests include parallel and distributed simula- ulty Member with the Department of Information
tion systems, virtual worlds, games, the Internet of and Communication, Yeungnam University, South
Things, and grid and cloud computing. Korea. His research interests include non-volatile
memory and storage systems.

33322 VOLUME 9, 2021