Automated optimized classification

techniques for magnetic resonance brain

images

Ahmed S. Elkorany & Zeinab

F. Elsharkawy

Multimedia Tools and Applications

An International Journal

ISSN 1380-7501
Volume 79
Combined 37-38

Multimed Tools Appl (2020)

79:27791-27814
DOI 10.1007/s11042-020-09306-6

1 23
Your article is protected by copyright and
all rights are held exclusively by Springer
Science+Business Media, LLC, part of
Springer Nature. This e-offprint is for personal
use only and shall not be self-archived in
electronic repositories. If you wish to self-
archive your article, please use the accepted
manuscript version for posting on your own
website. You may further deposit the accepted
manuscript version in any repository,
provided it is only made publicly available 12
months after official publication or later and
provided acknowledgement is given to the
original source of publication and a link is
inserted to the published article on Springer's
website. The link must be accompanied by
the following text: "The final publication is
available at link.springer.com”.

1 23
 Author's personal copy
Multimedia Tools and Applications (2020) 79:27791–27814
https://doi.org/10.1007/s11042-020-09306-6

Automated optimized classification techniques

for magnetic resonance brain images

Ahmed S. Elkorany 1,2 & Zeinab F. Elsharkawy 3

Received: 24 May 2019 / Revised: 29 June 2020 / Accepted: 9 July 2020 /

Published online: 2 9 July 2020
# Springer Science+Business Media, LLC, part of Springer Nature 2020

Abstract
This paper presents automatic tumor detection and classification approaches for brain
magnetic resonance images (MRI). These approaches are based on hybrid-optimized
classification techniques and classify brain MRI to healthy, benign or malignant. The
proposed system implements three-optimization techniques combined with Artificial
Neural Network (ANN). Multi-Verse Optimizer (MVO), Moth-Flame Optimizer
(MFO) and Salp Swarm Algorithm (SSA) are used and compared to examine how these
techniques could be successfully employed to enhance the classification accuracy via
selecting the optimal parameters of ANN. The proposed techniques are applied to the
Harvard database and BRATS challenge dataset to evaluate the performance via Receiver
Operation Characteristics (ROC) analysis. The approaches are tested against geometric
transformations such as scaling, rotation and warping to show how much the proposed
system resists these transformations. Experimentally, the proposed algorithms achieve the
highest classification accuracy as compared to the other published ones. Also, the MVO-
ANN algorithm outperforms the other proposed algorithms.

Keywords Image processing . Pattern recognition . Artificial neural networks . Multi-verse

optimizer . Moth-flame optimizer . Salp swarm algorithm

* Zeinab F. Elsharkawy
[email protected]

1
Department of Electronics and Electrical Comm. Engineering, Faculty of Electronic Engineering,
Menoufia University, Menouf 32952, Egypt
2
High Institute of Electronic Engineering, Ministry of Higher Education and Scientific Research,
Belbeis, Elsharkia, Egypt
3
Engineering Department, Nuclear Research Center, Atomic Energy Authority, Cairo, Egypt
 Author's personal copy
27792 Multimedia Tools and Applications (2020) 79:27791–27814

1 Introduction

Brain diseases like brain tumors raised very fast. It is one of the major causes of cancer-related
death among adults and children in the world. In the brain tumor, the abnormal cells grow
inside or around the brain. Several imaging techniques are used for brain tumor detection such
as X-Ray, Portion Emission Tomography (PET), Computed Tomography (CT) and MRI [1].
MRI is widely used among all imaging techniques to visualize brain soft tissues. Compared
with other techniques, MRI results in high contrast and high-quality images with detailed
information that is needed for brain tumor detection, plus it is a nondestructive imaging
technique [2]. The early detection of brain tumors allows the prompt treatment that increases
the possibility of healing.
Recent work shows that the brain tumor detection can be done via supervised classification
techniques like k-nearest neighbors (k-NN) [3], Support Vector Machine (SVM) [4] and ANN
[3], Probabilistic Neural Network (PNN) [5], Convolution Neural Network (CNN)[6], Deep
Neural Network [7]. Also, the unsupervised techniques like Fuzzy Clustering Means (FCM)
[8] and self-organization map (SOM) [9] are used. While the obtained classification accuracy
(CA) of the unsupervised classifier is good, the CA and the performance of the supervised
classifier is higher. The classifier worked successfully by the proper selection of the its
parameters. This can be done by combining the classifier with the optimization techniques.
Recently, three new population-based stochastic optimization algorithms namely, the MVO
[10–12], MFO [13, 14], and SSA [15, 16] are introduced. MVO is used for global optimization
problem solutions. It is inspired by the multiverse theory in physics and big bang theory. The
main motivation of the MVO is a white hole, black hole, and wormhole that used to estimate
exploration, exploitation and local search [11]. The MVO has been used in many applications
[10–12]. In these studies, the MVO algorithm gives a better CA compared to other state-of-the-
art algorithms i.e., Gravitational Search Algorithm (GSA), Grey Wolf Optimizer (GWO) [17,
18], Particle Swarm Optimization (PSO) [19, 20], and Genetic Algorithm (GA) [21]. It has the
ability to get the global optimal solution by avoiding the local optima that makes it suitable for
real applications.
On the other hand, MFO is a metaheuristic algorithm based on the special behavior of the
moths at night [13, 14]. Moths in nature, are flying using the moonlight around the moon with
a fixed angle and hence, they are flying in straight lines. The power of the MFO algorithm is
appeared by increasing the exploration of the search space and decreasing the possibility of
being trapped in local optima. The results of the MFO algorithm proved high exploitation
capability, applicability to solve real problems when compared with other swarm algorithms
such as GA, GSA, BAT optimization algorithm (BAT) [22], and PSO. It also has the ability of
search spaces optimization with infeasible regions.
Moreover, SSA is a meta-heuristic algorithm used for solving the optimization problems
[15, 16]. It is inspired by the swarming behavior of salps in oceans when navigating and
foraging. SSA algorithm exhibits better result and it outperforms other optimization algorithms
like BAT, PSO, GA, and GSA. Its benefits are high exploitation and convergence speed.
In addition, the main advantages of these algorithms are simple frameworks, flexibility, and
local optima avoidance. There is no need for structural adjustments for solving different global
optimization problems because they have a small number of parameters to be adapted. They
are used for several real applications. Based on the success of the proposed MVO, MFO and
SSA algorithms. They are examined in this work with ANN.
 Author's personal copy
Multimedia Tools and Applications (2020) 79:27791–27814 27793

In this paper, hybrid optimized classification techniques for brain MRI are proposed to
classify brain images into healthy, benign or malignant. The presented technique implements
the MVO, MFO, or SSA combined with ANN to increase the CA and decrease the Error Rate
(ER) by selecting the ideal ANN parameters, i.e., to prevent ANN to be stuck in native optima.
The features of brain MRI datasets (Harvard and BRATS 2015) are extracted from whole
images after and before segmentation and their results are compared. Rescaled, rotated and
warped brain MRI are examined to evaluate the resistance of our approaches against these
geometric transformations.
In compared to well-known brain MRI optimized classifiers, i.e., GWO-ANN, PSO-
ANN,PSO-SVM, Whale Optimization Algorithm (WOA) WOA-ANN [14, 18], GA-SVM,
FKNN-BAT [22] and Sine Cosine optimization Algorithm (SCA)SCA-ANN [20], the pro-
posed algorithms provide very good outcomes in terms of improved native optima avoidance,
CA and ER.
The paper has been organized as follows: section 1 is the introduction. The second section
reviews the research work on brain MRI classification. The proposed hybrid method is
presented in the third section. The case study and results are discussed in the fourth section.
Finally, the conclusion is found in section 5.

