Multimedia Tools and Applications (2022) 81:8021–8057

https://doi.org/10.1007/s11042-022-12106-9

Hybrid deep neural network with adaptive rain

optimizer algorithm for multi-grade brain tumor
classification of MRI images

V. V. S. Sasank 1 & S. Venkateswarlu 1

Received: 12 April 2021 / Revised: 2 June 2021 / Accepted: 3 January 2022 /

Published online: 29 January 2022
# The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2022

Abstract
Classification of brain tumor is highly significant in the medical field in real-
world to improve the progress of treatments. The seriousness behind the tumors
are normally graded based on the size into grade I, grade II, grade III and grade
IV. This is where the process of multi-grade brain tumor classification gains
attention. Thus, the article focusses on classifying the brain MRI images into four
different grades by proposing a novel and a very efficient classification strategy
with high accuracy. The acquired images are pre-processed with the help of an
Extended Adaptive Wiener Filter (EAWF) and then segmented using the piece-
wise Fuzzy C-means Clustering (piFCM) technique. Then the most ideal features
such as the texture, intensity and shape features that can best explain the growth
of tumors are extracted using the Local Binary Pattern (LBP) and the Hybrid
Local Directional Pattern with Gabor Filter (HLDP-GF) techniques. After
extracting the ideal features, the Manta Ray Foraging Optimization (MRFO)
method has been introduced to optimally select the most relevant features.
Finally, a Hybrid Deep Neural Network with Adaptive Rain Optimizer Algorithm
(HDNN-AROA) is proposed to classify the grades of brain tumors with high
accuracy and efficiency. The proposed technique has been compared with the
existing state-of-the-art techniques relevant to brain tumor classification in terms
of accuracy, precision, recall and dice similarity coefficient to prove the overall
efficiency of the system.

Keywords Multi-grade brain tumor classification . Segmentation . Feature extraction . Feature

selection . Classification

* V. V. S. Sasank
[email protected]

1
Department of Computer Science and Engineering, Koneru Lakshmaiah Education Foundation,
Vaddeswaram, AP 522502, India
8022 Multimedia Tools and Applications (2022) 81:8021–8057

1 Introduction

Diseases associated with the brain are evolving as a big issue in modern society specifically
malignant brain tumors which seriously affect the human lives [9]. In brain tumor disease,
Gliomas are considered as highly occurring malignant brain tumor that is caused by the
abnormal transformation of brain cells also, it is basically classified into Low-Grade Gliomas
(LGG) and High-Grade Gliomas (HGG) [28]. Growth of HGG is highly worsened also
considered as malignant tumors and need surgery due to its complex nature. Whereas, LGG
has low influences in human life and the life time of the patient is extended over the treatment
[29]. These tumors are differentiated by various techniques like X-ray, Single-Photon Emis-
sion Computed Tomography (SPECT), Ultrasound, Magnetic Resonance Imaging (MRI),
Positron Emission Tomography (PET), Magnetic Brain Wave Graph (MEG), Computed
Tomography (CT), and Electroencephalogram (EEG) [1, 19, 27].
The exact size and volume of the malignant tumor is accurately identified by using MRI
scan and this MRI technique is considered as most widely used comprehensive method than
any other medical imaging techniques [14, 39]. In addition to that, historical screening is done
to perform grade labelling in images using MRI. Biopsy is a method which is required by this
invasive labelling technique. 3 Tesla (3 T) MRI is a non-invasive method developed to create a
robust in alternative for biopsy [10, 33]. Without a doubt, grade estimation in brain tumor is
performed by a MRI screening method that provides large amount of information to the
radiologists and neuro-surgeons. For instance, anatomical sequences can be related with
perfusion and diffusion imaging for an improved analysis. To advance molecular facts of
cancerous brain substances, MR Spectroscopy (MRS) can be used as additional technique.
This diversity of acquirements brands MRI a multimodal device for tumor oncology [2, 21].
With the significant efforts, automatic identification and segmentation of irregularities is
still a great challenge because of the inadequacies in image acquisition, scanner settings
variation, variations in the anatomy of brain morphology and pathological variations in
appearance [16, 22]. Early stage detection of brain tumor position, recognition and classifica-
tion is considered as a serious problem in medical science. When compared to the previous
conventional machine learning methods, deep learning methods have accuracy to detect the
tumor in brain MRI over the large amount of datasets [15, 36]. This kind of enhancement helps
the physicians to observe and track the occurrence and growth of tumor-affected areas at
various stages so that, they can offer proper analysis with these image scanning [12, 18, 34].
Overall, key issue in the multi-grade tumor identification is early stage detection which
helps to treat people with proper treatment. Treatments like radiation, therapy, surgery or
chemotherapy can be initiated by the physicians to cure the disease. When the tumor is
detected in its early stage, it is obvious that the survival chances of tumor-infected patient
can be increased significantly. In recent years, there has been a proliferation of machine and
especially deep learning techniques in the medical imaging field [19–21]. Deep learning
algorithms like deep neural networks [13, 35], or the convolutional neural networks [11] are
built using large stacks of individual artificial neurons, each of which performs primitive
mathematical operations of multiplication, summation, and thresholding. One of the key
reasons for the success of these modern deep neural networks is the idea of representation
learning; the process of learning useful features automatically from the data as opposite to
manual selection by expert humans [14, 20, 37].
Multi-grade tumor classification is highly significant to determine the growth and stage of
the tumor inside the brain. To reduce the mortality rate due to brain tumors and to improve the
Multimedia Tools and Applications (2022) 81:8021–8057 8023

quality of life for every patients, an efficient and highly accurate classification strategy is much
required. A descent classification strategy with reduced error can easily identify the rate of
growth of the tumor as well as helps in detailed analysis of the tumors paving way for efficient
remedies. There are a vast count of techniques available to categorize the tumors identified as
harmful and harmless whereas the technological side still lags in obtaining an efficient and
accurate classification results based on the stages (grades) of the tumors. The immediate
requirement of a powerful classification strategy and the lag in the research side in accurate
identification of the grades of tumors motivated to propose this work. In this work, the major
texture, intensity and shape features that can distinguish the different sizes of the tumors are
extracted using Uniform Local Binary Pattern (ULBP) and Hybrid Local Directional Pattern
with Gabor Filter (HLDP-GF) methods. The hybrid features are chosen as the Local Direc-
tional Pattern (LDP) is highly discriminative and can well-define the tumor shape and size.
The major objective of the proposed work is to accurately classify the brain tumors based
on the grades. For efficient classification, the deep learning (DL) is highly significant as it
learns the high-level features from the input images and performs classification effectively
based on the image features. The proposed article uses the Deep Neural Network (DNN) for
performing the classification task. In general, the results of the DNN classifier has an influence
of the random weights chosen while training the network. Hence it is necessary to optimize the
weight values in order to improve the classification performance. Therefore, the proposed
model uses an efficient optimization algorithm called as the Rain Optimization Algorithm
(ROA) to fine-tune the random weights in the hidden layers of DNN. This algorithm is a
recently introduced meta-heuristic algorithm applicable for any engineering applications with
reduced complexity. The major motivation of choosing this algorithm is that it is simple to
implement and has the ability to converge faster than many other meta-heuristics providing
reduced time and computational complexities. Also, to the best of our knowledge, this
algorithm has not been implemented with any other classifier which makes our proposed
work novel and unique. The efficiency of this algorithm in terms of accuracy, stability and
convergence has been proved through simulations in [23]. This combined hybrid classification
model is utilized to label the MRI tumor images into different grades based on the input
features and the obtained knowledge of the brain tumors through the analysis of the training
samples. The major contributions of the proposed work include the following:

& A hybrid Local Directional Pattern with Gabor filter (HLDP-GF) along with uniform
Local Binary Pattern (ULBP) has been utilized for extracting the ideal features related to
texture, intensity and shape of the images that can best describe the tumor size.
& An efficient Manta Ray Foraging Optimization (MRFO) is introduced for selecting the
most relevant features with the elimination of high dimensionalities to improve the training
process and to reduce the time complexity.
& Constructing a novel and unique Hybrid Deep Neural Network with Adaptive Rain
Optimizer Algorithm (HDNN-AROA) to classify the grades of brain tumors with better
accuracy and convergence.
& The performance of the proposed approach has been evaluated in terms of accuracy,
precision, recall, and dice similarity to prove its efficiency against other existing state-of-
the-art techniques.

The remaining sections of the paper are organized as follows: Sections 2 provides the survey of
recently published papers relevant to brain tumor classification. Section 3 presents the
8024 Multimedia Tools and Applications (2022) 81:8021–8057

proposed methodologies. Section 4 demonstrates the simulation part along with results and
discussion. Section 5 provides the conclusion and future scopes of the proposed article.

