Multimedia Tools and Applications (2020) 79:14009–14030

https://doi.org/10.1007/s11042-020-08643-w

Combining optimal wavelet statistical texture

and recurrent neural network for tumour detection
and classification over MRI

S. Salma Begum 1 & D. Rajya Lakshmi 2

Received: 31 January 2019 / Revised: 7 December 2019 / Accepted: 2 January 2020 /

Published online: 6 February 2020
# Springer Science+Business Media, LLC, part of Springer Nature 2020

Abstract
Brain tumor is one of the major causes of death among other types of the cancer because
brain is a very sensitive, complex and central part of the body. Proper and timely diagnosis
can prevent the life of a person to some extent. Therefore, in this paper, an efficient brain
tumor detection system is proposed using combining optimal wavelet statistical texture
features and recurrent neural network (RNN). The proposed system consists of four phases
namely; feature extraction feature selection, classification and segmentation. First, noise
removal is performed as the preprocessing step on the brain MR images. After that, texture
features (both the dominant run length and co-occurrence texture features) are extracted
from these noise free MR images. The high number of features is reduced based on
oppositional gravitational search algorithm (OGSA). Then, selected features are given to
the Recurrent Neural Network (RNN) classifier to classify an image as normal or
abnormal. After the classification process, abnormal images are given to the segmentation
stage to segment the ROI region with the help of modified region growing algorithm
(MRG). The performance of the proposed methodology is analyzed in terms of different
metrics and experimental results are compared with existing methods.

Keywords Brain tumor . Wavelet statistical texture . Recurrent neural network . Feature extraction
. Segmentation . Dominant run length . Co-occurrence texture features

1 Introduction

An abnormal growth of cells within the brain or the central spinal canal is a brain tumor or an
intracranial solid neoplasm. Brain tumor, one of the most common and deadly diseases in the

* S. Salma Begum
[email protected]

1
Department of Computer Science and Engineering, JNTUK University, Kakinada, India
2
Principal, Professor of CSE, JNTUK, Narasaraopet, India
14010 Multimedia Tools and Applications (2020) 79:14009–14030

world can be cured if detection is done in its early stage. There are various types of brain
tumors that make the decision very complicated [3]. The classification of the type of brain
tumor really suffered by patient is very important because a good classification process leads to
the right decision thereby providing good and right treatment [35]. In most cases, the physician
gives the treatment for the strokes rather than the treatment for the tumor. Detection of the
tumor is essential for the treatment since we find in most cases we find the physician giving the
treatment for the strokes rather than the treatment for the tumor. The lifespan of a person
affected by the brain tumor is increased when detected early [31]. Early stage brain tumor
diagnose mainly includes Computed Tomography (CT) scan, Magnetic Resonance Imaging
(MRl) scan, Nerve test, Biopsy etc. [7]. With the rapid growth of the Artificial Intelligence (AI)
development in Biomedicine, computer-aided diagnosis and Magnetic Resonance Imaging
(MRl) has gained more attention [26].Brain cancer is a very serious type of malignancy that
occurs when there is an uncontrolled growth of cancer cells in the brain. Brain cancer though
caused by a malignant brain tumor. All brain tumors are not malignant (cancerous). Some
types of brain tumors are benign (non-cancerous). Brain cancer is also known as glioma and
meningioma [35].
Two important steps in breast cancer detection and classification are Feature extraction and
selection. Mostly an optimum feature set should have effective and discriminating features and
also reduce the redundancy of features pace to avoid “curse of dimensionality” problem [26]
.Feature selection strategies are applied to explore the effect of irrelevant features on the basis
of their performance of classifier systems [1, 39, 41]. An optimal subset of features which are
necessary and sufficient to solve a problem is selected in this phase. Feature selection improves
the accuracy of algorithms in two ways-by reducing the dimensionality and removing irrele-
vant features [18] [24]. Feature extraction of image is an important step in tumor classification
since there are several types of feature extraction from digital mammograms including position
feature, shape feature and texture feature etc. Several techniques are developed for feature
extraction from tumor image. After the process of feature extraction, feature selection process
is more important. Feature selection (also known as subset selection) a process commonly used
in machine learning, utilizes a subset of the features available from the data. This is selected for
application of a learning algorithm. Feature selection algorithms may be divided into filters [9],
wrappers [2] and embedded approaches [6]. Filters method is used to evaluate quality of
selected features, independently from the classification algorithm. But wrapper methods
require application of a classifier to evaluate this quality. Embedded methods perform feature
selection during learning of optimal parameters Embedded methods perform feature selection.
Different classification methods that range from statistical and machine learning area are
applied to cancer classification. Classification is a basic task applied in data analysis and
pattern recognition and requires the construction of classifiers. There are many machine
learning techniques applied to classify the tumor, including Fisher linear Discriminate analysis
[10], k-nearest neighbour [23] decision tree, multilayer perceptron [20], and support vector
machine [17]. A recent comparison done on classification and feature selection algorithms
applied to tumor classification can be found in [8, 38]. In order to perform brain tumor
detection and classification numerous algorithms have been developed. These methods include
thresholding and morphological techniques [4, 5], watershed method [11], region growing
approach [13, 30], asymmetry analysis [15, 33], atlas-based method [27, 40], contour/surface
evolution method [32, 37], interactive algorithm [12, 16], and supervised unsupervised [13]
learning methods. At the same time, Probabilistic Neural Network (PNN) is used for computer
aided brain tumor classification which utilized the feed-forward neural network. This helps to
Multimedia Tools and Applications (2020) 79:14009–14030 14011

identify the type of brain tumor suffered by patient regarding to the image of brain tumor from
the Magnetic Resonance Imaging (MRl) [35].
In this paper, we explain a brain tumor image classification and segmentation based on
combining optimal wavelet statistical texture features and RNN. Here, texture features are
extracted from each image and important features are selected with the help of optimization
algorithm. Then, selected features are given to the RNN classifier to classify an image as
normal or abnormal. Then, in abnormal image ROI part is segmented. The main contribution
of the paper is explained below;

& For classification, wavelet statistical features are extracted from each image. The feature is
a combination of both dominant run length and co-occurrence texture features.
& To reduce the complexity and time, important features are selected with the help of OGSA.
Here, oppositional based learning (OBL) strategy is combined with GSA to increase the
searching ability.
& Then, selected features are given to the RNN to classify an image as normal or abnormal.
& The ROI region of abnormal images is separated with the help of MRG algorithm. This
algorithm overcomes the difficulties present in the region growing algorithm (RG).