2 Related works

Several brain MRI classification approaches are developed to differentiate the healthy from
tumorous ones [3, 4, 17, 19, 22–29], distinguish between benign and malignant ones [6, 18,
35–42], or classify the three types [5, 21, 43, 44, and].
A hybrid technique of Principle Component Analysis (PCA) and Discrete Wavelet Trans-
form (DWT) are used for features extraction and reduction [3]. The DWT is used to extract the
features and PCA is used to reduce the feature space. These features are fed to k-NN and
forward back-propagation ANN (FP-ANN) to classify the brain MRI to healthy or tumorous.
The resultant CAs are 98% and 97% for k-NN and FP-ANN, respectively. An automatic
method for tumor detection in brain MRI is presented in [4]. The SVM with different cross-
validation is applied on texture, shape, and intensity features and average CA of 97.1% was
achieved. In [23], the authors used fused textural and semantic features with hybrid kernel-
based SVM (HKSVM) for the segmentation and classification of brain MRI with CA of
94.12%. Intensity, textural and fractal descriptors are proposed in [24] for brain tumor
detection. The extracted features are first optimized and fed to SVM classifier for the detection
task. The obtained CA is 96.04%. Pseudo Zernike moment and kernel SVM (KSVM) are used
in [25] for brain MRI classification resulting in 99.75% CA. The computer-assisted system is
utilized in [26] for brain tumor recognition based on gray-level co-occurrence matrix (GLMC)
and completed local binary pattern (CLBP). The features selection approach is performed
using a Fisher score-based filter and SVM classifier is used for classification. The overall CA
is 99%. A fully automatic tumor detection system of brain MRI is presented in [27]. This
system performs orientation analysis on images for enhancement of the forging regions in the
brain. After that, the extracted GLCM features from these regions are used for classification by
the adaptive neuro-fuzzy Inference System (ANFIS). In [28], co-active adaptive neuro-fuzzy
inference system (CANFIS) is proposed for brain tumor detection. The k-means clustering is
used for segmentation. The achieved CA is 99.58%. Some optimization techniques are used in
classification methods to obtain the classifier’s optimal parameters hence, enhancing the CA.
 Author's personal copy
27794 Multimedia Tools and Applications (2020) 79:27791–27814

One of them proposed a hybrid approach of genetic algorithm (GA) and SVM [21] for
automatic brain MRI classification. The spatial gray-level dependence (SGLDM) method
and DWT were used for feature extraction that fed to the SVM classifier. The achieved
accuracy range is 94.44% to 96.29%. Modified PSO and discrete ripplet-II transform based
on Extreme Learning Machine (ELM) are used for brain MRI classification [29]. A hybrid
system of SVM with particle swarm optimization (PSO) is used to classify brain MRI [15]. In
this system, DWT, curvelet transform and shearlet transform are used to detect brain abnor-
malities. Then, extracted features are selected using PSO for the optimal solution. The system
achieved 97.38% CA percentage. Our group proposed a hybrid optimized algorithm based on
GWO combined with ANN classifier [17]. The GWO-ANN classification technique achieved
98.91% average CA and higher performance than the traditional NN classifier. In [22], the
authors presented a combined method of Fisher+ Parameter-Free BAT (PFreeBAT) to obtain
the best features set with a fuzzy K-nearest neighbor (FKNN) to classify MRI of the brain to
normal and abnormal images. This method resulting in 98% CA.
Computer-aided detection (CAD) of brain MRI is presented in [30] to detect the type of tumor
(benign/malignant). It is performed in three stages; image pre-processing, image segmentation
and (texture and statistical) features extraction. After that, the SVM classifier is utilized for
detection and CA of 92% is obtained. Fuzzy c-means clustering is used in [31] for brain tumor
detection and classification based on SVM. The hybridized approach of the cuckoo search
algorithm (CSA) and Markov random field technique is used in [32] to segment and classify
brain MRI to benign or malignant. Authors in [33] are also used CSA to extract the tumor from an
MRI of the brain based on histogram thresholding. Harmony Search Optimization (HSO)
technique is presented in [34] for brain tumor segmentation. After that, (statistical and GLCM)
extracted features classified with Naive Bayes (NB) and Radial Basis Function Network (RBFN)
classifiers. The obtained CAs are 83.33% and 98.75%, respectively. Automated tumor detection
and classification of brain MRI based on the integration of modified adaptive sine cosine
optimization algorithm (MASCA) with PSO presented in [20]. This integrated model based on
a local linear radial basis function neural network (LLRBFNN) achieved CA of 98.75%. All the
above researches are applied to Harvard dataset [3, 4, 17, 19–34]..
Non-Sub sampled Contourlet Transform (NSCT) is presented in [35] for brain image
enhancement. Texture features are extracted to train and test the ANFIS classifier to classify
Glioma brain image into low-grade tumor (LGT) that called benign or high-grade tumors
(HGT) that called malignant. This method achieved 96.15% accuracy, 94.25 sensitivity and
95.65% specificity for the BRATS dataset. In [18], Gabor, moment invariant and GLCM
features are extracted and fed to an optimized Whale Optimization Algorithm-ANN (WOA-
ANN) to enhance the CA of MRI brain images. The obtained CA is 98%. Brain MRI
segmentation and classification approach of the BRATS dataset is proposed in [36]. It is
performed in three stages; K-means clustering, ANN to choose the correct object and texture
features extraction. Then, ANN and SVM are used for the classification task. The CA of
94.07% is recorded for ANN. In [6], Maxpool methodology is used in convolutional NN
(CNN) architecture to enhance the accuracy of the brain tumor detection system. In this
system, DWT is used for image fusion and GLCM are extracted features that fed to CNN
classifier for Glioma image classifications. The accuracy of 92.7%, the sensitivity of 89.85%
and the specificity of 92.5% are obtained. An automated segmentation method based on CNN
is applied to the brain MRI BRATS dataset in [37]. The CA of 93.7%, the sensitivity of
90.65% and the specificity of 91.75% are obtained. Dong et al. [38] presented U-Net based
deep CNN for brain lesion detection and the Dice Similarity Coefficient (DSC)of 86% is
 Author's personal copy
Multimedia Tools and Applications (2020) 79:27791–27814 27795

obtained on BRATS 2015 dataset. Amin et al. proposed deep NN (DNN) based architecture
for brain tumor detection and segmentation [39]. The DSC of 95% is achieved. An optimiza-
tion model for the automatic tumor detection of brain MRI is presented in [40]. This model
integrates the small kernels two-path CNN (SK-TPCNN) and random forests (RF). The
learned features from SK-TPCNN are classified using RF classifier. The wide residual network
and pyramid pool net-work (WRN-PPNet) is introduced in [41] for glioma segmentation and
classification of BRATS dataset images. The features are extracted from WRN and fed to the
PPNet for the detection task. The average DSC and sensitivity of the presented method are
91% and 94% respectively. In [42], residual cyclic unpaired encoder-decoder network
(RescueNet) is used for automatic brain tumor detection. The measured DSC of 94% and
the sensitivity of 88% are achieved.
Abd-Ellah et al., [43] developed a two-stage CAD system to classify normal or abnormal
images in the first stage, then, classify the abnormal images into benign or malignant. In [5] an
efficient classification method was presented to classify the brain MRI to normal, benign or
malignant based on PNN with a radial basis function. Abdullah et al. introduced a brain MRI
classification approach based on soft computing techniques [44]. In this approach, global
threshold and watershed segmentation techniques with ANN classifiers are utilized to detect
the type of tumor.
According to the related work, it can be concluded that, the GLCM is one of the main texture
analysis methods utilized with statistical features to obtain statistical properties for further
classification [5, 6, 17, 18, 20, 23, 24, 26, 27, 34], and they are employed in the presented
research work. Also, the classifier performance is good on training images but is poor on new
testing images. Hence, in this study, the statistical and texture (GLCM) features have been used
and two different combinations of testing and training image sets are used. The presented work
aims to classify brain MRI to healthy or tumorous images and recognize the type of tumor
(benign or malignant) using the hybridized classification approaches, i.e., MVO-ANN, MFO-
ANN, and SSA-ANN to enhance the CA (the lowest ER) in an automated manner.