2 Literature review

Brain tumor causes significant damage to health and leads to death due to its severe nervous
system disorder. For the treatment and analysis of glioma, MRI becomes a most widely used
principle diagnostic system. In brain tumor MRI, segmentation and classification process is
considered as a complex task. So, C. Narmatha et al. [26] developed a fuzzy brain-storm
optimization method for brain tumor segmentation and classification. Cluster center calcula-
tion was determined by brain-storm optimization algorithm and it provides them the highest
priority. Optimal network structure was obtained by performing several iteration of fuzzy
algorithm. In terms of various results, brain storm algorithm outperforms existing methods.
Nowadays, automatic diagnosis of brain tumor is contemplated as an important research
topic. Due to the various characteristics of tumors, it is very difficult to classify the different
grades of the tumors. Neurologists or neurosurgeons are assisted by the recent deep neural
networks which normally classify the tumors in an efficient manner. But, deep learning
algorithms are affected by overfitting and vanishing gradient problem. In this research [17],
R. Lokesh Kumar et al. suggested a deep network model named as ResNet-50 which solved
the overfitting and vanishing gradient problems. A three-tumor brain MRI dataset consisting of
3064 images were utilized to estimate the efficiency of the proposed method.
Wrong classification or misclassification of tumors can lead to bad consequences in people
lives. For correct treatment, it is very essential to recognize the correct category and grades of
tumors in early stage. In addition to that, manual detection of tumor grade is more time
consuming and may lead to human error, thanks to the huge amount of data as well as several
kind of brain tumor types. So, it requires automatic detection using computer assisted
diagnosis (CAD). In this research [3], Wadhah Ayadi et al. deep convolution neural network
(CNN) is developed to classify the tumor in an accurate manner. This deep CNN model is
evaluated against three different kind of datasets.
When it comes to classification of tumors, analyzing the tumor types gains advantage.
Benign and malignant are two different types of brain tumor in which both types have different
properties. Henceforth, it is very essential to segment and classify two different kind of tumors
from MRI so that, doctors can easily provide a correct treatment to the patient. In this research
[31], V.V. Sasank and S. Venkateswarlu developed a machine learning based classification in
brain MRI. In addition to that, pre-processing was done with the help of Contrast Limited
Adaptive Histogram Equalization (CLAHE) as well as Laplacian of Gaussian (LoG) methods.
After that feature extraction was performed and optimization based method was applied to
select the correct features from huge number of features. Finally, classification of brain tumor
was performed by proposed Kernel based Softplus Extreme Learning Machine (KSELM)
technique. Different types of dataset like BRATS 2014, 2015, 2018 and BRT (Brain Tumor)
were taken to evaluate the proposed methodology.
Classification of brain tumor grade is considered as an essential process to assist the neuro-
surgeons during clinical analysis. Hiba Mzoughi et al. [25] developed a deep multi-scale 3D
CNN for MRI gliomas brain tumor classification. This algorithm efficiently classified the
glioma brain tumor into high grade gliomas (HCG) and low grade gliomas (LCG). The
proposed method had the possibility to join both global as well as local contextual data with
Multimedia Tools and Applications (2022) 81:8021–8057 8025

reduced weights over the small kernels in 3D convolutional layer and a deep network. In
addition to that, intensity normalization and adaptive contrast enhancement techniques were
applied to perform the pre-processing stage in proposed methodology. BRATS 2018 dataset
was utilized to evaluate the proposed methodology in which it creates discriminative feature
map to differentiate the LCG and HCG variants.
Among the different cancer classification techniques, microarray technology is highly
significant as it is capable of simultaneously analyzing huge count of genes and samples.
An approach had been introduced by Santos Kumar Baliarsingh and Swati Vipsita [4] for the
microarray cancer classification. The approach hybridized the Chaotic Emperor Penguin
Optimization (CEPO) algorithm with the Extreme Learning Machine (ELM) for efficient
classification. Initially, the relevant genes were selected using the Fisher score and ReliefF
filters. Then, the CEPO algorithm was used to train the ELM model with the aid of optimizing
the hidden weight and bias values. With the use of the selected genes, the classification was
performed efficiently. The performance of the approach was evaluated under different exper-
iments over seven binary-class and multi-class datasets. The results demonstrated the efficien-
cy of the approach over other compared approaches in terms of accuracy due to the
hybridization utilized.
Another similar approach using the MapReduce had been presented by Santos Kumar
Baliarsingh [5] with the analysis of high-dimensional genomic data. The relevant genes were
initially selected using the MR-based Fisher Score (mrFScore) and MR-based ReliefF
(mrReliefF) algorithms. Then, based on the selected genes, the microarray classification had
been performed using the MR-based probabilistic neural network (mrPNN) with weighted
chaotic grey wolf optimization (WCGWO). The approach had been simulated under the
Hadoop framework with four different microarray datasets to identify the efficiency of the
feature section and classification techniques. The results of the experiments suggested that the
approach performed well in classifying the genes even with large datasets. The WCGWO
algorithm worked efficiently in optimizing the neural network and improved the overall
stability.
To deal with the high-dimensional biomedical data, a hybrid approach had been suggested
by Santos Kumar Baliarsingh [6] named as SARA hybridizing Simulated Annealing (SA) with
the Rao Algorithm (RA). The approach was formulated with SA working for the local search
and RA for global optimization where the RA’s exploitation capability was improved through
hybridization. The framework included two stages where the first stage was meant for gene
selection and the second stage for classification. The relevant gene subsets were selected from
the microarray dataset using the minimum redundancy maximum relevance (mRMR) method.
The efficiency of the framework was proved through experiments using five different binary-
class and multi-class datasets. The accuracy of the model was improved with the help of the
utilization of RA optimization into the training algorithm.
From the surveys taken, it has been identified that the available strategies for brain tumor
classification is not so accurate and efficient. Most of the reviewed methods took a very long
time to train and produce classification results. Moreover, the learning time of the classifiers
are too high when the classification is performed for multiple images with large quantity of
features. Though some of the surveyed techniques are seen to be efficient in providing better
classification results, those techniques are computationally complex and are impractical to be
implemented in real-world. To compromise the immediate requirement of an accurate and
efficient classification strategy, the article has been proposed with the most efficient techniques
known so far. Selection of optimization techniques influences the results of the proposed
8026 Multimedia Tools and Applications (2022) 81:8021–8057

approach and hence the most efficient algorithms with high convergence rates are chosen for
generating accurate classification results. The use of AROA algorithm efficiently reduced the
computational complexity of the learning process with minimized error. The main aim of the
proposed article is to improve the classification accuracy along with reduced time and
computational complexities.

3 Proposed methodology

The article proposes a hybrid framework for multi-grade brain tumor classification to classify
the grades of brain tumors accurately. The brain tumor images are initially acquired, and the
pre-processing steps are introduced to remove the noise present in it. The EAWF filter has
been utilized with the aid of removing the unwanted noise present in the images as well as to
improve the quality of the images. After the pre-processing step, the images are subjected to
segmentation, where the piFCM technique has been employed to efficiently segment the
images based on the image features applicable in the domain of real-world applications. The
main aim in using the piFCM clustering technique is to reduce the uncertainty arising while
clustering the images employing a powerful uncertainty absorbent.
Succeeding the clustering process, feature extraction is carried out on the segmented images
using ULBP, LDP and Gabor filter. In some cases, the extracted features might cover both
relevant and irrelevant features in large dimensions. Thus an efficient feature selection method
has been introduced to identify the required and sufficient features for classification. The
proposed technique presents an efficient optimization algorithm called MRFO for optimal
feature selection from the extracted set of features. The selected features are given as input to
the classification phase where the tumors get classified into different grades. For classification,
an efficient HDNN-AROA has been proposed to provide better prediction results with high
accuracy. Figure 1 depicts the proposed framework for multi-grade brain tumor classification.

3.1 Pre-processing

Pre-processing is the initial step followed in the proposed work to improve the image quality
by making it sharper and to remove the undesirable noise present in the image. The images
acquired from the databases like Radiopaedia dataset and brain tumor dataset consists of MRI
brain tumor images that are labelled into four grades based on the tumor’s intensity and
growth. Usually, the images of MRI consists of noise owing to various factors like sensors,
temperature, lighting levels, granularity, etc. These images are pre-processed using the EAWF
[30] filter to improve the clarity of the images by removing the noise while preserving the edge
information of the images. To improve the original adaptive wiener filter, a dispersion index
has been introduced in place of variance. This helps in identifying the clustered and dispersed
occurrences.
Normally the pre-processing strategy eliminates the noise present in the image. The use of
EAWF filter reduces any kind of noise identified in the image along with improving the image
quality. Considering a pixel location p1, p2 of an input image Ip, the normal AWF determines
the noise using the mean and noise variance. The extended technique uses a dispersion index
to identify the noise approximately from the image. The additive noises present in the
image are retarded to smoothen the image with high pixel quality. Dispersion index
can be identified as:
Multimedia Tools and Applications (2022) 81:8021–8057 8027

Dataset

Pre-
Processing

Feature Extraction Optimal Features

Intensity Texture Shape Feature 1 … Feature n

Segmentation

.. Classification

HGT LGT

Fig. 1 Proposed Framework for Multi-Grade Brain Tumor Classification

σ2
di ¼ ð1Þ
μ
Applying the dispersion index onto the normal AWF filter, the equation gets modified like the
following:

 di  σ2n

W f I p ðp1 ; p2 Þ ¼ μ þ I p ðp1 ; p2 Þ  μ ð2Þ
di
Upon simplification, the overall output after pre-processing can be identified through the
following formula:
 2 

 σn

W f I p ðp1 ; p2 Þ ¼ I p ðp1 ; p2 Þ  μ 2 I p ðp1 ; p2 Þ  μ ð3Þ
σ
where, μindicates the mean value and σn indicates the variance of noise.
From Eq. (3), the enhanced quality of the input image can be identified. The noise present
in the image can be efficiently removed using the improved filter. The noise variance indicates
the noise present in the image and the value obtained for the noise variance demonstrates the
influence it caused to the image. Reducing the noise improves the image quality and provides
an image that is more clear and visible than the original image.
8028 Multimedia Tools and Applications (2022) 81:8021–8057

The major benefit of using this improved filter is that the speckle noise present in the image
will be keenly removed. Speckle noise is seen mostly in the MRI images and it occurs possibly
due to environmental conditions. This noise has a direct impact on the contrast and texture of
the images as it gives a grainy appearance and hence removing this noise is highly significant.
AWF filter is popularly known to remove the additive noise identified in the images and the
improved AWF filter removes the speckle noise more efficiently, meanwhile improving the
image quality.