The rest of the paper organized as follows: Section 2 presents the review of related works and
the proposed heart disease prediction is explained in section 3. The result and discussion
explained in section 4. The conclusion part is presented in section 5.

2 Related works

In the literature survey, several methods are proposed for brain tumor classification in image
processing. Among them the most recently published works are presented here: In, Meiyan Huang
et al. [14] explained the Brain Tumor Segmentation Based on Local Independent Projection-based
Classification which treats tumor segmentation as a classification problem. Each voxel was classified
into different classes using the local independent projection-based classification (LIPC) method. By
introducing the local independent projection into the classical classification model a novel classifi-
cation framework was derived. Locality was important in the calculation of local independent
projections for LIPC, which was taken into consideration in determining whether local anchor
embedding was more applicable in solving linear projection weights compared with other coding
methods. Moreover, LIPC considers the data distribution of different classes. This is done by
learning a softmax regression model, which further improved classification performance.
Moreover, Zhan-Li Sun et al. [36] have explained the tumor classification using Eigengene
-based classifier committee learning algorithm. Here, we find Eigengene extracted by inde-
pendent component analysis (ICA) was one kind of effective feature for tumor classification
which used Eigen gene and support vector machine based classifier committee learning (CCL)
algorithm. In order to improve the diversity of weaker classifiers the random feature subspace
division was designed. Gene expression data constructed by different feature subspaces were
modeled by ICA and the corresponding eigengene sets extracted by the ICA algorithm were
used as the inputs of the weaker SVM classifier. A strategy of Bayesian sum rule (BSR) was
designed to integrate the outputs of the weaker SVM classifiers, used to provide a final
decision for the tumor category. Experimental results on three DNA microarray datasets
demonstrate the fact that it is effective and feasible for tumor classification.
14012 Multimedia Tools and Applications (2020) 79:14009–14030

Additionally, Kailash D.Kharat et al. [21] have explained the two Neural Network tech-
niques for the classification of the magnetic resonance human brain images. The Neural
Network technique constitutes three stages namely, feature extraction, dimensionality reduc-
tion, and classification. In the first stage, the features obtained were related with MRI images
using discrete wavelet transformation (DWT). In the second stage, the features of magnetic
resonance images (MRI) was reduced making use of principles component analysis (OGSA)
to the more essential features. In the classification stage, two classifiers based on supervised
machine learning were developed. The first classifier was based on feed forward artificial
neural network (FF-ANN) and the second classifier based on Back-Propagation Neural
Network. The classifiers functioned to classify subjects as normal or abnormal MRI brain
images. Artificial Neural Networks (ANNs) were developed using a wide range of applications
such as function approximation, feature extraction, optimization, and classification particularly
developed for image enhancement, segmentation, registration, feature extraction, and object
recognition and classification.
Likewise, Pankaj Sapra et al. [34] explained the Brain Tumor Detection Using
Neural Network where modified image segmentation techniques were applied on MRI
scan images for the detection of brain tumors. Using MRI-scans a modified Probabi-
listic Neural Network (PNN) model that was based on learning vector quantization
(LVQ) with image and data analysis and manipulation techniques were explained to
carry out an automated brain tumor classification. The assessment of the modified
PNN classifier performance was measured in terms of three factors namely training
performance, classification accuracies and computational time. The simulation result
showed the modified PNN giving rapid and accurate classification compared with the
image processing and published conventional PNN techniques. Simulation results also
showed that the system out performs the corresponding PNN system. This was so
when it was presented to successfully handle the process of brain tumor classification
in MRI image with 100% accuracy.
In, A.P. Nanthagopal and R.Sukanesh [25] have explained the Wavelet statistical
texture features-based segmentation and classification of brain computed tomography
images where the authors presented a method to select both dominant run length and co-
occurrence texture features of wavelet approximation. The tumor region of each slice
was to be segmented by a support vector machine (SVM In order to remove the noise
two-dimensional discrete wavelet decomposition was performed on the tumor image.
Seventeen features are extracted and six features are selected using Student’s t test. This
method was constructed using the SVM and probabilistic neural network (PNN) classi-
fiers with the selected features and the classification accuracy of both classifiers was
evaluated using the k fold cross validation method. The segmentation result was also
compared with the experienced radiologist ground truth. Quantitative analysis between
ground truth and the segmented tumour presentation was done in terms of segmentation
accuracy and segmentation error.
Moreover, Quratul Ain et al. [29] explained the Fuzzy anisotropic diffusion based
segmentation and texture based ensemble classification of brain tumor. The system
was used to classify multi-stage system for brain tumor diagnosis and tumor region
extraction. As the preprocessing step on the brain MR images noise removal was
performed. Texture features were extracted from these noise free brain MR images
and in the next phase of their system classification was done based on these extracted
features. Ensemble based SVM classification was used which helped to achieve more
Multimedia Tools and Applications (2020) 79:14009–14030 14013

than 99% accuracy After classification, using multi-step segmentation their system
extracts tumor region from tumorous images. Here the first step was skull removal
and brain region extraction and the next step was separating tumor region from
normal brain cells using FCM clustering. Results of their technique show that the
extraction of the tumor region was quite accurate.