3 Proposed approach

The proposed approach consists of three steps; namely: preprocessing, features extraction and
optimized ANN classification (MVO-ANN, MFO-ANN or SSA-ANN). First, for tissue
enhancement in brain MRI, Contrast Limited Adaptive Histogram Equalization (CLAHE)
[45] is employed and thresholding is used for tumor segmentation. Second, statistical and
texture features are extracted and used as inputs to ANN. Finally, the ANN’s parameters are
optimized using MVO, MFO or SSA. The proposed approach flow diagram is shown in Fig. 1.

3.1 Preprocessing and enhancement phase

3.1.1 Image normalization and enhancement

In this phase, all images are converted from RGB to grayscale first. Then, the intensity of the
brain MRI is normalized to obtain the same range of grey levels for all images. The most
common and simplest normalization technique is used, which normalizes the intensity of
pixels between 0 and 1 using Eq. 1.
 Author's personal copy
27796 Multimedia Tools and Applications (2020) 79:27791–27814

Fig. 1. The flow diagram of the presented approach

I a ðx; yÞ−minðI a Þ
I N ðx; yÞ ¼ ð1Þ
maxðI a Þ−minðI a Þ

Where (x, y)is the pixel position, IN(x, y) is the normalized pixel intensity, Ia(x, y) is the actual
pixel intensity, max(Ia) and min(Ia) are the maximum and minimum intensities over all pixels.
Next, the tissue enhancement step is performed to illustrate (isolate) the tumor area from the
image remained. Different image processing techniques are used in this step like histogram
equalization (HE), adaptive HE (AHE), Two-Stage AHE (TSAHE) and CLAHE [45]. The
CLAHE is an image processing technique that is used for improving contrast image. While HE
works on the entire image, CLAHE operates on small regions in the image called tiles and the
enhancement process are applied over each tile and adjust its contrast according to their
neighbor pixels. CLAHE combines neighboring tiles using bilinear interpolation to eliminate
artificially induced boundaries. It is different from AHE because of its contrast limitation that
prevent noise amplification in different image regions especially, homogeneous ones. When an
image has closed contrast regions (i.e., both the background and foreground are dark (or
bright) at the same time), CLAHE is useful method that changing the intensity of each pixel
according to all neighboring pixels and transformation function is derived [52].
In the tumor edge detection process, the edge appears extremely dark, which is very
disturbing. This problem is overcame using CLAHE that is more suitable for this task because
of its adaptive contrast enhancement over all neighboring pixels.
 Author's personal copy
Multimedia Tools and Applications (2020) 79:27791–27814 27797

3.2 Threshold based segmentation

The image is partitioned into its constituent parts using Segmentation. Thresholding is used for
image segmentation so as to get a binarized image with gray level 1 (white color) representing
the tumor and gray level 0 (black color) representing the background. The main purpose of the
binarization step is to discriminate the contrast value of tumor from normal regions in the
images by thresholding. The parameter which controls the segmentation is known as intensity
threshold (T) which is calculated using the following equation [46].
∑M−1 N−1
i¼0 ∑ j¼0 ei; j *M i; j
T¼ ð2Þ
∑M−1 N−1
i¼0 ∑ j¼0 M i; j

Where M is the number of MRI images. Mi, j, and ei, j represent the gray level image’s pixels in
vertical and horizontal directions. N is the number of pixels in vertical or horizontal direction.
After computing this threshold value, every pixel intensity is compared with. If the pixel
intensity is less than the threshold, it is established to be black in the output and vice versa.
Figure 2 shows the pre-processing step applied on three MRI of the brain (healthy, benign and
malignant). Image segmentation is an important process in early brain tumors diagnosis and it
is preferred to be done before feature extraction process.

Fig. 2 The preprocessing steps of brain MRI

 Author's personal copy
27798 Multimedia Tools and Applications (2020) 79:27791–27814

3.3 Feature extraction phase

The feature extraction is the procedure to represent a raw image in order to facilitate decision
making such as pattern classification. Features will be extracted from the whole brain MRI before
and after segmentation. Features extraction includes decreasing the amount of data required to
describe a large set of data correctly. The statistical and texture (GLCM) features are the most
generally applied method because of its high accuracy, simplicity and less computational time. In
this work, only ten features are extracted from each brain MRI by applying statistical and texture
(GLCM) models, hence, no need for features reduction. These features are the inputs to the
optimized ANN classifier which assigns them to the class which they represent.
1) Statistical features: Statistical features that are extracted are the mean, variance, skew-
ness, standard deviation, and kurtosis.


1 m−1 n−1
Mean ¼ ∑ ∑ f ðx; yÞ ð3Þ
m*n x¼o y¼0


1 m−1 n−1
Variance ¼ ∑ ∑ ð f ðx; yÞ−μÞ2 ð4Þ
m*n x¼0 y¼0

m−1 n−1
3

 ∑ ∑ ð f ðx; yÞ−μÞ
1 x¼0 y¼0
Skewness ¼ ð5Þ
m*n SD3

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
s

1 m−1 n−1
Standard Deviation ¼ ∑ ∑ ð f ðx; yÞ−μÞ2 ð6Þ
m*n x¼0 y¼0

 m−1 n−1
1 ∑x¼0 ∑y¼0 ð f ðx; yÞ−μÞ4
kurtosis ¼ ð7Þ
m*n SD4

2) Texture features: The textures which describe the spatial relationship between pixels of
different gray levels are found using GLCM. This method follows two stages for the features
extraction from the medical images. In the first stage, the GLCM is calculated, and in the next
stage, the texture features based on the GLCM are calculated. Contrast, entropy, Homogeneity,
Energy, and Correlation are some of the texture features and they are shown below.

m−1 n−1
Contrast ¼ ∑ ∑ ðx−yÞ2 f ðx; yÞ ð8Þ
x¼0 y¼0

m−1 n−1
Entropy ¼ − ∑ ∑ f ðx; yÞ log2 f ðx; yÞ ð9Þ
x¼0 y¼0
 Author's personal copy
Multimedia Tools and Applications (2020) 79:27791–27814 27799

m−1 n−1 f ðx; yÞ

Homogeneity ¼ ∑ ∑ ð10Þ
x¼0 y¼0 1 þ ðx−yÞ2

m−1 n−1
Energy ¼ ∑ ∑ ð f ðx; yÞÞ2 ð11Þ
x¼0n−1
y¼0
∑m−1
x¼0 ∑y¼0 ðx; yÞ f ðx; yÞ−μx μy
Correlation ¼ ð12Þ
σx σy

3.4 Optimized classification phase

Three optimization techniques are combined with ANN to obtain the optimal parameters of
ANN and achieve the best CA. The optimization algorithms (MVO, MFO, and SSA) and the
optimized ANN are indicated as follows.