3.2 Image segmentation

Succeeding the pre-processing step, the segmentation step through clustering has been
carried out. The pre-processed MRI images obtained are subjected to segmentation
based on the piecewise Fuzzy C-means Clustering technique (piFCM). This technique
has been employed for clustering as the data objects that are to be segmented might
be connected with more than one clusters with diverse memberships, and this tech-
nique can efficiently cluster such objects exhibiting relationships between clusters.
Segmentation prevails as a major task as it clusters the images into groups based on
the features which makes the extraction process simpler.
Normal fuzzy clustering clusters the images based on the membership between the image
clusters. The piecewise image centroid are identified to compute the segmented parts using the
piFCM. The identified centroid in the object determines the shortest distance between the
segmented images to identify the membership among various parts of the object. In this way,
the images are segmented identifying the relationships between them.
Initially, to cluster the images, the data points in the images are considered as a
fuzzy set determining the relationships between each other. The fuzzy clustering
technique is followed by almost all the segmentation techniques in image processing
systems but the difficulty lies in identifying the data objects and cluster centroids
from the distance computation. The distance formulation of the standard fuzzy clus-
P  
tering can be given as rc¼1 f μck ; qcl . The formula denotes the distance between pk
and ql where c denotes the count of pieces that are segmented. From Eq. (4), the
centroids and the data objects are difficult to identify and hence piecewise FCM is
adopted to overcome this issue with the introduction of multi-membership function.
Let the MRI brain image dataset be represented as D = {d1, d2, .…dn} which is a
multi-membership dataset imitated from a Bootstrap Probability (BP) set and the
kthdata object in the dataset can be represented as:
 
d k ¼ μ1k ; ::μik ; ::μrk ð4Þ

The data object identified through Eq. (5) aligns the membership degree of the identified object
Pr 
to all the clusters obtained from the available BPs. Thus the vector dk is a c¼1 Lc
dimensional vector where‖dk‖1 = r ∀ k. The centroid for the clustering can be computed
using the piFCM technique and can be represented as:
 
ql ¼ q1l ; :qcl ; :qrl ð5Þ

The centroid that are identified through Eq. (6) is also a vector and the vector dimension for ql
Pr 
can be identified as c¼1 Lc casing r pieces. Every centroid computed through Eq. (6) is
Multimedia Tools and Applications (2022) 81:8021–8057 8029

known to be the piecewise centroid as the centroid is computed for the pieces segmented from
the image. Thus qlcan be termed as the piecewise centroid for the dataset D for fuzzy
clustering. Considering the fuzzy clustering in terms of multi-membership data, the distance
between the centroid and the data object can be given as:
X
r  
disðd k ; ql Þ ¼ f μck ; qcl 8k; l ð6Þ
c¼1

The distance computed is the Euclidean distance between the centroid and the respective data
object and the minimization of the distance between them will provide clusters that are more
relevant.
The brain MRI images are mostly uncertain, and reducing the uncertainty of the
images are essential while passing on values to the membership functions. Uncer-
tainties reduce the efficiency of clustering and the relevance between the data objects
cannot be identified accurately due to the presence of various kinds of uncertainties.
Some types are present in the input image itself and many others might get added
later. To obtain an efficient clustering output and to improve the classification with
high accuracy, the uncertainties are required to be reduced using a suitable uncertainty
absorbent. With the use of these absorbents, the uncertainty present in every image
can be reduced and the clustering becomes more relevant and efficient. To eliminate
the uncertainty present in the image, a parameter to reduce the uncertainty can be
added to every pixel values subjected to the membership function. This parameter can
be termed as the hesitation degree denoted as follows:
f ðdÞ ¼ 1  μf ðdÞ  qf ðdÞ ð7Þ

In the above equation, μf(d) and qf(d) symbolizes the multi-membership function and the
centroid for the data objects respectively. Therefore, the piFCM clustering function along with
the uncertainty absorbent can be given as follows:
n o
F ¼ μf ðdÞ; qf ðdÞ; f ðdÞ 8d 2 D ð8Þ

From the above Eq. (8) the clusters that are segmented can be identified along with
the negation of uncertainty. The piFCM clustering mechanism enhances the segmen-
tation and the ideal clusters are obtained without uncertainties through the use of the
uncertainty absorbent.

3.3 Feature extraction

Feature extraction is one of the most significant functionality followed in the proposed
article for brain tumor classification. This process indicates the collection of most
required information from the images to get a deep idea about the image and to
obtain an accurate output in classification. Any kind of data objects are identified or
get classified only based on the features associated with the data objects and the
learning techniques highly relies on the features of the image to classify more
efficiently. Extraction of relevant and most important features related to shape and
texture paves way for efficient classification of the brain images. The clustered images
from the segmentation process are subjected to the feature extraction phase where the
8030 Multimedia Tools and Applications (2022) 81:8021–8057

most important features are identified based on the clusters itself. The article uses a
uniform local binary pattern (ULBP) and hybrid local directional pattern with gabor
filter (HLDP-GF) to extract the required features. The features like texture, shape and
intensity signifies the image with keen contrast. Extracting these features enable the
identification of tumor based on the grades more accurately.
The LBP features related to texture are initially extracted and then the hybrid LDP with
Gabor filter is introduced to extract the left out features that constitutes the intensity and shape
features of the image. Moreover, the edge information of the images are preserved through
ULBP and if the features present in the edge of the images are found significant, those are
extracted. Then, the intensity and shape features are extracted with the use of the hybridized
feature extraction method to reduce the complexity as well as to improve the efficiency.
Intensity features like mean, variance, standard deviation, skewness, kurtosis, etc., constitute
the major part of the image and identifying and extracting these features will improve the
classification accuracy with best results. Also, the most important shape features that can
discriminate different tumor grades like shape index, irregularity, perimeter, area and circular-
ity are extracted with the use of HLDP-GF.

3.3.1 ULBP based feature extraction

This phase constitutes the major part in extracting the texture based features. The ULBP
feature offers robustness, it is computationally simple and can be practically implemented.
This is highly significant and the major features are extracted on the basis of a pattern. Initially,
the histogram is constructed for the image and then the features are identified through the
circular domains used in the histogram. The ULBP pattern associates almost all the required
texture patterns of the image and the information identified in the edges, end points, isolated
points, corners, etc. are extracted more efficiently. This method is highly robust and ensures
efficient scaling of the images with simplicity. Using ULBP, the uniform patterns are
recognized through merging of 2luniform patterns with (L + 1)uniform and non-uniform
patterns where the spatial transitions will not be greater than 2 and the calculation for ULBP
can be identified using the following equation:
 PL1
U LBP ðl; mÞ ¼ l¼0 pðbl  bc Þ; if U  2 ð9Þ
L þ 1; otherwise

The computation of the ULBP can be formulated as follows:

X
l1
U ¼ j pðbL1  bc Þ  pðbo  bc Þj þ j pðbl  bc Þ  pðbl1  bc Þj ð10Þ
l¼1


1; iffx  0
pðxÞ ¼ ð11Þ
0; otherwise

In the above equation, ULBP(l, m) represents the ULBP computation of a circular domain with l
pixels of radiusm.bl(l = 1, 2, .. …l − 1) represents the pixel values identified in the
neighborhood of the circular domain and bc is the central gray value of the pixel in the circular
domain.
Multimedia Tools and Applications (2022) 81:8021–8057 8031

The ULBP operator is more efficient than the standard LBP operator due to its simple
computation and due to its capability in identifying only the required information thus
reducing the storage requirements.
Along with the ULBP features, some of the common and most important texture features
[24, 32] are extracted such as energy, correlation, contrast, dissimilarity and homogeneity are
indicated as the following:
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u a1 d1
uX X 2
Energyen ¼ t f ð j; k Þ ð12Þ
j¼0 k¼0

Pa1 Pd1
j¼0 k¼0 ð j; k Þf ð j; k Þ  M jM k
Correlation cr ¼ ð13Þ
σx σy

a1 X
X d1
Contrastct ¼ ð j  k Þ2 f ð j; k Þ ð14Þ
j¼0 k¼0

a1 X
X d 1
Dissimilarityd s ¼ ð j j  k jf ð j; k ÞÞ ð15Þ
j¼0 k¼0

!
a1 X
X d 1
1
Homogeneityhm ¼ f ð j; k Þ ð16Þ
j¼0 k¼0 1 þ ð j  k Þ2

Thus the texture feature ULBP together with other extracted texture features can be indicates
as a vectorF1 as follows:
F 1 ¼ ½U LBP ; ðen Þ; ðcr Þ; ðct Þ; ðd s Þðhm Þ ð17Þ

3.3.2 HLDP-GF based feature extraction

The LDP is used to identify the intensity features present in the input MRI image(h,
v) through the use of the Kirsch masks that can recognize the pixel values from all
the eight directions. The major use of this feature is to make a tradeoff between the
performance and the computational time. This feature shows stronger discrimination
power in extracting the texture based features. Moreover, this feature is highly robust
to noisy images than many other feature descriptors and are invariant to image
rotations. Initially, a binary code holding eight bits is assigned to every pixel of the
input image. The Kirsch mask is applied in eight different orientations to obtain the
information present in the edges. This technique highly preserves the edge information
and the relevant intensity features identified in the edges of the images are extracted.
The LDP feature for the image segment can be identified as follows:
8032 Multimedia Tools and Applications (2022) 81:8021–8057

X
7 
1; iffx  0
LDP ðh; vÞ ¼ f ðK m  K A Þ2 ; f ðxÞ ¼
m
ð18Þ
m¼0
0; otherwise

From the above equation, (h, v)indicates the input segmented image subjected to feature
extraction, Km indicates the Kirsch mask and KA indicates the Kirsch activation function.
The required LDP codes are calculated for all the available pixels of the input images and
then the input image (I) can be represented using the histogram as:
XM X N 
1; x ¼ I
H g ðIÞ ¼ ðLDP ðh; vÞ; I Þ; f ð x; I Þ ¼ ð19Þ
h¼1 v¼1
0; otherwise