3 Proposed brain tumor detection system

The basic idea of our research is to detect and segment tumor from MRI brain images using
multiple stages. A feature extraction method along with segmentation and classification
methods is linked in the presented method in order to find out the brain as normal or tumours.
The overall proposed framework is illustrated in Fig. 1. Basically, MR images are given to the
system for the diagnosis purpose. Initially, these images are processed for noise removal. After
elimination of noise, texture features are taken out from these images with the help of Wavelet
statistical texture features (combining dominant run length and co-occurrence texture). After
that, the important features are selected with the help of OGSA. Due to adopting this feature
selection technique higher accuracy for classification is achieved by this system. Then, these
selected features are given to the RNN classifier. Finally, the classified tumors images are
given to the MRG, which is segment the ROI part from the image. The step by step process of
proposed brain tumor detection system is explained in following section.

Fig. 1 Overall diagram of the proposed MRI brain tumor detection system
14014 Multimedia Tools and Applications (2020) 79:14009–14030

3.1 Noise removal

The Noise removal of MR image is a significant stage to segmentation and classification

procedure. Deliberate the gray scale input image Iin, which has few of the noise like a Gaussian
noise etc. the noise image certainly disturb the output. So noise removal is necessary part
before any further processing on images, because the removal of noise from the image gives
the better accuracy for classification stage. In this work, for noise removal stage we utilize the
Gaussian filter, which is used to remove the noise from the image. The noise removal image I
is used for the further processing.

3.2 Feature extraction stage

Feature extraction is a significant phase in the medical image classification. The alter-
ation of an image into its set of features is called as feature extraction. The objective of
feature extraction is to abstract expressive features from the image, as they can be
categorized depending on their source. It is a stimulating mission to remove good feature
set for classification. Currently, there are various methods for feature extraction is
presented. In this work, we are utilizing Wavelet statistical texture features (combining
dominant run length and co-occurrence texture). Texture analysis is a quantifiable
technique that can be castoff to enumerate and perceive the structural irregularities in
various tissues. The texture feature extraction is originated to be very vital for additional
cataloguing because the tissues present in brain are tough to categorize with the help of
the form or the strength level of data. Here, we extracted17 spatial features from input
tumor image with the help of both dominant gray-level run length and gray-level co-
occurrence matrix.
The dominant gray-level run length matrix M (d, θ) is as follows;
M ðd; θÞ ¼ ½I ði; j j d; θÞ 0 < i ≤N g 0 < i ≤Rmax ð1Þ

Where; Ngis the Maximum gray level and Rmax is the Maximum run length. The element I(i,
j | θ)indicated the projected amount of runs that a given image contains a run length j for a
gray-level i in the direction of angle θ. Every image consuming four dominant gray-level run
length matrices conforming to θ = 0∘, 45∘, 90∘, 135∘. In this study, we analyze four dominant
run length texture features such as short run low gray-level emphasis (SRLGE), short run high
gray-level emphasis (SRHGE), Long run low gray-level emphasis (LRLGE) and long run high
gray-level emphasis (LRHGE). These four features are calculated in all the four directions for
each image. The extracted four features are given Table 1.
In the above Table 1, Nris the total number of runs, Mis the amount of gray level and Nis the
maximum run length. These four types of features compute all the four dominant gray-level
run length matrices and then yield the average of all the features extracted from four dominant
gray-level run length matrices.
The gray level co-occurrence matrix M(d, θ) is as follow;
M ðd; θÞ ¼ ½P ði; j j d; θÞ 0 < i ≤N g ð2Þ

Where, Ngis the maximum gray level. The utility P(i, j | d, θ)is the probability matrixes of two
pixels, which are situated with inter sample distance d and direction θ has a gray level i and
gray levelj. Every image is consuming gray-level co-occurrence matrices conforming to θ = 0∘,
Multimedia Tools and Applications (2020) 79:14009–14030 14015

Table 1 Four type of dominant run length texture features

Features name Computation

Short run low gray-level emphasis (SRLGE), M N

I ði; jÞ
SRLGE ¼ N1r ∑ ∑ i2 : j2
i¼1 j¼1
Short run high gray-level emphasis (SRHGE) M N
I ði; jÞ:i2
SRHGE ¼ N1r ∑ ∑ j2
i¼1 j¼1
Long run low gray-level emphasis (LRLGE) M N
I ði; jÞ: j2
LRLGE ¼ N1r ∑ ∑ i2
i¼1 j¼1
Long run high gray-level emphsis (LRHGE) M N
LRHGE ¼ N1r ∑ ∑ I ði; jÞ:i2 :j2
i¼1 j¼1

45∘, 90∘, 135∘. In this work, 15 textual based haralick features are extracted from the input
tumor image. The extracted 15 types of haralick features are shows in Table 2.
The above Table 3 illustrates the extracted features from MRI image using gray-level co-
occurrence matrix. These 15 types of features are calculated all the four gray-level co-
occurrence matrix and take the average of all the features extracted from four gray-level co-
occurrence matrices.

3.3 Feature selection using oppositional gravitational search algorithm (OGSA)

After the feature selection process, important features are selected with the help of OGSA.
GSA is inspired based on the newton’s law. To increase the searching ability of the GSA,
opposition based learning (OBL) strategy is included with GSA. The step by step process of
feature selection is explained below;

Step 1: Solution initialization: Solution initialization is an important process for optimiza-

tion problems. In this problem, initially, the important features are randomly selected.
The solution initialization format is given in Table 3.

In Tables 3, 1 represent corresponding feature is selected and 0 represent, that feature is not
selected.