3.4.1 Multi-verse optimizer

The MVO is a new nature-inspired algorithm presented in [10–12]. The three main concepts of
MVO are white hole, wormhole, and blackhole. These concepts are formulated in a mathe-
matical model to estimate exploration, exploitation, and local search, respectively. The
algorithm procedures are given in Fig. 3 as a pseudocode. Traveling distance rate (TDR)
and Wormhole existence probability (WEP) are the main coefficients that are used in the MVO
and they have updated the position of the universe. The WEP is required to increase the
optimization process linearly to emphasize its exploitation. The TDR is defined as the distance

Fig. 3 Pseudocodes of the MVO algorithm

 Author's personal copy
27800 Multimedia Tools and Applications (2020) 79:27791–27814

rate that an object can travel to reach the best universe. It is decreased during the optimization
process for increasing the local search accuracy of the best universe. Each universe (search
agent) contains the inflation rate and the fitness value that is calculated by the inflation rates.
Eqs. 13 and 14, gives the formula of WEP and TDR coefficient.


max−min
WEP ¼ min þ l  ð13Þ
L

l1=p
TDR ¼ 1− ð14Þ
L1=p

Where l is the current run, L is the maximum run. P is the accuracy of exploitation in each
run. The best universe contains more WEP values and fewer TDR values it.

3.4.2 Moth-flame optimizer

Mirjalili is the first one that presents the MFO in 2015 [13, 14]. The MFO is a nature-inspired
optimization algorithm that simulates the special navigation methods of moths in nature at
night. Moths fly around the moon with a fixed angle, which is a very effective mechanism for
traveling long distance in straight line. The algorithm consist of three-parts that approximates
the global and optimal solution of the optimization problems. It defined as follows:
MFO ¼ ½I; P; T  ð15Þ

The random population of moths is generated by I function that corresponding fitness values,
the main function is P that moves the moths around the search space. It received the matrix of
M (set of moths in a matrix) and returns its updated one. The T function returns true or false if
the termination criterion is satisfied or not, respectively.
The position of each moth was updated with respect to flame as follows:

 
M i ¼ P M i; F j ð16Þ

Where Mi is the ith moth and Fj is the jth flame

 
P M i ; F j ¼ Di  ebt  cosð2πtÞ þ F j ð17Þ

Where b is a constant that defines the shape of P, t is a random value between[r, 1], r is the
convergence constant, e is exponentials and Di is the distance of ith moth and the jth flame
which calculated by:
 
Di ¼  F j −M i  ð18Þ

In the search space, the position updating of moths with respect to different locations may
degrades the exploitation of the best promising solutions. To resolve this problem, an adaptive
mechanism is proposed for flames number.
The number of flames is adaptively decreased over the course of an iteration. It calculated
as follows:
 Author's personal copy
Multimedia Tools and Applications (2020) 79:27791–27814 27801

Fig. 4 Pseudocodes of the MFO algorithm


F n −1
flame number ¼ round F n −h* ð19Þ
Tn

Where h is the iterations current number, Fn indicate the maximum number of flames and Tn is
the maximum number of iterations. Figure 4 indicates the pseudocode of the MFO algorithm.

3.4.3 Salp swarm algorithm

The SSA is a novel meta-heuristic algorithm [15, 16] used to solve the optimization problems.
The swarming behavior of salps when navigating and foraging in oceans is the main
inspiration of SSA. The pseudocode of the SSA is illustrated in Fig. 5. The population is
 Author's personal copy
27802 Multimedia Tools and Applications (2020) 79:27791–27814

Fig. 5 Pseudocodes of the SSA algorithm

firstly divided into a leaders’ group and followers’ group. The leader position at the front of the
chain, whereas the remainder of salps, are the followers. The location of the leader is updated
according to Eq. 20.


F i þ si ððU bi −Lbi Þs2 þ Lbi Þ s3 ≥0
Xh1i ¼ : ð20Þ
F i −si ððU bi −Lbi Þs2 þ Lbi Þ s3 ≥0

Where Xh1i shows the leader’s location and Fi is the food location in the ith dimension, Ubi and
Lbi are the upper and lower bands respectively and the random numbers are s1, s2 and s3. The
most important parameter is s1 which responsible for balancing exploration and exploitation. It
defined as

s1 ¼ 2e−ð R Þ
r 2
ð21Þ

Where r is the current iteration and R is the maximum number of iterations. Equation 22 is
used for updating the follower’s location.
1 2
Xhij ¼ e c þ v0 c ð22Þ
2
Where, Xhij indicate the position of ith follower Salp in jthdimension, the initial speed is
vfinal
v0 ¼ v0 and c is the time. At v0 = 0, the follower’s location is calculated as follows:
1 i

Xhij ¼ Xh j þ Xhi−1
j ð23Þ
2

3.4.4 The optimized ANN

The above optimization algorithms are combined with ANN to obtain the best parameters of
ANN and achieve the optimal CA. The weights and biases are the most important parameter
variables for training ANN. So, the optimal values of weights and biases should be obtained to
provide the best classification rate (lowest ER). The common metric to evaluate the training
 Author's personal copy
Multimedia Tools and Applications (2020) 79:27791–27814 27803

Table 1 The initialization parameters of MVO, MFO, and SSA

Parameter value

No. of Search agent 200

Maximum iteration 500
[Lower band upper band] [−10 10]

ANN system is the average of Mean Square Error (MSE) over all the training samples, as
follows:

1 s 1 k  n n 2
T
MSE ¼ ∑ ∑ A −D ð24Þ
T s i¼1 T s i¼1 i i

Where k is the number of outputs. Ts is the number of training samples, the error between the
actual output is Ani and the desired output is Dni of ith input unit and nth training sample. The
fitness function is to obtain the highest CA by minimizing the average MSE. Hence, MVO,
MFO, or SAA are used to optimize the ANN by obtaining its optimal biases and weights so
that the MSE is minimum. Table 1, shows the initial parameters of MVO, MFO, and SSA. At
the beginning of the optimization process, it is assumed that a random biases and weights are
generated in range between the lower and upper bands.
The architecture of the optimized ANN is illustrated in Fig. 6(a). It is clear that, the ANN
consists of three layers; input, output and one hidden layer. The input {x1, x2, ……, x10} are
ten features data assigned to each node of the ANN’s input layer and Z1, Z2, …, ZN denotes
the sigmoid activation function in hidden nodes. In this network, the biases and weights are
initialized and optimized using MVO, MFO or SSA techniques.
The procedures of the proposed optimized- ANN model is shown in Fig. 6(b) and outlines
as follows:
1-Initialize the optimizer (MVO, MFO or SSA) parameters.
2-Train ANN classifier and the fitness of each search agent (universe of MVO, moth of
MFO or Salp of SSA) is evaluated. The fitness here is CA.
3-If the fitness of each search agent reaches the best CA, the position vector is saved and
goes to step5, else go to the next step.
4-The search agent’s position is updated until it reaches the optimal CA.
5-Get the optimal parameters (the biases and weights) of ANN and the optimal CA.
6-Training dataset to obtain learning model and use this model to predict test data and get
the CA.
In total, the pseudocode shown in Fig. 7. describes the proposed MVO-ANN system to
obtain the ideal parameters of ANN which result in a superior CA and ER.