Along with the LDP features, some of the other important intensity features are extracted such
as mean, standard deviation, skewness and kurtosis which are formulated as:
  a1 X
1 X d 1
MeanM ¼ f ð j; k Þ ð20Þ
a*d j¼0 k¼0

Standard Deviation
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u
u 1 X a1 X d 1
SD ¼ t ð f ð j; k Þ  M Þ2 ð21Þ
ða*dÞ j¼0 k¼0

Skewness
  Pa1 Pd1
1 j¼0 k¼0 ð f ð j; k Þ  M Þ3
SK ¼ ð22Þ
a*d SD 3
Kurtosis
  Pa1 Pd1
1 j¼0 k¼0 ð f ð j; k Þ  M Þ4
KR ¼ ð23Þ
a*d SD 4
All the above mentioned features are some examples of the intensity features considered in this
article. The gabor filter is introduced in combination with the LDP to enhance the feature
extraction process. The features that are required to be extracted using the gabor filter are the
shape features such as area, shape index, circularity, perimeter, irregularity, etc. Gabor filter is
chosen as it has the ability to represent tumor shapes over multiple scales, orientations and
locations. But extracting these features only through gabor filter holds a huge amount of high
dimensional and redundant information. Also, the gabor filter is not efficient in discriminating
the features and is computationally complex. It requires more filters to be installed to extract
the features from the pixels, thus increasing the processing complexity. Thus the gabor filter is
combined with the LDP to obtain the shape features along with the intensity features.
Initially, the gabor filter is applied to the segmented MRI images to obtain the
feature graphs and then the histograms for the pixels are constructed through the use
of LDP operator mentioned in (18). The SFG (Superimposed Feature Graph) of the
gabor filter defines the gabor features relevant to the shape and the graphs are
constructed with the help of the identified shape features. After extraction, the
Multimedia Tools and Applications (2022) 81:8021–8057 8033

histograms are defined only for the identified pixels through the SFG where the above
mentioned intensity features are extracted. The Gabor filter for a pixel can be
represented as follows:

w; kw; k2
2
kzk2 h i
eðiw;uz Þ  e 2
σ2
G; ðh; vÞ ¼ e 2σ ð24Þ
σ2
In the above equation, αdefines the orientation and βdefines the scale of gabor filter, σdefines the
radius of the envelope determined by Gaussian function, z = (h, v) andwα, β defines the wave
vector. The convolution operation to represent the gabor filter for the input image can be given as:
C ; ðh; vÞ ¼ I ðh; vÞ*G; ðh; vÞ ð25Þ

The magnitude value for the gabor filter is computed that corresponds to different response
values for every pixel position of the image. The magnitude value corresponding to the gabor
filter can be identified using the following equation:
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
 2  2ffi
M; ðh; vÞ ¼ Re C ; ðh; vÞ þ I m C ; ðh; vÞ ð26Þ

The other major shape features [32] of the image along with area and perimeter extracted are as
follows:
l10 l01
Centroidcr ¼ ðx1 ; y1 Þ; x1 ¼ ;y ¼ ð27Þ
l00 1 l00
Length of major axis
2  qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 3 12
2  20 þ  02 þ ð 20   02 Þ2 þ 4 211
6 7
Q ¼ 26
4
7
5 ð28Þ
00

Length of minor axis

2  qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 3 12
2  20 þ  02  ð 20   02 Þ2 þ 4 211
6 7
R ¼ 26
4
7
5 ð29Þ
00

Central Moments
XX
 pq ¼ ðh  x1 Þp ðv  y1 Þq f ðh; vÞ ð30Þ
h v

Image Moments
XX
impq ¼ hp vq f ðh; vÞ ð31Þ
h v

The above features are the features corresponding to the shape of the image. The LDP operator
works on finding the histogram for the feature graphs constructed by the Gabor filter to extract
the shape features. The output generated from the histogram will be the required intensity
8034 Multimedia Tools and Applications (2022) 81:8021–8057

features. The hybrid form representation to extract the shape and intensity features can be
given as follows:
X
7
LDP G ¼ f ðGm  GÞ2m ð32Þ
m¼0

After the feature graph construction, the histogram operator works on building histograms on
the graphs to identify the edge features and to extract them. The histogram operator to
construct the histogram on pixels can be modified as:
M X
X N X
7
H g ðIÞ ¼ ½ð f ðGm  GÞ2m Þ; I ðh; vÞ ð33Þ
h¼1 v¼1 m¼0

Thus the features extracted through the HLDP-GF along with other intensity and shape
features can be represented as a vector as follows:
h  i
F 2 ¼ LDP G ; ðM Þ; ðS D Þ; ðS K Þ; ðK R Þ; ðcr Þ; ðQÞ; ðRÞ;  pq ; impq ð34Þ

This hybridized form of feature vector is highly capable of extracting the most discriminative
intensity and shape features from the images with reduced computational complexity. There-
fore, the overall features that are extracted from the MRI brain image can be obtained by
adding the feature vectors in (16) and (34). This is described as:
h n   oi
F ¼ fðU LBP Þðen Þ; ðcr Þ; ðct Þ; ðd s Þðhm Þg; LDP G ; ðM Þ; ðS D Þ; ðS k Þ; ðK R Þ; ðcr Þ; ðQÞ; ðRÞ;  pq ; impq ð35Þ

3.4 Feature selection using MRFO

Feature selection is the prime segment in the proposed brain tumor multi-grade
classification approach. The feature selection process indicates the selection of best
features from the identified set of features. In this article, the MRFO algorithm [38]
has been introduced for the application of feature selection to optimally select the
most important features. This is a bio-inspired algorithm mimicking the foraging
process of manta rays. This algorithm is highly efficient in both exploration and
exploitation to identify the global optimal solution which makes it highly suitable for
selecting the optimal features. Manta Rays are marine creatures popularly known to be
one of the largest known sea animal. These creatures follow a very efficient foraging
strategy to identify and eat the sea planktons found in the ocean. The three unique
and intelligent foraging strategies of manta rays can be identified as follows:

Chain foraging In this foraging strategy, group of individuals move towards the food
following a lead. Also, the individuals find the food that are missed by the lead and hence
best food are also identified by every individual in the group.

Cyclone foraging The group of individuals in this foraging strategy make a spiral movement
towards the patch of food recognized. Along with following the lead in a spiral pattern, each
individual in the group forages separately along the path.
Multimedia Tools and Applications (2022) 81:8021–8057 8035

Somersault foraging The angle of the identified food is represented as a central point and
every individual of the group somersaults to reach the food and thus the positions of each
individual are updated frequently.
The MRFO can be imitated into the proposed technique for feature selection by
taking the features extracted through LBP and HLDP-GF as the individuals in the
search space. The major aim of the feature selection phase is to optimally select the
features with high discriminative nature that can define the tumor size more effec-
tively. The optimal features in the group are selected based on the fitness function
that signifies the tumor based on the grade. Initially, the group of features (n)
extracted are represented as the individuals in the D-dimensional search space in
matrix formulation. It can be described as follows:
2 3
f 1;1 f 1;2 :f 1;d
6 f 2;1 f 2;2 f 2;d 7
6 7
FS ¼ 6
6 ::: 7
7 ð36Þ
4 ::: 5
f n;1 f n;2 f n;d

The random placement of features signifies that there is no experience in the fitness
evaluation and the optimal features are unknown. Therefore, an objective function is
formulated to evaluate the initial placement. For this purpose, the minimization of the
features is enclosed to shorten the amount of features and to discard the features with
greater dimensions. The error in classification of the features into optimal and worst
influences the feature selection process and hence it is embedded in the fitness
evaluation function as follows:
jfj
Fit ¼ Ω*Ψ þ ð1  ΩÞ* ð37Þ
jF j

In the fitness evaluation, f resembles the count of features extracted in the feature
extraction process, Fresembles the total count of features identified in the dataset,
Ψsymbolizes the classification error and Ωresembles the parameter where (Ω ∈ {0,
1}) chosen to indicate the influence of the classification error over the extracted
feature set. Then the movement of the features towards the optimal features in
imitated based on the foraging strategy of manta rays.
The three different foraging strategies can be imitated to identify the best features
from the three different feature set. The optimal features among the intensity, texture
and shape feature sets are identified based on the chain, cyclone and somersault
foraging strategies respectively. In the chain foraging, the optimal intensity feature
is identified as the foraging point and the other secondary optimal intensity (LDP)
features move towards it to form an optimal intensity feature subset. The position of
every individual feature is identified and updated based on the following chain
position update equation:
8
< xdj ðtÞ þ ra xdb ðtÞ  xdj ðtÞ þ ! xdb ðtÞ  xdj ðtÞ j ¼ 1
xdj ðt þ 1Þ ¼ ð38Þ
: xd ðtÞ þ ra xd ðtÞ  xd ðtÞ þ ! xd ðtÞ  xd ðtÞ j ¼ 1; 2; :n
j j1 j b j
8036 Multimedia Tools and Applications (2022) 81:8021–8057

pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
! ¼ 2*ra jlogðra Þj ð39Þ

The Eq. (39) signifies the movement of intensity features towards the optimal features where
xdj ðtÞindicates the position of jth individual in the dth space at time t, ωindicates the weight
factor, ra is a random vector chosen between 0 and 1, xdb ðtÞis the best position identified by the
foraging features where it is close to optimal one. The most optimal intensity features are
identified from the chain foraging strategy.
Then the cyclone foraging strategy is imitated in this article to identify the optimal
texture based feature (ULBP) set from the extracted texture features. The optimal
texture feature that best defines the texture of the tumor is chosen the foraging point
and the other texture features are set as the individuals that travels towards the
optimum. The position of every foraging feature in the group is updated based on
the following position update equation:
8
< xdb þ ra xdb ðtÞ  xdj ðtÞ þ " xdb ðtÞ  xdj ðtÞ j ¼ 1
xj ðt þ 1Þ ¼
d
ð40Þ
: xd þ ra xd ðtÞ  xd ðtÞ þ " xd ðtÞ  xd ðtÞ j ¼ 1; 2; ::n
b j1 j b j