Step 2: Opposite solution generation: After the random solution generation, we immedi-
ately calculate the opposite solution of initial solution based on eq. (3).
0
S i ¼ Li þ U i −S i ð3Þ

Step 3: Fitness calculation: After the solution generation process, we calculate the fitness of
entire solution. Based on the fitness value only we can select the optimal features. In
this paper, maximum accuracy is considered as the fitness function.

Fitness ¼ maxðAccuracyÞ ð4Þ

14016 Multimedia Tools and Applications (2020) 79:14009–14030

Table 2 Extracted gray-level co-occurrence matrix features

Features name Computation

Angular Second Moment GL −1 GL −1

ASM ¼ ∑ ∑ fpðx; yÞg2
x¼0
( y¼0
)
Contrast GL −1 GL
C ¼ ∑ n2 ∑ ∑ pðx; yÞ ; jx−yj ¼ n
n¼0 x¼1 y¼1
Inverse difference moment GL −1 GL −1
IDM ¼ ∑ ∑ 1
1þðx−yÞ2
pðx; yÞ
x¼0 y¼0
entropy G −1 G −1
L L

E¼− ∑ ∑ pðx; yÞ  log ðp ðx; yÞÞ

x¼0 y¼0
correlation G −1 GL −1 fx;yg p ðx;yÞ−fμ μ g
L

COR ¼ ∑ ∑ σi σy
i y

x¼0 y¼0
Variance GL −1 GL −1
VAR ¼ ∑ ∑ ðx−μÞ2 p ðx; yÞ
i¼0 j¼0
Sum average 2GL −1
Aver ¼ ∑ x piþ j ðxÞ
x¼0
Sum variance 2 GL −1
S Var ¼ ∑ ðx−AverÞ2 piþ j ðxÞ
x¼0
Sum entropy 
2GL −2 
Sent ¼ − ∑ Piþ j ðxÞ log piþ j ðxÞ
x¼0
Difference entropy 
GL −1 
Dent ¼ − ∑ Piþ j ðxÞ log piþ j ðxÞ
x¼0
Inertia GL −1 GL −1
I¼ ∑ ∑ ðx−yÞ  pðx; yÞ
x¼0 x¼0
Cluster shade GL −1 GL −1 n o3
C shade ¼ ∑ ∑ x þ y−μi −μy  pðx; yÞ
x¼0 y¼0
Cluster prominence GL −1 GL −1 n o4
C Pro
¼ ∑ ∑ x þ y−μi −μy  pðx; yÞ
x¼0 y¼0
Dissimilarity Dsim ¼ ∑ jx−yj pðx; yÞ
x;y
Homogeneity 1
H¼∑ pðx; yÞ
x;y 1−ðx−yÞ2

TP þ TN
Accuracy ¼ ð5Þ
T P þ FP þ FN þ T N

Table 3 Solution initialization

format F1 F2 F3 F4 …. FM

R1 1 0 1 0 …. 1
R2 0 1 0 1 …. 0
R3 1 1 0 1 …. 1
…. 1 0 1 0 …. 1
RN 0 1 0 1 …. 0
Multimedia Tools and Applications (2020) 79:14009–14030 14017

Step 4: Computation of gravitational constant (G): The gravitational constant G is

calculated using eq. (6);

Gðk Þ ¼ G0 eð−αk=K Þ ð6Þ

G0 and α are initialized at the beginning and will be reduced with time to control the search
accuracy. K is the total number of iterations.

Step 5: Masses of the agents’ calculation: In GSA, the mass of each agent is a performance
of iteration and hinge on the fitness of an agent. The mass shrewdness of each agent
as follows

fnsi ðk Þ−worst ðk Þ
mi ðk Þ ¼ ð7Þ
best ðk Þ−worstðk Þ

mi ð k Þ
M i ðk Þ ¼ N
ð8Þ
∑ m j ðk Þ
j¼1

Where, Mi(k) is the standardized mass of ith agent at kthiteration andworst (k), best (k) are the
worst and best fitness of all agents at kthiteration.

Step 6: Accelerations of agents’ calculation: The acceleration Adi ðk Þ acting on ith agent at
iteration k is evaluated as follows

M i ðk Þ  d 
adi ðk Þ ¼ ∑ rand j Gðk Þ x j ðk Þ−xdi ðk Þ ð9Þ
j∈gbest; j≠i Rij ðk Þ þ ε

Where gbest is the set of first2%agents with the best fitness value and biggest mass, randj is the
uniform random number between interval [0,1], Rij(k) is the Euclidean distance between two
agents ith and jthat tth iteration and ε is a small positive constant.

Step 7: Agent’s velocity and position updates: The agent’s velocity and position for next
(k + 1)thiteration are calculated as follows;

V di ðk þ 1Þ ¼ rand i  V di ðk Þ þ adi ðk Þ ð10Þ

14018 Multimedia Tools and Applications (2020) 79:14009–14030

Y di ðk þ 1Þ ¼ Y di ðk Þ þ V di ðk þ 1Þ ð11Þ

Where, randi is the random number between interval [0,1]. V di ðk þ 1Þ is the velocity of ith the
agent at dth dimension during tthiteration and Y di ðk þ 1Þ is the position of ith agent at
dthdimension during kthiteration.

Step 8: Termination: Steps 3 to 7 are repeated until the iterations reach their maximum
limit. The best fitness solution is chosen as the optimal features. The selected
features are given to the classification process.