4 Experiments and discussions

4.1 Database

To evaluate the performance of the proposed approach, two datasets are used. 110
images are obtained from the Harvard dataset [47], in which 30 images are healthy,
 Author's personal copy
27804 Multimedia Tools and Applications (2020) 79:27791–27814

Fig. 6 (a) The architecture of optimized ANN, (b) The flowchart of the optimized ANN model
 Author's personal copy
Multimedia Tools and Applications (2020) 79:27791–27814 27805

Fig. 7 The pseudocode of the proposed MVO-ANN approach

20 images are benign tumors experiencing a meningioma, LGT and 60 images are
malignant tumors experiencing a Sarcoma, Bronchogenic, Glioblastoma Multiform and
Carcinoma. Harvard brain images are directly downloaded as two-dimension (2D)
images in *.jpg format. BRATS 2015 dataset [48] consists of 80 images: 60 are
HGT (malignant) and 20 are LGT (benign). Its images are three-dimension (3D)

Table 2 Partition of training and testing set of the Harvard and the BRATS database

Database Training-testing Total No. of image No. of images in the training No. of images in the testing

Healthy Benign Malignant Healthy Benign Malignant

Harvard (50–50) 110 15 10 30 15 10 30

5-CV 110 24 16 48 6 4 12
BRATS (50–50) 80 – 10 30 – 10 30
5-CV 80 – 16 48 – 4 12
 Author's personal copy
27806 Multimedia Tools and Applications (2020) 79:27791–27814

Table 3 The proposed approach results when features are extracted from original images

Database classes method Sensitivity% Specificity% DSC% CA% ER% AUC

Harvard H/T MVO 0.9333 100 0.9655 98.18 1.82 0.9915

MFO 1 0.9750 0.9677 98.18 1.82 0.9867
SSA 1 0.950 0.9375 96.36 1.82 0.9752
ANN 93.33 95.12 90.32 94.64 5.36 0.9589
B/M MVO 0.9667 100 0.9831 97.50 2.5 0.9967
MFO 0.9667 100 0.9831 97.50 2.5 0.9867
SSA 0.9667 100 0.9831 97.50 2.5 0.9833
ANN 70 98.33 80 91.25 8.75 0.9081
BRATS 2015 B/M MVO 86.7 80 89.66 85 15 0.8667
MFO 90 70 90 85 15 0.8267
SSA 90 70 90 85 15 0.8900
ANN 83.33 50 76.92 70 30 0.7083

images in *.mha format and converted to 2D images in *.jpg format using DIACOM
software [53]. However, the two processed images datasets are in *.jpg format.
Harvard images dataset have higher resolution, higher contrast and large size [47].
Hence, better segmented images are obtained.

4.2 Performance measurement

Classification rate (CA), Area Under the ROC Curve (AUC), Specificity, Sensitivity,
and DSC parameters are used for evaluating and visualizing the classifier perfor-
mance, which are given as follows.
TP þ TN
Classification rateðCAÞ ¼ ð25Þ
T P þ T N þ FP þ FN

∞
0
AUC ¼ ∫ T PR ðxÞ F PR ðxÞ dx ð26Þ
−∞

Table 4 The proposed approach results when features are extracted from segmented images

Database classes method Sensitivity% Specificity% DSC% CA% ER% AUC

Harvard H/T MVO 100 100 100 100 0 1

MFO 100 100 100 100 0 1
SSA 100 100 100 100 0 1
ANN 100 97.50 96.77 98.18 1.82 0.9765
B/M MVO 100 100 100 100 0 1
MFO 100 100 100 100 0 1
SSA 100 100 100 100 0 1
ANN 100 70 95.24 92.5 7.5 0.9257
BRATS 2015 B/M MVO 96.77 100 98.36 97.62 2.38 0.9736
MFO 96.67 100 98.31 97.5 2.5 0.9809
SSA 96.77 100 96.67 95.24 4.76 0.9631
ANN 83.87 90.91 89.66 85.71 14.29 0.9384
 Author's personal copy
Multimedia Tools and Applications (2020) 79:27791–27814 27807

Fig. 8 ROC curves of proposed approaches when features are extracted from original images

TN
Specificity ¼ ð27Þ
T N þ FP

TP
Sensitivity ¼ ð28Þ
T P þ FN

2T P
DSC ¼ ð29Þ
2T P þ F P þ F N
Where TP, TN, FP and FNare true positive, true negative, false positive and false negative
respectively. TPR and FPR are true and false positive rate respectively.

4.3 The classifier setup

The brain MRI dataset is partitioned by one of two ways. First way, a standard partition
(50–50) training-testing is used, 50% of images are used for training the ANN classifier
and the rest are used for testing. The other way, 5-fold-cross validation (5-CV) is used to
determine the robustness of the proposed approach. The partitioning process is applied to
the databases (Harvard and BRATS 2015) as illustrated in the Table. 2.

Fig. 9 ROC curves of proposed approaches when features are extracted from segmented images
 Author's personal copy
27808 Multimedia Tools and Applications (2020) 79:27791–27814

Fig. 10 Geometric transformation on brain MR.

In the beginning, the features are extracted from the whole (original) and segment-
ed images of Harvard and BRATS datasets and examined using the proposed ap-
proaches. These features are used to train and test the proposed classifiers (MVO-
ANN, MFO-ANN, and SSA-ANN) and their results are depicted in Tables 3 and 4.
The ROC curves are shown in Figs. 8 and 9. As shown in these results, the
performance of the classifiers with the extracted features from segmented images is
better than that of the original image as expected. As the main aim of the segmen-
tation step is to isolate tumor regions from the normal ones. Figure 2c shows the
example of segmented healthy, benign and malignant images. It is clear from the
figure that, the efficient discrimination between these images can be done easily.
Hence the extracted features from segmented images are the best way for exact
diagnosis of brain tumor.
After that, the performance of the presented approaches is evaluated against the
geometric transformations. Different types of geometric transformations have been
considered to transform test images including scaling by 70% of original image size,
rotation by angle of 30o and warping. The extracted features from the transformed
images are used to test the proposed classifiers. Figure 10 shows the geometric
transformation performed on brain MRI. The resultant CA and ER for the proposed
three classifiers are listed in Table 5. The results indicate that the Harvard dataset
does not affected by these transformations when classifying the dataset to healthy and
tumorous(H/T) images. The CAs are decreased (higher ERs) but it is still acceptable
when recognizing the type of tumor (benign or malignant (B/M)). This is because of
its segmented images are better as explained above, and its training sets are

Table 5 The CA and ER values of the proposed approaches performed on transformed images

Database classes method Scale Rotation Warping

CA% ER% CA% ER% CA% ER%

Harvard H/T MVO-ANN 100 0 100 0 100 0

MFO-ANN 100 0 100 0 100 0
SSA-ANN 100 0 100 0 100 0
B/M MVO-ANN 92.5 7.5 95 5 92.5 7.5
MFO-ANN 92.5 7.5 95 5 92.5 7.5
SSA-ANN 95 5 90 10 87.5 12.5
BRATS 2015 B/M MVO-ANN 70 30 70 30 82.5 17.5
MFO-ANN 70 30 70 30 82.5 17.5
SSA-ANN 70 30 70 30 82.5 17.5
 Author's personal copy
Multimedia Tools and Applications (2020) 79:27791–27814 27809