Ttþ1
" ¼ 2era 1 T sinð2ra 1Þ ð41Þ

The above Eq. (42) signifies the weight factor ε with ra1 indicating the random number
between 0 and 1 and T indicating the maximum count of iterations. The cyclonic foraging
strategy has the ability to find the global optimum among the features recognized.
Finally, the most unique and efficient somersault foraging strategy is imitated to
identify the optimal shape feature (Gabor) set among the extracted shape features. The
most optimal shape feature is identified more efficiently through this strategy and the
position update takes place more frequently for every foraging feature which is
depicted as follows:

xdj ðt þ 1Þ ¼ xdj ðtÞ þ  ra 2*xdb  ra 3*xdj ðtÞ j ¼ 1; 2; ::n ð42Þ

The somersault position update Eq. (43) introduces the somersault factor representing the
range and it is set to 2, ra2 and ra3 are two random numbers introduced that ranges between 0
and 1. For every movement, the best and optimal feature is identified based on the fitness
evaluation between the identified solutions. The optimal fitness evaluation is represented as
follows:
 0  0
P ; if the position P is better thanP
FE ¼ ð43Þ
P; otherwise

The above equation is used by all the foraging strategies to identify the best feature out of the
available intensity, texture and shape features. The output of the feature selection process will
be the optimal features covering the features related to intensity, texture and shape enclosed in
the MRI brain images. The high dimensional features are discarded efficiently through this
Multimedia Tools and Applications (2022) 81:8021–8057 8037

feature selection strategy. The pseudo-code for the feature selection process is depicted in
algorithm 1.

The selected features from the MRFO based feature selection process are then subjected to
classification process where the grades of the tumors are categorized.

3.5 Multi-grade classification using HDNN-AROA

The selected features with reduced dimensionality are vectorized and taken as the input of the
classifier and the grades of the tumors are guessed based on certain prediction rules. For tumor
grade classification, this article proposes the hybrid DNN-AROA technique to efficiently cate-
gorize the brain tumor with high classification accuracy. The classification strategy mainly aims to
reduce the error rate with improved accuracy. The tumor classification depends on the appropriate
features that are given as input that better signifies the growth of the tumors. The identified optimal
features in the feature selection phase constitutes the most important characteristics that deter-
mines the tumor growth and these features are directly given as input to the DNN. The DNN
classifier is composed of three major layers such as the input, hidden and output layers where the
hidden layers are responsible to learn the input features to carry out the classification of images.
The major use of the DNN is its adapting capability that can efficiently extract the input features
and categorizes the input through the identification of similarities between the features. The
architecture of the DNN for multi-grade brain tumor classification is depicted in Fig. 2.
8038 Multimedia Tools and Applications (2022) 81:8021–8057

The DNN is an efficient neural network that classifies the tumors based on the grade more
accurately than the other neural networks. The proposed DNN is a feed-forward kind
constituting one input layer, more than two hidden layers and one output layer. The hidden
layers are chosen accordingly based on the size of the input and is preferred to be more than 2
to improve the depth and the learning ability of the classifier. The selected features with high
discriminative nature are fed through the input layer in the DNN where the features are passed
to the hidden layers to get classified. The hidden layer possess neurons that learn from the
input features and classifies it accordingly. The DNN classifies the image based on the grades
into two parts as low-level grade tumors (LGT) and high-level grade tumors (HGT). Initially,
the input features that are fed to the neural network can be represented as follows:
X ¼ ðx1 ; x2 ; xn Þ ð44Þ

The sequence of features that are optimally selected constitutes the major part in classification
to accurately determine the grade of tumors. The input layer receives the input and passes it
directly to the hidden layer without any changes where the classification procedure is started
and performed. While processing, the input inside the hidden layer is mapped with certain
weight (wj) and bias (bk) values that are randomly chosen for every associated hidden layers.
The inputs for every sequential hidden layers can be framed as follows:
X
hk ¼ w j xj þ b k ð45Þ
j

In the above mentioned equation, wj determines the weight value of the predecessor added while
transmitting the input to the successor layer and j indicates the count of hidden layers given ask =
1, 2, . ……l. For every hidden layer, the weight and bias values are chosen randomly and updated
with the input values for further processing in the succeeding hidden layers.
The hidden layers in the DNN are associated with activation function to decide on the input
that has to be passed to the other hidden layer. Moreover, the training process can be efficiently
boosted up with the activation function that helps the neural network to learn any kind of
complex problems. There are many different activations like ReLU, sigmoid,tanh etc. followed
by various neural networks for various purposes. Out of all the activation functions, the ReLU
has proved to improve the learning process by maintaining a balance even after several
iterations. Thus the ReLU activation function has been chosen to trigger the output of every

I1
x1 I1
I2 Input nodes

o1 HGT
I3

H1
x2 I2 H2 Hidden Layers
Hn
o2
….

LGT
o1
….

….

….

xn I3 o2 Output Nodes

H1 H2 Hn

Fig. 2 Architecture of DNN for multi-grade brain tumor classification

Multimedia Tools and Applications (2022) 81:8021–8057 8039

hidden layer to the next layer and also to improve the efficiency of learning. The overall output
obtained from the hidden layer can be expressed based on the activation function as follows:
H ¼ f ðhk Þ ð46Þ
where, f(hk) = ReLU(hk) symbolizes the influence of the activation function in processing the
inputs. The overall output from the hidden layer comprises the weight and bias values added
along with the activation functions. This output is transmitted to the output layer of DNN to
achieve the classification output of brain tumor grade. In order to attain the classification output,
the most preferred non-linear softmax activation function is introduced for obtaining the multi-
grade classification output with high accuracy. The addition of softmax function helps in
converting the input into a probability set for tumor classification. The overall classification
output from the DNN can be demonstrated as follows:

eXk
σðX Þk ¼ Pl ð47Þ
k¼1 eXk

The output from the classifier will be a probability value indicating the tumor based on its
growth. The output comprises two labels as LGT and HGT where the intensity measures of the
tumor categorizes it into four different grades. The DNN comprises a loss function to indicate
the deviation of the classification accuracy from the original input. Minimization of loss
function is highly preferred to obtain a keen and accurate classification and to make the
system more efficient. The error function of the DNN classifier can be given as:

1X n
Lf ¼  yj logpj þ 1  yj log 1  pj ð48Þ
n j¼1

The DNN gets trained with vast and diverse range of input features and the classification accuracy
determines the quality of the system in identifying the grade. But, the DNN comprises weight
values that are changed randomly for every hidden layer which highly influences the output and
the accuracy of classification. To avoid this problem, the article proposes a hybridized strategy
where the optimization technique is used in updating the weight value of the DNN so that optimal
weight value will be chosen with an aid of minimizing the error rate of classification.

3.5.1 Updating weight based on AROA

To update the weight value by means of improving the classification accuracy, the article imitates the
rain’s behavior where the radius of the droplets are considered for parameter tuning. The training of
the DNN is a challenging task and introducing the ROA algorithm into the classifier is risky due to
various challenging factors based on the layers involved. The problem of hybridizing the ROA into
the DNN can be termed as NP-complete due to the risk factors involved in training. The algorithm is
imitated by taking the droplets as the search agents to identify the optimal weight factor. The droplets
indicated as the search agents can be demonstrated asΟ = (d1, d2, .…dr). The variables identified in
the droplet moves gradually downward and the limits of the variables are determined. The process is
repeated for all the identified variables inside the droplet. Meanwhile, the cost function is updated by
every single droplet. The radius of the droplet determines the weight factor of the hidden layer and the
minimal cost will be taken as the optimal weight factor. An objective function is formulated to
determine the optimality of the chosen weight factor. This can be given as follows:
8040 Multimedia Tools and Applications (2022) 81:8021–8057

 
Fit ¼ min Lf ð49Þ

For every new finding of the variables, the cost function is determines and the fitness function
is evaluated based on the above Eq. (50). The radius between the droplets influences the
weight optimization based on the distance between them. When two droplets with different
radii ν1andν2are located very close to each other, these two can combine to form a large
droplet with radius R as follows:
 1
R ¼  n1 þ  n2 n ð50Þ

while, n indicating the count of variables idealized in each droplet. Similarly, the droplet might
face changes based on the absorption by the soil and the properties of the soil. Assuming that
the droplet ν1 is absorbed by the soil, the droplet radius can be updated as follows:
 1
R ¼ λ n1 n ð51Þ

In the above Eq. (53) λ indicates the absorption rate of the droplet for every iterations. The search
agents perform the search to find the optimal weight out of all the randomly chosen weight factors
while classification. For every new finding by the droplet, the fitness function is evaluated to identify
the optimal weight to minimize the error rate and to achieve better and efficient classification
accuracy. The overall algorithm for updating the weight factor is depicted in algorithm 2.
Multimedia Tools and Applications (2022) 81:8021–8057 8041

Thus the brain tumor gets classified more accurately based on the grades. The output of the
classification process labels the trained images into LGT and HGT where the four different
grades of the tumors are identified. Moreover, the DNN with the help of AROA produces
more accurate output with reduced error.