3.4 Brain tumor classification using RNN

After the feature selection, the selected features are given to the classification stage.
Classification is the procedure where a given test model is allocated a class based on
information grew by the classifier at the time of training and which classifies the
unidentified information samples. Selection of a suitable classifier requires consider-
ation of many factors like computational resources it used, accuracy of the classifier
for several datasets, and performance of the algorithm. Based on the requirement, in
this we adapt the recurrent neural network (RNN) for classification. The basic corner
stone of a RNN are the neurons effectively linked by the synaptic links (connections)
whose synaptic strength is appropriately coded by a weight. Basically, it is possible to
effectively differentiate between the input units, internal (hidden) units, and the output
units. At a specified time, a unit is subject to activation. Here we use, feed-forward
neural networks trained with the back-propagation algorithm to brain tumor classifi-
cation. The network consists of an input layer, an output layer, with one or more
hidden layers in between the input and output layer.
Let us consider, N number of image features are furnished to the input of the RNN, which
yields the L number of output unit and of units in hidden layer m isNm. The weight of the jthunit
in layer m and the ithunit in m + 1 is represented byWij. The activation of the ith unit in layer mis
xm
i (for m = 0 this is an input value, for m = k + 1an output value). The training data for a feed
forward network training task comprises the T input-output (vector-valued) data pairs.

 t  t
uðnÞ ¼ x01 ðnÞ; …; xok ðnÞ d ðnÞ ¼ d kþ1
1 ðnÞ; …; d kþ1
L ðnÞ
ð12Þ

Where, n represents the training instance, not the time. The activation of non-input units is
evaluated in accordance with the following Equation.
!
xmþ1
i ðnÞ ¼ F ∑ ij x j ðnÞ
wm ð13Þ
j¼1;:::N m
Multimedia Tools and Applications (2020) 79:14009–14030 14019

The prior modernization equation is employed to evaluate the activations of units in successive
hidden layers, till a network response is represented as shown below.

 0
1 ðnÞ; …; xL ðnÞ
yðnÞ ¼ ykþ1 kþ1
ð14Þ

Equation (5) is attained in the output layer. The objective of training is to locate a set of
network weights in such a way that the summed squared error is represented as follows.

E¼ ∑ kd ðnÞ−yðnÞk2 ¼ ∑ E ðnÞ ð15Þ

n¼1;:::T n¼1;:::T

The error value Eis reduced. This is carried out by incrementally varying the weights in the
direction of the error gradient w.r.t. weights

∂E ∂E ðnÞ
¼ ∑ ð16Þ
∂wmij n¼1;::;T ∂wij

New weight is estimated as illustrated in Eq. (17).

∂E
ij ¼ wij −γ
new wm ð17Þ
m
∂wmij

This is the formula employed in the batch learning mode, where new weights are
evaluated after furnishing the entire training samples. One such pass through all the
samples is known as an epoch. Prior to the first epoch, weights are initialized,
classically to trivial arbitrary numbers. A variant represents the incremental learning,
where weights are modified after presentation of the individual training samples:

∂E ðnÞ
ij ¼ wij −γ
new wm ð18Þ
m
∂wmij

The vital subtask in this technique is the evaluation of the error gradients∂E ðnÞ
∂wm . The back
ij

propagation technique represents a novel mechanism to carry out the related evaluations. A
steps for one epoch of batch processing is furnished below.
14020 Multimedia Tools and Applications (2020) 79:14009–14030
Multimedia Tools and Applications (2020) 79:14009–14030 14021

Subsequent to every such epoch, the fault is evaluated in accordance with (8). The function
has to be stopped when the fault becomes lesser than a fixed threshold, or when the variation in
error goes below a different fixed threshold, or when the number of epochs is in excess of a
fixed maximal number of epochs. A multitude of similar epochs (amounting in thousands in
respect of significant functions) is likely to be needed till a satisfactorily minor fault is attained.
In the long run, the score value is achieved which effectively decides whether the specified
image is tumour or not. If the score value exceeds the threshold value, it indicates that the
specified data is a case of intrusion. If it is less than or equal to the threshold value, the
specified image is treated as tumour image. Thus the obtained score value is assessed with the
condition (12) which is given in below for the purpose of categorizing the data.

T h ≥score; no tumour
Decision ¼ ð19Þ
T h < score; tumour image

3.5 ROI segmentation using modifier region growing algorithm

After the image classification process, the tumor images are selected and given to the
segmentation stage. In this work, for segmentation stage we utilize the Modified Region
Growing (MRG) algorithm. Region growing method is a popular technique for image
segmentation which involves seed point selection. In the segmentation process, the neighbor-
ing pixels are compared with the initial seed points to check whether the neighboring pixels
can be added to the region or not. Seed point selection is important task in the segmentation.
But, this normal Region Growing method selects the seed points by setting the intensity
threshold, which has drawbacks of noise or variation in intensity that leads to over-
segmentation or holes. Moreover, the shadings of real images may not be differentiated by
this method. To overcome these difficulties, we modify the Region growing method by
considering intensity and orientation thresholds from the input images to utilize those features
in the selection of seed points. The process of MRG method is given in steps which are shown
below:

Step 1: At first, we calculate the gradient of the image I for both x axis (IR x) and yaxis(IR y).
Step 2: After that, we calculate gradient vector GVby combining the gradient values using the
following Eq. (20).

1
GV ¼   ð20Þ
1 þ I R x2 þ I R y2

Step 3: Change the gradient vector values that are usually in radians into degrees to get the
values of orientation.
Step 4: Separate the image into gridsGi.
Step 5: Set intensity threshold (TIN) and orientation threshold (TOR).
Step 6: For every gridGi, continue the following processes in step 7 until the number of grids
reached total number of grids for an image.
14022 Multimedia Tools and Applications (2020) 79:14009–14030

Step 7: (a): Find the histogram H of each pixel in Gi.

(b): Determine the most frequent histogram of the Gith grid and denote it asFH.
(c): Prefer any pixel, according to FH and assign that pixel as seed point which has the
intensity INp and OrientationORp.
(d): Consider the neighboring pixel having the intensity INn and orientationORn.
(e): Find the intensity and orientation difference of those pixels p and n.