Table 6 the proposed approach results of Harvard and BRATS datasets

Database classifier ER% CA%

(50–50) 5-CV (50–50) 5-CV

Harvard MVO-ANN 0 0.68 100 99.23

MFO-ANN 0 1.54 100 98.46
SSA-ANN 0 2.31 100 97.69
BRATS MVO-ANN 2.5 1.25 97.5 98.75
MFO-ANN 2.5 2.5 97.5 97.5
SSA-ANN 5 3.75 95 96.25

containing different tumor sizes and with different angles. So, Harvard images dataset
are slightly affected by the transformation. Concerning BRATS dataset, the classifi-
cation does not achieve the required high CA (low ER) because of its lower
resolution, lower contrast and smaller size compared to Harvard images dataset. So,
Its more affected by the geometric transformation.
The performance of the ANN with MVO, MFO, and SSA of each partition way
(standard and 5-CV) for the two databases are presented in Table 6. The ROC curves of
the proposed approaches of standard and 5-CV partitioning ways are illustrated in Figs. 9
and 11, respectively. The computational results listed in Table 6 denote that the optimal
CA and ER are obtained when using ANN with each optimizer i.e. MVO, MFO or SSA.
The highest CA of 100% and the lowest ER of 0% are achieved for the Harvard database
for each optimizer. The optimal CA of 98.75% is also achieved for the BRATS database
using MVO. All the listed results indicate that the proposed hybrid approaches outper-
form the traditional ANN approach.
This means that a successful combination between ANN and each optimizer (MVO,
MFO, or SSA) has occurred and it is considered an effective hybrid approach. The
proposed optimizers are effectively successful in selecting the optimal weights and
biases of the ANN. The performance of the presented approaches is compared with the
recently published results in the state of art methods for the same MRI datasets, which
are shown in Tables 7 and 8 for the Harvard database and in Table 9 for the BRATS
dataset. In addition, the proposed optimizers are compared to the most popular
optimizers, i.e., GA, BAT, SCA, WOA, GWO and PSO as illustrated in Tables 7, 8
and 9. It is clear that the performance of the proposed optimizers outperforms these

Fig. 11 ROC curves of proposed approaches (5-CV)

 Author's personal copy
27810 Multimedia Tools and Applications (2020) 79:27791–27814

Table 7 Comparison of classification rate on healthy and tumorous of the Harvard dataset

Paper Methods CA % ER%

[3], 2010 DWT + PCA + ANN 97 3

DWT + PCA + K-NN 98 2
[21], 2010 SGLDM+GA + SVM 94.44 5.56
WT + SGLDM+GA + SVM 96.29 3.71
[19], 2013 PSO + KSVM 97.78 2.22
[4], 2017 shape, texture, and intensity +SVM 97.1 2.9
[23],2017 textural, semantic +HKSVM 95.45 4.55
[24], 2017 Intensity, textural, fractal +PCA + SVM 96.04 3.96
[27],2017 GLCM +ANFIS 99.58 0.42
[28], 2018 K-mean + CANFIS 99.8 0.2
[25],2018 Pseudo Zernike moment+SVM 99.75 0.25
[29] 2018 MPSO+ELM 99.62 0.38
[17], 2018 GWO + ANN 98.91 1.09
[26], 2019 CLBP+GLCM+ Fisher score + SVM 99 1
[22], 2019 PFreeBAT+ FKNN 98 2
[49], 2019 PSO + SVM 97.38 2.62
Proposed (SSA, MFO and MVO) (50–50) 100 0
5-CV 100 0

popular ones. The methods that classify H/T images of the Harvard dataset are
compared in Table 7. In which our approaches achieved the optimal CA of 100%
and optimal ER of 0%. Then the methods that classify tumor images into B/M for
Harvard and BRATS datasets are compared in Table 8 and 9respectively. It is obvious
that the proposed classification approaches outperform the other published methods in
CA, ER, sensitivity, specificity, DSC, and AUC. The proposed MVO-ANN algorithm

Table 8 Comparison of performance analysis value on benign and malignant of the Harvard dataset

Paper Method Sensitivity% Specificity% DSC% CA % ER% AUC

[21], 2010 SGLDM+GA + SVM 91.87 100 – 94.44 5.56 –

WT + SGLDM+GA + SVM 94.6 100 – 96.29 3.71 –
[31]2014 Fuzzy c-means+SVM 94.68 5.32
[50], 2015 Wavelet-Entropy+SVM 78.05 100 – 82.69 17.31
[51], 2015 Wavelet-Entropy+GA + SVM 94.5 91.7 – 92.60 7.4
[32]2015 CSA+ Markov Random Field 98 2
[30], 2018 SVM 92 8
[33], 2019 histogram thresholding+ CSA 99.06 72.92 – 98.03 1.97
[20],2019 PSO + LLRBFNN 95.5 4.5
SCA + LLRBFNN 97.76 2.24
ASCA+PSO + LLRBFNN 98.75 1.25
[34],2020 HSO + NB 83.33 16.67
HSO + RBFN 98.75 1.25
ANN (50–50) 70 98.33 80 91.25 8.75 0.9257
5-CV 70 96.67 77.78 90 10 0.9376
Proposed ANN + MVO (50–50) 100 100 100 100 0 1
5-CV 100 98.33 97.56 98.75 1.25 0.9927
ANN + MFO (50–50) 100 100 100 100 0 1
5-CV 100 96.67 95.24 97.50 2.5 0.9886
ANN + SSA (50–50) 100 100 100 100 0 1
5-CV 100 95 93.02 96.25 3.75 0.9806
 Author's personal copy
Multimedia Tools and Applications (2020) 79:27791–27814 27811

Table 9 Comparison of performance analysis value on benign and malignant of the BRATS dataset

Paper Method Sensitivity% Specificity% DSC CA % ER% AUC

[37],2016 Deep CNN with 3 × 3 kernel 90.65 91.75 78 93.7 6.3

[6],2017 DWT + CLCM+CNN 89.85 92.5 – 92.7 7.3 –
[38], 2017 U-Net CNN – – 86 – – –
[35], 2018 NSCT +ANFIS 94.25 95.65 – 96.15 3.85 –
[39], 2018 DNN with 7 layers 95 95.2 95 95.1 4.9 –
[18], 2018 WOA-ANN 96 98.94 54.3 98 2 0.94722
GWO-ANN 88 94 56.17 92 8 0.8833
[40], 2019 SK-TPCNN 95 86 -
SK-TPCNN+RF 96 89 -
[36], 2019 K-Mean- ANN 90.09 96.78 94.07 5.93
K-Mean- SVM 87.54 91.32 90.72 9.28
[41], 2019 WRN-PPNet 92 94 –
[42], 2020 RescueNet 88 94 –
ANN (50–50) 83.87 90.91 89.66 85.71 14.29 0.9384
5-CV 87.10 90.91 91.53 88.10 11.9 0.9275
Proposed ANN + MVO (50–50) 96.77 100 98.36 97.62 2.38 0.9736
5-CV 98.33 100 99.16 98.75 1.25 0.9903
ANN + MFO (50–50) 96.67 100 98.31 97.5 2.5 0.9809
5-CV 96.67 100 98.31 97.5 2.5 0.9818
ANN + SSA (50–50) 96.77 100 96.67 95.24 4.76 0.9631
5-CV 95 100 97.44 96.25 3.75 0.9827

achieves the best results compared to other proposed algorithms as obvious in Tables 4,
6, 7, 8 and 9 and ROC curves in Figs. 8, 9 and 11. The CAs of 99.23%, 98.75% and
ERs of 0.68, 1.25 are obtained using the MVO-ANN approach for Harvard and
BRATS 2015, respectively.