4 Simulation analysis

This section presents the simulation analysis and results of the proposed approach to prove the
performance efficiency in comparison with the other existing state-of-the-art approaches. The
existing techniques like Gabor-modulated CNN [33], SVM + D-SEG [12], RFE-SVM [36],
mRMR+SVM [34], logistic regression [16], ensemble method [7], U-net NN [8] and hybrid
model [20] has been chosen for comparison for the metrics such as precision, recall, accuracy
and Dice Similarity Coefficient (DSC). The detailed description of the datasets used and the
experimental scenarios are presented in the next part and the performance analysis and
comparison are presented in the upcoming sections.

4.1 Simulation scenario

The proposed technique used the images obtained from two different datasets relevant to brain
tumor such as Radiopaedia dataset and Brain Tumor detection dataset (BTS). Many different
experiments are carried out to ensure the accuracy and efficiency of the proposed approach
against the other compared techniques. Among the dataset images, 80% of the total images are
considered for training and the remaining 20% are used for testing. The implementation of the
work is carried out in Python with the use of the Keras Tensorflow library.

4.1.1 Radiopaedia dataset

The radiopaedia dataset is an online dataset commonly used for the training and testing
purposes for brain tumor grade classification. The images in this dataset are classified into
four different grades that helps in efficient classification of any kind of brain tumor depending
on the growth of the tumor. This dataset comprises 121 MR brain tumor images that are
classified into four major grades like meningiomas, gliomas grade II, gliomas grade III and
glioblastomas. The tumor images are classified and labelled based on the guidelines described
by WHO. The brain images are acquired from the patients belonging to diverse racial
discriminations having different grades of tumors. This dataset comprises tumor images
partitioned into four grades as follows: 36 images with meningioma, 32 images with grade
II gliomas, 25 images with grade III gliomas and 28 images with glioblastomas.

4.1.2 BTS dataset

To overcome the deficiency of the images, another online dataset with more images called as
the BTS dataset is combined with the radiopaedia dataset. The BTS dataset is more appropriate
for brain tumor grade classification and it covers grade images such as meningioma, glioma
and pituitary tumor. These grade images are used along with the radiopaedia images in the
proposed technique for efficient classification. This dataset comprises 3064 T1-weighted MR
images with high contrast derived from 233 patients. The MRI images obtained from this
8042 Multimedia Tools and Applications (2022) 81:8021–8057

dataset is gray-scale in nature. The resolution of each of the MRI images range into 512×512
with 0.49×0.49 pixel size. The thickness of the image slice is 6 mm and the gap between every
image slice is identified as 1 mm. The dataset images are collected from two chief hospitals in
China such as the Nanfang Hospital and Tianjing Medical University. The cancer patients in
this hospital belong to different racial distributions and from different land areas. The images
of the patients are acquired between the time period 2005 and 2010. The grades of the tumors
in this dataset are partitioned based on the following specifications. For the three different
grades, a total of 1426 samples of MRI are collected from 92 Glioma patients, 930 samples are
collected from 62 Pituitary patients and 708 samples are collected from 82 Meningioma
patients.
The simulations of the proposed approach are carried out in the python toolkit with Intel i7
3.4GHz processor having 32GB RAM and NVIDIA GPU Geforce GTX 1080 Ti having
16GB VRAM operating on Windows 10. The DNN classifier has been built with one input
layer, three hidden layers and one output layer to improve the efficiency of classification.

4.2 Performance metrics

This part covers the performance metrics considered for evaluation of the proposed approach
against the other approaches. The major performance metrics such as precision, recall,
accuracy and DSC are considered for evaluation. The analysis are noted for all the other
compared techniques with the proposed HDNN-AROA technique to verify the efficiency.

(i) Precision: Precision is the most significant metric considered to evaluate the classification
approaches. It defines the rate of acceptably classified images based on the label i.e. the
total count of true positives. It is obtained as the sum of exactly labelled images belonging
to a particular class divided by the total count of images classified into the given label.
The formulation can be given as follows:

Tp
Pn ¼ ð52Þ
Tp þ Fp
In the above Eq. (52), Tpdenotes the true positive values and Fpdenotes the false positive
values. It helps in determining the properly classified images and thereby verifies the accuracy
of the overall system.

(ii) Recall: Recall is another important metric that helps in identifying the overall system
accuracy. It can be defined as the total count of exactly classified images out of the total
count of images that should have been classified for a particular label. The formulation
can be given as follows:

Tp
Rl ¼ ð53Þ
Tp þ Fn
Multimedia Tools and Applications (2022) 81:8021–8057 8043

where, Fnsignifies the false positives which means that the actual value is true but the result
obtained is false. The error in classification can be identified through this metric.

(iii) Accuracy: Accuracy is the most important metric that evaluates the efficiency and the
improvement rate of any proposed classification strategy compared with other ap-
proaches. The accuracy obtained should be eventually high than any other approaches
taken for comparison to prove the efficiency. The formulation for accuracy can be given
as follows:

Tp þ Tn
Ay ¼ ð54Þ
Tp þ Tn þ Fp þ Fn
The accuracy plays a vital role in determining the performance of the classification technique
in classifying the images into HGT and LGT. All the images are subjected to the classification
to identify the intensity of the tumor and accurate classification empowers the technique to
outperform the other techniques.

(iv) DSC: DSC is a statistical measure to identify the similarity between the input attributes.
The overlap between the two images can be identified using the following formula:

2Tp
DSC ¼ ð55Þ
2Tp þ Fp þ Fn
DSC metric relates both the precision and recall measures and helps in identifying the accuracy
of the system. The highest value of DSC can be 1 indicating maximum precision and recall and
the lowest value can be 0 indicating either the precision or the recall is 0.
All these metrics are computed for the two different datasets used and the method is
compared with the existing approaches to evaluate the performance of the approach with
different images.

4.3 Performance analysis

This part covers the comparative analysis of the proposed classification technique against other
approaches. From the overall simulations performed, it has been identified that the proposed
approach worked well in classifying the images accurately based on the growth of the tumors.
The images form the two different datasets are used to evaluate the proposed approach and the
classification strategy classifies the acquired images into HGT and LGT depending on the
tumor growth. The simulations proved the efficiency and the improvement in the accuracy rate
of the system against the other approaches. Moreover, the system clearly implied the efficiency
of the system in appropriate feature selection that accurately determines the growth of the
tumor.
The simulations performed suggested the importance of the features for the betterment of
any classification model. The behavior of the classification model over an input image depends
on the properties and efficiency of the features identified in the images. The input features of
8044 Multimedia Tools and Applications (2022) 81:8021–8057

the classification model should be able to characterize the tumor in all aspects. Based on the
nature of the input features, the performance of classification is identified. In the proposed
method, highly optimal and relevant features that can exactly define the brain tumor are
selected with the help of an efficient optimization algorithm. With the help of these optimal
features, the layers of DNN are able to perform the computations in an effective way.
The training parameters used for the training of the proposed architecture is depicted in
Table 1.
The performance evaluation of the proposed approach on two different datasets are as
follows:

4.3.1 Analysis of the proposed approach on Radiopaedia dataset

The images of the radiopaedia dataset are used where the 80% are used for training the
classifier and the remaining 20% are taken for testing. Though the number of images are
minimum, this dataset consists of images with all the four grades of tumor making it highly
suitable for the proposed grade-based brain tumor classification approach. For the tumor
images, the performance evaluation is done for the metrics such as precision, accuracy, recall
and DSC.
The summarization of the values obtained for the performance evaluation of the proposed
technique with other techniques for radiopaedia dataset is given in Table 2. From the table, it
has been clearly identified that the proposed technique shows a higher classification accuracy
than the other techniques. The gabor-modulated CNN technique also showed a higher and a
better precision value and the recall rate was also found to be good. The technique more or less
showed an equivalent accuracy rate to the proposed technique. Better accuracy has been
achieved by the proposed technique with the hybridization of the ROA onto DNN. For every
hidden layer, the high convergence nature of ROA helped in optimally selecting the weight
parameter to train the network. Since the random weight value for every iteration of the
training process influences the classification performance, the parameter has been optimized to
obtain higher accuracy rate. Therefore, the simulation of the proposed method yielded better
results for the radiopaedia dataset as depicted in Table 1.
The graphs demonstrate the performance improvement rate of the proposed approach
against the other approaches. The HDNN-AROA classifier has been used for classification
of the acquired images and the efficiency of the system has been improved by the addition of
the optimizer. The learning rate of the classifier has been improved and hence efficient
classification has been provided for all the images of the dataset. This is validated with the
help of the performance metrics and the hybridization of the optimizer with the classifier

Table 1 Hyper-parameters for classification

Training Parameters HDNN-AROA

Initial learning rate 0.001

Mini batch size 30
Max epochs 10
Learning algorithm AROA
Validation frequency 20
Drop factor for learning rate 0.9
Multimedia Tools and Applications (2022) 81:8021–8057 8045

Table 2 Summarization of the performance metrics for classification for Radiopaedia

Classification Techniques Precision (%) Recall (%) Accuracy (%)