ði:e:Þ DIN ¼ ‖IN p −IN n ‖ ð21Þ

and DOR ¼ ‖IN p −IN n ‖ ð22Þ

(f): If DIN ≤ TIN && DOR ≤ TOR, then add the corresponding pixel to the region and the region is
grown, else move to step 7(h).
(g): Check whether all pixels are added to the region. If true go to step 6 otherwise go to
step 7(h).
(h): Re-estimate the region and find the new seed points and do the process from step 7(a).

Step 8: Stop the whole process.

Using this Modified Region Growing process, the input images are gets segmented. The
segmented image output is shows in Fig. 2.

4 Result and discussion

In this section, we discuss the result obtained from the proposed brain tumor classification and
segmentation technique. For implementing the proposed technique, we have used Mat lab
version (7.12). This proposed technique is done in windows machine having Intel Core i5
processor with speed 1.6 GHz and 4 GB RAM. The proposed system has been tested on the
data set available at web. We have utilized the size of the image “512 × 512” which images are
publicly available.

Fig. 2 Segmentation output (a) Input image (b) Segmented output

Multimedia Tools and Applications (2020) 79:14009–14030 14023

4.1 Evaluation metrics

We need various assessment metric values to be calculated in order to analyze our proposed
technique for the efficient MRI brain tumor classification. The metric values are found based
on True Positive (TP), True Negative (TN), False Positive (FP) and False Negative (FN) with
the option of segmentation and grading. The usefulness of our proposed work is analyzed by
three metrics such as Accuracy, Sensitivity and Specificity. The demonstration of these
assessment metrics are specified in equations that given below.

Sensitivity The sensitivity of brain tumor detection is determined by taking the ratio of
number of true positives to the sum of true positive and false negative. This relation can be
expressed as;
Tp
St ¼
T p þ Fn

Specificity The specificity of the brain tumor detection can be evaluated by taking the relation
of number of true negatives to the combined true negative and the false positive. The
specificity can be expressed as;
Tn
Sp ¼
T n þ Fp

Accuracy The accuracy of brain tumor detection can be calculated by taking the ratio of true
values present in the population. The accuracy can be described by the following equation;
Tp þ Tn
A¼
T p þ Fp þ Fn þ T n

4.2 MRI dataset description

The MRI images are effectively employed in the innovative tumor segmentation and classifi-
cation. The equivalent collected image dataset have 1000 MRI images. Here, 750 lungs images

Fig. 3 Experimental used sample images

14024 Multimedia Tools and Applications (2020) 79:14009–14030

Fig. 4 Performance of accuracy plot based on classification stage

acquired from Star Diagnostics MRI and CT Scan Centre Andhra Pradesh and the remaining
images are gathered from web resources. Datasets are gathered through June 2018–July 2019. It
contains females, males, and an infant. Few of the input images are given in Fig. 3.

4.3 Experimental results on proposed approach

The performance of the proposed brain tumor classification and segmentation is analyzed with
the help of sensitivity, specificity and accuracy which are most significant performance
parameters. The effectiveness of the proposed technique is demonstrated by performing a
comparison between the matching results of the proposed method with other approaches. The
result section is split into two phases such as classification phase and segmentation phase. We
firstly, verify the improvement of the classification phase. In classification phase we used the
recurrent neural network classifier to recognize the tumor part is present in the image or not. In
segmentation, we obtain the ROI and background region separately and also we measure the
accuracy of the proposed approach of MRG when other segmentation approaches.

Fig. 5 Performance of sensitivity plot based on classification stage

Multimedia Tools and Applications (2020) 79:14009–14030 14025

Fig. 6 Performance of specificity plot based on classification stage

4.3.1 Performance of classification phase

The basic idea of our research is to efficient MRI brain tumor detection and classification
based on Combining Wavelet statistical texture and recurrent neural network. In the classifi-
cation stage, at first we extract texture features from the image using combination of dominant
run length and co-occurrence texture features. After that, we reduce the features based on
sparse principle component analysis approach (OGSA). Then, the relevant features are given
to the recurrent neural network (RNN) classifier for classification. Our proposed classification
approach is compare to other well known classifiers such as SVM, KNN and NN. The
performance of the approach is shows in Fig. 4, 5 and 6.
The above Fig. 4 shows the performance of accuracy plot based on the classification stage,
as per the analysis, the accuracy gradually increases when the compared to the other
approaches. Here, in classification stage, we used recurrent neural network (RNN). When
analyzing Fig. 4, we obtain the maximum accuracy of 96% for using proposed RNN, 84% for
using KNN, 82% for using SVM and 85% for using NN. Moreover, KNN and SVM are
producing the almost same output value. The Fig. 5 shows the Performance of sensitivity plot
based on classification stage. Here, our proposed approach achieves the maximum sensitivity
of 100%. Similarly, Fig. 6 shows the performance of specificity plot based on classification
stage. Here, also we obtain the maximum output. In this classification stage, we obtain the

(a) (b) (c) (d)

Fig. 7 Experimental results, (a) MRI image with tumor, (b) gridded image, (c) seed point selected image, (d)
segmented MRI image
14026 Multimedia Tools and Applications (2020) 79:14009–14030

Fig. 8 Tumor region is extracted from the original brain MR image

maximum accuracy because of feature extraction based on WST and selection based on OGSA
methods. In Feature extraction, we obtained N number of features using combination of
dominant run length and co-occurrence texture. The large number of features affects the
classification accuracy. Therefore, we reduce the features OGSA. Using these two methods
are used to improve the accuracy of our proposed approach compare to other approaches.
Multimedia Tools and Applications (2020) 79:14009–14030 14027

Fig. 9 Performance analysis of segmentation process

4.3.2 Performance of segmentation phase

Segmentation is the important stage for tumor detection system. After classification stage, the
tumors images are given to the modified region growing algorithm (MRG). The explanation of
MRG algorithm is explained in section 3.5. The experimental results are illustrated in Fig. 7.
Classification stage provides the images which are tumour to the segmentation stage as
input. Segmentation stage of the proposed system quite accurately extract the tumour region
from these tumours brain images. Table 3 shows the results of segmentation phase. In all
images first brain portion is extracted and then tumour boundary is detected from this brain
portion. All images show that tumor region which is very high in contrast is quite accurately
identified and extracted by the proposed system.Figure 8
Figure 9 shows the performance of segmentation process. Here, we compare our proposed
MRG based ROI region segmentation with region growing (RG) based segmentation and k-
means based segmentation. Here, our proposed approach achieves the maximum accuracy of
95%, sensitivity of 97% and specificity of 98%. From the results, we clearly understand our
proposed approach attains the better results compare to other approaches.