5 Conclusion

In this study, automated hybrid classification approaches are presented to discriminate

between healthy, benign and malignant brain MRI. The proposed technique combined
the MVO, MFO or SSA with ANN in order to achieve the best MRI classification
rate via choosing the ideal ANN parameter. The introduced system has been conduct-
ed on two benchmark datasets i.e. Harvard and BRATS 2015. The optimal CA(ER) of
100(0)% is achieved for each optimized classification techniques at standard and 5-CV
partitions with H/T classification, and it is also achieved at standard partitions with B/
M classification when Harvard datasets are tested. The best performance is achieved
when MVO-ANN is used and CA (ER) of 99.23 (0.68) % and 98.75 (1.25) % is
obtained for Harvard and BRATS 2015, respectively. The experimental results denote
that the proposed approach is the best when compared with traditional ANN and
previously published ones. For future improvement, larger dataset can be evaluated
using an advanced machine learning as deep neural networks. Also, the T2-weighted
images are only evaluated using the proposed approach. Hence, more experiments are
needed to evaluate other types of medical images (i.e. three-dimensional images, T1-
weighted images and CT images).
 Author's personal copy
27812 Multimedia Tools and Applications (2020) 79:27791–27814

References

1. Fink JR, Muizi M, Peck M, Krohn KA (2015) Continuing education: multi-modality brain tumor imaging-
MRI, PET, and PET/MRI. Journal of nuclear medicine: official publication, Society of Nuclear Medicine
56(10):s
2. Wu Ming-Chi, Chin Wen-Chi, Tsan Ting-Chen and Chin Chiun-Li (2016) The benign and Malignant
Recognition System of Nasopharynx in MRI image with Neural-Fuzzy based Adaboost classifier. 2nd
International Conference on Information Management (ICIM), London: 47–51.
3. El-Dahshan E, Hosny T, Salem A (2010) Hybrid intelligent techniques for MRI brain images classification.
Digital Signal Processing 20(2):433–441
4. Amin J, Sharif M, Yasmin M, Fernandes SL (2017) A distinctive approach in brain tumor detection and
classification using MRI. Pattern Recogn Lett:1–10. https://doi.org/10.1016/j.patrec.2017.10.036
5. Preethi G and Sornagopal V (2014) sMRI image classification using GLCM texture features. International
conference on green computing communication and electrical engineering (ICGCCEE), Coimbatore, India:
1-6, https://doi.org/10.1109/ICGCCEE.2014.6922461
6. Anitha R, Siva Sundhara Raja D (2017) Segmentation of glioma tumors using convolutional neural
networks. International Journal of Imaging Systems and Technology 27(4):354–360. https://doi.
org/10.1002/ima.22238
7. Amin J, Sharif M, Yasmin M, Fernandes S (2018) Big data analysis for brain tumor detection: deep
convolutional neural networks. Futur Gener Comput Syst 87:290–297. https://doi.org/10.1016/j.
future.2018.04.065
8. Benson C, Deepa V, Lajish V, and Rajamani K (2016) Brain tumor segmentation from MR brain images
using improved fuzzy c-means clustering and watershed algorithm. International Conference on Advances
in Computing, Communications and Informatics (ICACCI), Jaipur, India: 187–192. https://doi.org/10.1109
/ICACCI.2016.7732045.
9. Chaplot S, Patnaik L, Jagannathan N (2006) Classification of magnetic resonance brain images using
wavelets as input to support vector machine and neural network. Biomed Signal Processing Control 1(1):
86–92. https://doi.org/10.1016/j.bspc.2006.05.002
10. Mirjalili S, Mirjalili S and Hatamlou A (2016) Multi-verse optimizer: a nature inspired algorithm for global
optimization. Neural Computing Appl, 27(2): 495–513, 2016. https://doi.org/10.1007/s00521-015-1870-7
11. Al-Madi N, Faris H, and Mirjalili S (2019) Binary multi-verse optimization algorithm for global optimiza-
tion and discrete problems. Int J Mach Learn Cybernetics: 1-21, 2019. https://doi.org/10.1007/s13042-019-
00931-8
12. Karthikeyan K, Dhal P (2017) Multi verse optimization (MVO) technique based voltage stability analysis
through continuation power flow in IEEE 57 bus. Energy Procedia 117:583–591. https://doi.org/10.1016/j.
egypro.2017.05.153
13. Mirjalili S (2015) Moth-flame optimization algorithm: a novel nature-inspired heuristic paradigm. Knowl-
Based Syst 89:228–249. https://doi.org/10.1016/j.knosys.2015.07.006
14. Abd El Aziz M, Ewees A, Hassanien A (2017) Whale optimization algorithm and moth-flame optimization
for multilevel thresholding image segmentation. Expert Syst Appl 83:242–256. https://doi.org/10.1016/j.
eswa.2017.04.023
15. Mirjalili S, Gandomi A, Mirjalili S, Saremi S, Faris H, Mirjalili S (2017) Salp swarm algorithm: a bio-
inspired optimizer for engineering design problems. Adv Eng Softw 114:163–191. https://doi.org/10.1016/j.
advengsoft.2017.07.002
16. Sayed G, Khoriba G, Haggag M (2018) A novel chaotic salp swarm algorithm for global optimization and
feature selection. Appl Intell 48(10):3462–3481. https://doi.org/10.1007/s10489-018-1158-6
17. Ahmed H, Youssef B, Elkorany A, Saleeb A, Abd El-Samie F (2018) Hybrid gray wolf optimizer–artificial
neural network classification approach for magnetic resonance brain images. Appl Opt 57(7):B25–B31.
https://doi.org/10.1364/AO.57.000B25
18. Virupakshappa, Amarapur B (2018) Computer-aided diagnosis applied to MRI images of brain tumor using
cognition based modified level set and optimized ANN classifier. Multimed Tools Appl 79:3571–3599.
https://doi.org/10.1007/s11042-018-6176-1
19. Zhang Y, Wang S, Ji G, Dong Z (2013) An MR brain images classifier system via particle swarm
optimization and kernel support vector machine. The scientific world journal: 1-9, ID 130134. https://doi.
org/10.1155/2013/130134
20. Mishra S, Sahu P, Senapati M (2019) MASCA–PSO based LLRBFNN model and improved fast and robust
FCM algorithm for detection and classification of brain tumor from MR. Image Evol Intel 12:647–663.
https://doi.org/10.1007/s12065-019-00266-x
 Author's personal copy
Multimedia Tools and Applications (2020) 79:27791–27814 27813