Gabor Modulated CNN 99.77 96.39 98.68

SVM+D-SEG 97 90 94.7
RFE +SVM 95.5 84.6 87.8
mRMR +SVM 83.7 82.7 83
Logistic Regression 90 82 88
Proposed 99.6 99.38 99.5

showed better significance than the other approaches. The graphical representation of the
evaluation metrics are presented as follows:
The graphical representation in Fig. 3 depicts the performance of the proposed approach in
terms of precision against the other approaches. From the graph it has been identified that the
proposed approached shows high precision range than the other approaches. Gabor-modulated
CNN technique shows a high precision value that is closely equal to the proposed strategy. The
overall precision range of the proposed strategy for the radiopaedia dataset is found to be
99.6% whereas the gabor-modulated CNN strategy showed 99.77 precision rate. Better
precision value demonstrates the capability of the model in classifying the images accurately.
The proposed approach achieved a higher precision value with the help of the ROA algorithm
involved in the minimization of the random weight parameter. For every iteration, the weight
parameter has been minimized with the run of ROA, hence an optimal precision value has
been obtained. The overall precision rates scored by the compared methods like SVM + D-
SEG, RFE + SVM, mRMR+SVM and logistic regression over the radiopaedia dataset are
97%, 95.5%, 83.7% and 90% respectively.
From the graphical representation depicted in Fig. 4, it has been identified that the proposed
technique shows a higher recall value than the other techniques compared with it. In terms of
recall, the gabor-modulated CNN and SVM + D-SEG shows a close relevance with the
proposed technique but the recall value for the proposed technique is 99.38% whereas the
recall rates for gabor-modulated CNN and SVM + D-SEG are 96.39% and 90% respectively.
The recall rate ensures the capability of the model in classifying the images accurately with

Fig. 3 Graphical Representation of precision metric comparison for radiopaedia dataset

8046 Multimedia Tools and Applications (2022) 81:8021–8057

Fig. 4 Graphical Representation of recall metric comparison for radiopaedia dataset

respect to the input images. Also, this measure helps in identifying the behavior of the
classification model over the provided input features. Here, the proposed model is able to
score better recall value due to the hybridization of the approach in minimizing the error rates.
Therefore, the model avoids misclassification of the images leading to better recall rates than
the other approaches. The recall rates scored by the other compared methods like RFE +
SVM, mRMR+SVM and logistic regression over the radiopaedia dataset are 84.6%, 82.7%
and 82% respectively.
Accuracy is the most important measure in identifying the overall effectiveness of the
classification. In terms of accuracy, the proposed technique outperformed the other compared
approaches. This is depicted in Fig. 5 where the proposed technique shows a higher accuracy
rate which is more comparable to the gabor-modulated CNN approach. The accuracy rate of
the proposed technique is 99.5% and for gabor-modulated CNN, the accuracy rate is 98.68%.
Better accuracy values are obtained in the proposed model with the help of error minimization.
The ROA algorithm is made to run in the hidden layers of DNN to optimize the random
parameter for every iteration. Since the loss function is minimized for each and every iteration

Fig. 5 Graphical Representation of accuracy metric comparison for radiopaedia dataset

Multimedia Tools and Applications (2022) 81:8021–8057 8047

in the training process, the accuracy value has been improved than the other models. Also, the
simulation results have provided optimum classification results without any over-fitting
problems in the DNN. The proposed model is capable of yielding better results on both the
training and testing samples with the help of the hybrid features provided as input. The overall
accuracy yielded by the other compared methods like SVM + D-SEG, RFE + SVM,
mRMR+SVM and logistic regression over the radiopaedia dataset are found to be 94.7%,
87.8%, 83% and 88% respectively.
The summarized version of the performance values for DSC metric is depicted in Table 3.
From the table, it is clear that the proposed technique outperformed the other techniques
chosen for comparison. Higher DSC rate indicates higher performance efficiency of an
approach in classifying the images. The proposed approach shows higher DSC and hence it
can be concluded that the proposed scheme is highly efficient than the other approaches. This
is achieved due to the behavior of the model over the hybrid features and the capability of the
model in understanding the similarity of different pixels of images possessing the hybrid
features. The optimal choice of features has higher impact on the overall classification, as the
model trains itself based on the optimal features. Therefore, if the model has the ability in
understanding the similarities between the feature vectors, then it provides better and accurate
classification results.
The DSC of the proposed technique compared with the other techniques are represented
graphically in Fig. 6. The computation for DSC is based on the true positives identified by the
method along with penalizing the false positives identified. This measure indicates the
capability of a classification model in identifying the similarity between the input features of
all the input images. This is highly similar to the precision score and is expected to be high to
indicate optimum classification outcomes. The similarity between the input features are
identified and categorized efficiently by the proposed hybrid method than the other compared
methods. From the graph, it has been identified that the proposed technique provided a better
DSC rate than the other techniques. The gabor-modulated CNN is the only compared
technique that showed a significant relevance to the proposed technique in DSC rate. The
DSC rate of the proposed technique is 99.49% whereas the DSC rate of the Gabor-modulated
CNN is found to be 99.10%. Also, the DSC score secured by the other compared models like
ensemble method, U-net NN and hybrid model over the radiopaedia dataset are found to be
85%, 87% and 72% respectively.
Table 4 depicts the percentage of the tumor grades identified for the proposed and existing
techniques over the radiopaedia dataset. The values for each grades are computed based on the
classification results obtained through simulations. The overall accuracy percentage of the
proposed technique for grade I, grade II, grade III and grade IV are obtained to be 99.5%, 99%,
98.5% and 96% respectively.

Table 3 Summarization of the DSC performance metric for Radiopaedia

Classification Techniques DSC (%)

Gabor Modulated CNN 99.10

Ensemble Method 85
U-net NN 87
Hybrid Model 72
Proposed 99.49
8048 Multimedia Tools and Applications (2022) 81:8021–8057

Fig. 6 Graphical Representation of DSC metric comparison for radiopaedia dataset

4.3.2 Analysis of the proposed approach on BTS dataset

Apart from radiopaedia, similar performance metrics are also computed for the model using
the BTS dataset to understand the efficiency and performance of the model over diverse
datasets. Also, two different datasets are used so as to identify the differences in the values for
different data sizes. The images acquired from the BTS dataset are used for classification
where the 80% of the images are used for training and the remaining 20% are used for testing.
The same approaches are used for performance evaluation of the proposed approach against
the other approaches.
Table 5 summarizes the performance values obtained through simulation of the techniques
with BTS dataset. It is clear from the table that the proposed approach shows a higher accuracy
rate than any other techniques. This means that the true positive values of the proposed
classifier is high with minimized true negative rates. The accuracy of the system is highly
important to determine the classification efficiency of the system. However, the classification
accuracy is determined based on the values of precision and recall, hence these metrics are

Table 4 Overall accuracy for each class over radiopaedia dataset

Sl. No Methods Tumor Classes % of class accuracy

1. Gabor Modulated CNN (i) High-Grade Glioma (HGG) 97.5

(ii) Low-Grade Glioma (LGG) 99.9
2. SVM+D-SEG (i) Low Grade 93.2
(ii) High Grade 95.8
3. RFE+SVM (i) Metastasis 86.6
(ii) Meningioma 88.5
4. mRMR+SVM (i) Normal 82
(ii) Abnormal 84
5. Logistic Regression (i) Low-Grade Gliomas 87
(ii) Glioblastomas 89
6. Proposed (i) Grade I 99.5
(ii) Grade II 99
(iii) Grade III 98.5
(iv) Grade IV 96
Multimedia Tools and Applications (2022) 81:8021–8057 8049

Table 5 Summarization of the performance metrics for classification for BTS dataset

Classification Techniques Precision (%) Recall (%) Accuracy (%)

Gabor Modulated CNN 99.39 96.77 98.68

SVM+D-SEG 97 90 94.7
RFE +SVM 95.5 84.6 87.8
mRMR +SVM 83.7 82.7 83
Logistic Regression 90 82 88
Proposed 99.39 99.2 99.27

required to be optimized. Optimal values for these metrics are obtained by the proposed
approach through the use of an optimization algorithm in the layers of DNN with high
convergence and stability.
The graphical representation of the different performance metrics and the values obtained
indicating the performance improvement are as follows:
The graphical representation in Fig. 7 indicates the precision rate of the proposed approach
with other approaches. From the graph, it has been identified that the proposed approach has
shown higher precision rate than the other techniques. The gabor-modulated CNN showed an
even better precision rate than the proposed approach. The precision rate of the proposed
approach is found to be 99.39% whereas for gabor-modulated CNN it is found to be 99.5%.
The proposed technique performed well in identifying the true positives of the image samples
with minimum false rates. When the false rates are minimized, the classification accuracy can
be improved. The behavior of the classifier over the input features is based upon the domain
knowledge obtained by the classifier. Since the DNN is highly adaptable to any changes, it is
seen to be efficient enough in acquiring and understand the domain knowledge of different
data leaving efficient classification results. The precision rates scored by the other compared
methods like SVM + D-SEG, RFE + SVM, mRMR+SVM and logistic regression over the
BTS dataset are 97%, 95.5%, 83.7% and 90% respectively.
Figure 8 depicts the graphical representation of the recall rate of the proposed approach
compared with other approaches. From the graph, it has been identified that the proposed
approach showed a higher recall rate than the other techniques. The gabor-modulated CNN

Fig. 7 Graphical Representation of precision metric comparison for BTS dataset

8050 Multimedia Tools and Applications (2022) 81:8021–8057

Fig. 8 Graphical Representation of recall metric comparison for BTS dataset

and SVM + D-SEG techniques also showed a better recall rate when compared with the other
approaches. The recall rate of the proposed technique for BTS dataset is found to be 99.2%.
For the BTS dataset, the proposed model provided a better recall rate than the rest of the
techniques proving its efficiency in classifying datasets of higher sizes. The overall recall rates
for the compared techniques like gabor-modulated CNN, SVM + D-SEG, RFE + SVM,
mRMR+SVM and logistic regression over the BTS dataset are found to be 96.77% 90%,
84.6%, 82.7% and 82% respectively.
Figure 9 depicts the accuracy rate of the proposed technique for BTS dataset compared with
other approaches. From the graph, it is clearly identified that the proposed technique
outperformed the other techniques in terms of accuracy. The overall accuracy rate of the
proposed technique is found to be 99.27% which is significantly higher than any other
techniques. From the accuracy rate, it can be depicted that the proposed technique played
well in classifying the images based on the discriminative input features. In our proposed
scheme, the hybrid features are considered to train the layers of DNN to improve the capability
of the DNN in understanding the differences between the features and categorizing it