4.3.3 Comparative analysis with state of art-methods

To prove the effectiveness of the proposed methodology, we compare our work with three
published works namely, Kumar and Vijayakumar [22], Preethi and palaniappan [28] and
Kavitha and Chellamuthu [19]. In [22], tumour image classification and segmentation was
proposed. For classification, they used kernel based support vector machine and segmentation
they used region growing algorithm. In [28], Combining Wavelet Texture Features and Deep
Neural Network for tumour Detection and Segmentation over MRI was proposed. Here, they

Table 4 Comparative analysis

Methods Accuracy (%) Sensitivity (%) Specificity (%)

Kumar and Vijayakumar [22] 88 84 90

Preethi and palaniappan [28] 92 86 91
Kavitha and Chellamuthu [19] 93 94 92
Proposed 96.26 100 92.85
14028 Multimedia Tools and Applications (2020) 79:14009–14030

used deep neural network for classification and PFCM for used for segmentation. Similarly, in
[19], self-adaptive particle swarm optimization (SLPSO) algorithm based tumour detection is
proposed. Here, for feature selection SLPSO algorithm is used and classification feed forward
neural network was used. The comparison based result is given in Table 4.
Table shows the comparative results of proposed against existing. When analyzing Table 3,
our proposed methodology attain the maximum accuracy of 96.26% which is 93% for using
Kavitha and Chellamuthu [19], 92% for using Preethi and palaniappan [28] and 88% for using
Kumar and Vijayakumar [22]. From the table we clearly understand, our proposed method
attain the better result compared to other methods.

5 Conclusion

The proposed system is developed for the diagnosis of brain tumour from MRI images of the
brain. This system makes the diagnosis in several phases. At First, texture features are extracted
from the noise free MR images. The dominant run length and co-occurrence texture features are
extracted at 0°, 45°, 90° and 135°. In this we select the suitable features based on OGSA
algorithm. These selected features are used for classification stage. In classification stage
proposed system used recurrent neural network (RNN) for classify image is tumour or not.
Once the images are determined as tumour these are further processed for tumour extraction
from them. Last stage is segmentation that extracts the tumour region using modified region
growing algorithm. The experimental results are explained our approach achieves the maxi-
mum accuracy of 96.26% which is high compared to existing approaches. Determine the type
of the tumour; size and thickness of the tumour region are some future directions of this work.

References

1. Acır N, Ozdamar O, Guzelis C (2006) Automatic classification of auditory brainstem responses using SVM-
based feature selection algorithm for threshold detection. Eng Appl Artif Intell 19:209–218
2. Almuallim H, Dietterich TG (1991) Learning with many irrelevant features, in proceedings of AAAI- 91,
pp.547–552
3. Bankman IH (2009) Hand book of medical image processing and analysis. Academic Press in Biomedical
Engineering
4. Bhattacharyya D, Kim TH (2011) Brain tumor detection using MRI image analysis. Communications in
Computer and Information Science 151:307–314
5. Biji CL, Selvathi D, Panicker A (2011) Tumor detection in brain magnetic resonance images using modified
thresholding techniques. Communications in Computer and Information Sicence 4:300–308
6. Blum AL, Langley P (1997) Selection of relevant features and examples in machine learning. Artif Intell 97:
245–271
7. Borden NM, Forseen SE (2011) Pattern Recognition Neuro radiology. Cambridge University Press, New
York
8. Cho S, Won H (2003) Machine learning in DNA microarray analysis for cancer classification. In
Proceedings of the First Asia-Pacic bioinformatics conference on Bioinformatics:189–198
9. Domingos P, Pazzani M (1997) On the optimality of the simple Bayesian classifier under zero-one loss.
Mach Learn 29:103–130
10. Dudoit S, Fridlyand J, Speed TP (2002) Comparison of discrimination methods for the classification of
tumors using gene expression data. Journal of the American Statistical Association 97(457)
11. Gibbs P, Buckley DL, Blackband SJ, Horsman A (Nov 1996) Tumour volume determination from MR
images by morphological segmentation. Phys Med Biol 41(11):2437–2446
12. Hamamci A, Kucuk N, Karaman K, Engin K, Unal G (Mar 2012) Tumor-cut: segmentation of brain tumors
on contrast enhanced MR images for radiosurgery applications. IEEE Trans Med Imaging 31(3):790–804
Multimedia Tools and Applications (2020) 79:14009–14030 14029