21. Kharrat A, Gasmi K, Messaoud M, Benamrane N, Abid M (2010) A hybrid approach for automatic
classification of brain MRI using genetic algorithm and support vector machine. Leonardo J Sci 9(17):
71–82
22. Kaur T, Saini B, Gupta S (2019) An adaptive fuzzy K-nearest neighbor approach for MR brain tumor image
classification using parameter free bat optimization algorithm. Multimed Tools Appl 78:21853–21890.
https://doi.org/10.1007/s11042-019-7498-3
23. Bhuvaneswari K, Geetha P (2017) Segmentation and classification of brain images using firefly and hybrid
kernel-based support vector machine. J Exp Theor Artif Intell 29(3):663–678. https://doi.org/10.1080
/0952813X.2016.1212106
24. Bahadure N, Ray A, Thethi H (2017) Feature extraction and selection with optimization technique for brain
tumor detection from MR images. Int Conf Computational. Intel Data Sci (ICCIDS): 1–7. https://doi.
org/10.1109/ICCIDS.2017.8272635
25. Zhang Y-D, Jiang Y, Zhu W et al (2018) Exploring a smart pathological brain detection method on pseudo
Zernike moment. Multimed Tools Appl 77:22589–22604. https://doi.org/10.1007/s11042-017-4703-0
26. Patil D, Hamde S (2019) Brain MR imaging tumor detection using monogenic signal analysis-based
invariant texture descriptors. Arab J Sci Eng 44:9143–9158. https://doi.org/10.1007/s13369-019-03989-2
27. Sasikanth S, Suresh Kumar S (2017) Glioma tumor detection in brain MRI image using ANFIS-based
normalized graph cut approach. Int J Imaging Syst Technol 28(1):64–71. https://doi.org/10.1002/ima.22257
28. Rufus H, Selvathi D (2018) Performance analysis of brain tissues and tumor detection and grading system
using ANFIS classifier. Int J Imaging Syst Technol 28(2):77–85. https://doi.org/10.1002/ima.22258
29. Nayak D, Dash R, Majhi B (2018) Discrete ripplet-II transform and modified PSO based improved
evolutionary extreme learning machine for pathological brain detection. Neurocomputing 282:232–247
30. Devkota B, Alsadoon A, Prasad P, Singh A, Elchouemi A (2018) Image segmentation for early stage brain
tumor detection using mathematical morphological reconstruction. Procedia Computer Science 125:115–
123. https://doi.org/10.1016/j.procs.2017.12.017
31. Preetha R, Suresh G (2014) Performance analysis of fuzzy C means algorithm in automated detection of
brain tumor. IEEE world congress on computing and communication technologies (WCCCT),
Trichirappalli, India: 30-33.
32. Ben George E, Jeba Rosline G, Gnana Rajesh D (2015) Brain tumor segmentation using cuckoo search
optimization for magnetic resonance images. 8th IEEE GCC conference and exhibition, Muscat, Oman: 1-6.
33. Bhakat S, Periannan S (2019) Brain tumor detection using cuckoo search algorithm and histogram
Thresholding for MR images. Smart innovations in communication and computational sciences. Advanc
Intel Syst Comput 851:85–95
34. Kalpana R, Chandrasekar P (2020) An optimized technique for brain tumor classification and detection with
radiation dosage calculation in MR image. Microprocess Microsyst 72:1–7. https://doi.org/10.1016/j.
micpro.2019.102903
35. Selvapandian A, Manivannan K (2018) Fusion based Glioma brain tumor detection and segmentation using
ANFIS classification. Comput Methods Prog Biomed 166:33–38. https://doi.org/10.1016/j.
cmpb.2018.09.006
36. Arunkumar N, Mohammed MA, Abd Ghani MK, Ibrahim D, Abdulhay E et al (2019) K-means clustering
and neural network for object detecting and identifying abnormality of brain tumor. Soft Comput 23:9083–
9096. https://doi.org/10.1007/s00500-018-3618-7
37. Pereira S, Pinto A, Alves V, Silva CA (2016) Brain tumor segmentation using convolutional neural
networks in MRI images. IEEE Trans Med Imaging 35(5):1240–1251. https://doi.org/10.1109
/TMI.2016.2538465
38. Dong H, Yang G, Liu F, Mo Y, Guo Y (2017) Automatic brain tumor detection and segmentation using U-
net based fully convolutional networks. Medical image understanding and analysis. MIUA 2017. Comm
Comp Inform Sci 723:506–517. https://doi.org/10.1007/978-3-319-60964-5_44
39. Amin J, Sharif M, Yasmin M, Fernandes S (2018) Big data analysis for brain tumor detection: deep
convolutional neural networks. Futur Gener Comput Syst 87:290–297. https://doi.org/10.1016/j.
future.2018.04.065
40. Yang T, Song J, Li L (2019) A deep learning model integrating SK-TPCNN and random forests for brain
tumor segmentation in MRI. Biocybernetics Biomed Eng 39(3):613–623. https://doi.org/10.1016/j.
bbe.2019.06.003
41. Wang Y, Lia C, Zhua T, Zhang J (2019) Multimodal brain tumor image segmentation using WRN-PPNet.
Comput Med Imaging Graph 75:56–65
42. Nemaa S, Dudhanea A, Muralaa S, Naidu S (2020) RescueNet: an unpaired GAN for brain tumor
segmentation. Biomed Signal Process Control 55:1–8. https://doi.org/10.1016/j.bspc.2019.101641
 Author's personal copy
27814 Multimedia Tools and Applications (2020) 79:27791–27814

43. Abd-Ellah M. K , Awad A. I, Khalaf A. A. M, and Hamed H. F. A (2016) Design and implementation of a
computer-aided diagnosis system for brain tumor classification. 28th Int Conf Microelectron(ICM), Giza,
Egypt: 73–76. https://doi.org/10.1109/ICM.2016.7847911
44. Abdullah H. N. and Habtr M. A (2015) Brain Tumor Extraction Approach in MRI Images Based on Soft
Computing Techniques. 8th International Conference on Intelligent Networks and Intelligent Systems
(ICINIS), Tianjin, China: 21–24. https://doi.org/10.1109/ICINIS.2015.29
45. Shastri A, Tamrakar D, Ahuja K (2018) Density-wise two stage mammogram classification using texture
exploiting descriptors. Expert Syst Appl 99:71–82. https://doi.org/10.1016/j.eswa.2018.01.024
46. Ural B, Yüksek M, Muratdağ F (2017) An Expert Automated Preliminary Diagnostic System for
Identifying Brain Tumors’ Features and Types. Proceedings of the 6th MICCAI BraTS Challenge
(2017): 284–291
47. http://www.med.harvard.edu/AANLIB/
48. http://www.smir.ch/BRATS/Strat2015
49. Gudigar A, Raghavendra U, San T. R, Ciaccio E, Acharya U (2019) Application of multiresolution analysis
for automated detection of brain abnormality using MR images: a comparative study. Futur Gener Comput
Syst 90: 359–367. https://doi.org/10.1016/j.future.2018.08.008
50. Zhang G, Wang Q, Lee E, Ji G, Wang S, Yan J, Zhang Y (2015) Automated classification of brain MR
images using wavelet-energy and support vector machines. International Conference on Mechatronics,
Electronic, Industrial and Control Engineering (MEIC), Shenyang, China, pp 683–686
51. Zhou X, Wang S, Xu W, Ji G, Phillips P, Sun P, and Zhang Y (2015) Detection of Pathological Brain in
MRI Scanning Based on Wavelet-Entropy and Naive Bayes Classifier. International Work-Conference on
Bioinformatics and Biomedical Engineering (IWBBIO), Granada, Spain:683–686.
52. Yadav G, Maheshwari S, Agarwal A (2014) Contrast limited adaptive histogram equalization-based
enhancement for real time video system. International Conference on Advances in Computing,
Communications and Informatics (ICACCI), New Delhi, pp 2392–2397
53. https://www.microdicom.com/downloads.html

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