13. Hsieh TM, Liu YM, Liao CC, Xiao F, Chiang IJ, Wong JM (2011) Automatic segmentation of meningioma
from non-contrasted brain MRI integrating fuzzy clustering and region growing. BMC Medical Informatics
and Decision Making 11:54
14. Huang M, Yang W, Wu Y, Jiang J, Chen W, Feng Q (2014) Brain tumor segmentation based on local
independent projection-based classification. Biomedical Engineering, IEEE Transactions on 61(10)
15. Iscan Z, Dokur Z, Olmez T (2010) Tumor detection by using Zernike moments on segmented magnetic
resonance brain images. Expert Syst Appl 37(3):2540–2549
16. Jiang J, Wu Y, Huang M, Yang W, Chen WF, Feng QJ (2013) 3D brain tumor segmentation in multimodal
MR images based on learning population- and patient-specific feature sets. Computerized Medical Imaging
and Graphics 37(7–8):512–521
17. Jong K, Mary J, Cornuejols A, Marchiori E, Sebag M (2004) Ensemble feature ranking, In Proceedings Eur.
Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases
18. Karabatak M, Ince MC (2009) An expert system for detection of breast cancer based on association rules
and neural network. Expert Syst Appl 36:3465–3469
19. Kavitha AR, Chellamuthu C (2019) Brain tumour detection using self-adaptive learning PSO-based feature
selection algorithm in MRI images. International Journal of Business Intelligence and Data Mining 15(1)
20. Khan J, Wei JS, Ringner M et al (2001) Classification and diagnostic prediction of cancers using gene
expression profiling and artificial neural networks. Nature Medicine 7(673):679
21. Kharat KD, Kulkarni PP, Nagori MB (2012) Brain tumor classification using neural network based
methods. International Journal of Computer Science and Informatics 1(4)
22. Kumar P, Vijayakumar B (2015) Brain tumour MR image segmentation and classification using by PCA
and RBF kernel based support vector machine. Middle-East J Sci Res 23:2106–2116
23. Li D, Weinberg L, Pedersen (2001) Gene assessment and sample classification for gene expression data
using a genetic algorithm and k-nearest neighbor method. combinational chemistry and high throughput
screening 4(8):727–739
24. Mehmet Fatih A (2009) Support vector machines combined with feature selection for breast cancer
diagnosis. Expert Syst Appl 36:3240–3247
25. Nanthagopal AP, Sukanesh R (2013) Wavelet statistical texture features-based segmentation and classifica-
tion of brain computed tomography images. Image Processing, IET 7(1):25–32
26. Osareh A, Shadgar B (2011) A computer aided diagnosis system for breast cancer. IJCSI International
Journal of Computer Science Issues 8(2)
27. Prastawa M, Bullitt E, Ho S, Gerig G (Sep 2004) A brain tumor segmentation framework based on outlier
detection. Med Image Anal 8(3):275–283
28. Srinivasalu Preethi and Palaniappan Aishwarya, "combining wavelet texture features and deep neural
network for tumor detection and segmentation over MRI", journal of intellegent system, 2017
29. Quratul Ain M (August 2014) Arfan Jaffar and tae-Sun Choi, "fuzzy anisotropic diffusion based segmen-
tation and texture based ensemble classification of brain tumor". Volume 21:330–340
30. Ratan R, Sharma S, Sharma SK (2009) Brain tumor detection based on multi-parameter MRI image
analysis. International Journal on Graphics, Vision and Image Processing 9(3):9–11
31. Rohit M, Kabade S, Gaikwa (2013) Segmentation of brain tumour and its area calculation in brain MRI
images using K-mean clustering and fuzzy C-mean algorithm. International Journal Computer Science
Engineering Technology 4(5):524–531
32. Sachdeva J, Kumar V, Gupta I, Khandelwal N, Ahuja CK (2012) A novel content-based active contour
model for brain tumor segmentation. Magn Reson Imaging 30(5):694–715
33. Saha BN, Ray N, Greiner R, Murtha A, Zhang H (2012) Quick detection of brain tumors and edemas: a
bounding box method using symmetry. Comput Med Imaging Graph 36(2):95–107
34. Sapra P, Singh R, Khurana S (2013) Brain tumor detection using neural network. International Journal of
Science and Modern Engineering 1(9)
35. D. Sridhar, Murali Krishna (2013) “Brain Tumor Classification Using Discrete Cosine Transform and
Probabilistic Neural Network”, International Conference on Signal Processing, Image Processing and
Pattern Recognition
36. Sun Z-L, Zheng C-H, Gao Q-W, Zhang J, Zhan D-X (2012) Tumor classification using eigengene-based
classifier committee learning algorithm. IEEE Signal Processing Letters 19(8)
37. Taheri S, Ong SH, Chong VFH (2010) Level-set segmentation of brain tumors using a threshold-based
speed function. Image Vis Comput 28(1):26–37
38. Tan AC, Gilbert D (2003) Ensemble machine learning on gene expression data for cancer classification.
Appl Bioinforma 2:75–83
39. Valentini G, Muselli M, Ruffino F (2004) Cancer recognition with bagged ensembles of support vector
machines. Neuro computing 56:461–466
14030 Multimedia Tools and Applications (2020) 79:14009–14030

40. Wang ZJ, Hu QM, Loe K F, Aziz A, Nowinski WL (2004) Rapid and automatic detection of brain tumors in
MR images. in Proc. of SPIE
41. Zhang YL, Guo N, Du H, Li WH (2005) Automated defect recognition of C- SAM images in IC packaging
using Support Vector Machines. The International Journal of Advanced Manufacturing Technology 25:
1191–1196

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

Salma Begum Shaik obtained her Bachelor’s degree in Computer Science from Lakireddy Balireddy College of
Engineering, JNTUH.. Then she obtained her Master’s degree in Computer Science from Acharya Nagarjuna
University and doing PhD in computer Science majoring in image processing. Currently, she is Assistant
Professor at the Department of Computer Science and Engineering in Gudlavalleru Engineering College. Her
Specializations include image Processing, Machine Learning.

Dr D. Rajya Lakshmi , Vice –principal, Academics and Professor of Computer Science and Engineering in
JNTUK University College of Engineering, Narasaraopet, received her Doctorate from JNTU Hyderabad. She
presented many papers in International journals and conferences. She participated in many International
workshops. Areas of interest are Image processing and wireless sensor networks. She has a lot of experience
in teaching profession at various levels.