Fig. 9 Graphical Representation of accuracy metric comparison for BTS dataset

Multimedia Tools and Applications (2022) 81:8021–8057 8051

accordingly. The overall accuracy rate acquired by the other compared methods like gabor-
modulated CNN, SVM + D-SEG, RFE + SVM, mRMR+SVM and logistic regression over
the BTS dataset are found to be 98.68%, 94.7%, 87.8%, 83% and 88% respectively.
Table 6 illustrates the summarization of the DSC obtained by the proposed and existing
techniques for BTS dataset. From the table, it is clear that the proposed technique showed a
higher DSC value of 99.28% than any other proposed techniques. The existing techniques
compared with the proposed technique almost showed similar results for both the datasets
whereas the proposed technique showed higher accuracy in classifcation of radiopaedia
images. The precision and recall values for both the datasets are almost similar but the
accuracy values had a mild difference for the proposed technique. Based on the DSC scores
obtained for the different techniques over the BTS technique, it is clear that the proposed
technique is able to differentiate the wide variety of features of different pixels of the images
even with a large-sized dataset. Here, it can be summarized that the efficiency of DNN is really
high when compared with other conventional classification techniques in terms of differenti-
ating the vast count of features. Moreover, the feature selection phase played a major in
optimizing the feature vector set for the classifier to perform efficiently in classifying the brain
samples. The efficiency of the MRFO algorithm contributed in achieving such efficient results
through its global exploration capabilities and better convergence rates. Also, the chosen ROA
algorithm reduced the system complexity with simple implementation for weight optimization.
Thus, the DSC rate is effectively optimized with better values and with minimum time
complexity.
Figure 10 depicts the graphical representation of the proposed approach in terms of DSC for
the BTS dataset. From the graph, it is noted that the proposed technique and gabor-modulated
CNN techniques showed a higher DSC values than the other techniques. The DSC value of the
proposed technique for BTS dataset is 99.28% and that for gabor-modulated CNN is 99.10%.
The similarities of the tumor regions are identified by the proposed technique more accurately
than the other compared techniques. This capability helped the DNN to train itself with
advanced knowledge about the tumor features and sizes. This capability made the proposed
DNN classifier to identify the tumor grades and to classify it with better accuracy. Therefore,
from the simulations performed it is clear that the input features given to the classifier and the
ability of the classifier to train itself based on the differences in each and every input feature
has a major influence in the overall classification of the grades of tumors. The overall DSC
scored by the compared techniques like ensemble method, U-net NN and hybrid model over
the BTS dataset are found to be 85%, 87% and 72% respectively.
Table 7 depicts the percentage of the tumor grades obtained for the proposed and existing
techniques over the BTS dataset. The results for each class are obtained through simulations.
The overall accuracy of the proposed technique for grade I, grade II, grade III and grade IV are
obtained to be 99.3%, 98%, 97% and 96.5% respectively.

Table 6 Summarization of the DSC performance metric for classification for BTS dataset

Classification Techniques DSC (%)

Gabor Modulated CNN 99.10

Ensemble Method 85
U-net NN 87
Hybrid Model 72
Proposed 99.28
8052 Multimedia Tools and Applications (2022) 81:8021–8057

Fig. 10 Graphical Representation of DSC metric comparison for BTS dataset

Table 8 presents the intermediate results obtained through simulations based on the
different sizes of the radiopaedia dataset. The accuracy rate of the classification approach is
seen to reduce with respect to the increase in size of the dataset. This is due to the requirement
of training the DNN classifier with large number of images and the ROS is made to run for
every iterations with high dimensionality.
Table 9 presents the intermediate results obtained from simulations of the proposed method
over the BTS dataset. Compared to the radiopaedia dataset, the BTS dataset is seen to provide
lower accuracy in classification as the number of images in the BTS dataset is higher than the
radiopaedia dataset. But the optimization of the DNN classifier helped in achieving a better
accuracy for the proposed approach than the compared methods in every iteration.
An example of the training and testing images taken from the radiopaedia dataset are
depicted in Fig. 11. From the fig, the presence of the tumor on the brain can be identified. Both
the training and testing datasets consists of tumor regions and the proposed technique works on
classifying the images based on the tumor growth. The images are classified into HGT or LGT
depending on the growth and the HGT covers grade I and II and the LGT demonstrates the

Table 7 Percentage of tumor identified for each class over BTS dataset

Sl. No Methods Tumor Classes % of class accuracy

1. Gabor Modulated CNN (i) High-Grade Glioma (HGG) 97.5

(ii) Low-Grade Glioma (LGG) 99.9
2. SVM+D-SEG (i) Low Grade 93.2
(ii) High Grade 95.8
3. RFE+SVM (i) Metastasis 86.6
(ii) Meningioma 88.5
4. mRMR+SVM (i) Normal 82
(ii) Abnormal 84
5. Logistic Regression (i) Low-Grade Gliomas 87
(ii) Glioblastomas 89
6. Proposed (i) Grade I 99.3
(ii) Grade II 98
(iii) Grade III 97
(iv) Grade IV 96.5
Multimedia Tools and Applications (2022) 81:8021–8057 8053

Table 8 Validation of the accuracy rate for HDNN-AROA over different sizes of the radiopaedia dataset

Dataset Sizes Accuracy (%)

20% 99.9
40% 99.85
60% 99.75
80% 99.6
100% 99.5

grade III and grade IV tumors. Figure 12 depicts the four different grades of tumors taken from
the two different datasets. The location of the tumor on the brain region is identified and then
the basic features are extracted and then classified accordingly.

4.4 Comparison of the proposed framework with other classification approaches

From the simulations carried out, the proposed scheme is seen to dominate the other
approaches in terms of accuracy and other performance metrics. One of the compared scheme
that provided efficient resutls in our simulation outcomes is the Gabor-modulated CNN [33]
approach that dominated the proposed approach in terms of finding the true positives with
increased precision. The scheme resulted in 98.68% of accuracy over both the datasets. The
SVM_D-SEG [12] scheme is also much comparable to the proposed approach in terms of
classification accuracy. It resulted in a total of 94.7% overall accuracy for both the datasets.
The compared Logistic Regression [16] approach classified images with an average accuracy
rate slightly comparable to the proposed scheme. It showed an accuracy rate of 88% whereas
the RFE-SVM [36] scheme provided an accuracy rate of 87.8% over both the datasets. These
schemes accuracy are reduced due to error values in the training process that leaded to
misclassifications. The mRMR+SVM scheme [34] achieved a lower accuracy rate when
implemented over the radiopaedia and BTS datasets with an average accuracy rate of 83%.
The effectiveness and efficiency of the proposed approach has been compared with the
other state-of-the-art models relevant to brain tumor classification. For simulations we have
considered eight different classification models implemented over both the radiopaedia and
BTS datasets. It is clearly evident from the simulations that the proposed framework provided
optimal results compared to other classification approaches. A recent approach on classifica-
tion of brain tumors for MRI images using 3D features [14] has been studied where the
classification is performed through the KNN classifier. The authors presented features extract-
ed through 3D-CNN and GLCM for classification but it diminished the stability and accuracy
of the classification as the model is prone to errors under different sequential iterations.
Another approach using multiscale CNN [11] classified the tumor images into three different

Table 9 Validation of the accuracy rate for HDNN-AROA over different sizes of the BTS dataset

Dataset Sizes Accuracy (%)

20% 99.82
40% 99.69
60% 99.51
80% 99.38
100% 99.27
8054 Multimedia Tools and Applications (2022) 81:8021–8057

Training Dataset Testing Dataset

Fig. 11 An example of the training and testing images acquired from the radiopaedia dataset

classes but lacked performance efficiency due to its incapability in removing the skull and
vertebral column sections. An approach [13] combines meta-heuristic and learning method for
classification where the classification is performed using the Deep Belief Network (DBN).
Though the approach worked on effectively classifying the images, the poor convergence of
the meta-heuristic used in the layers of DBN decreased its overall performance and speed of
classification. The classification approach of Hiba Mzoughi et al. [25] using 3D CNN achieved
96.49% of accuracy rate with 284 images.
From the overall analysis, it is seen than the proposed model provided better results over the
radiopaedia dataset than the BTS dataset. With minimum count of images, the proposed model
is able to provide outstanding results compared to the existing models. The deviation in the
accuracy rate for the dataset with large images indicates the challenges faced by the system in
adapting to the dataset size. The ROA algorithm is able to offer fruitful results with minimum
images than that of a large dataset. The high dimensionality of the search space causes this
flaws to occur while dealing with the parameter tuning. Alternatively, the computational
complexity of the proposed approach has been optimized with the help of the feature selection
process using an efficient MRFO algorithm. The optimal features made the DNN to train itself
efficiently for different iterations.

4.5 Time complexity analysis

The time complexity of the proposed classification approach has been reduced with the help of
hybridization of an efficient ROA with the DNN classifier. Since the ROA algorithm is much
simple with minimum number of parameters, this algorithm is chosen to be implemented with
the DNN classifier with an ultimate aim of reducing the time complexity of the learning

Glioma Glioblastoma Pituitary Meningioma

Fig. 12 The sample of the tumor grade images acquired from the radiopaedia and BTS datasets
Multimedia Tools and Applications (2022) 81:8021–8057 8055

Table 10 Analysis of the time complexity for the proposed and existing methods for 100 images

Techniques Average Runtime (s)

Gabor modulated CNN 0.7526

SVM+D-SEG 1.6872
RFE +SVM 2.0056
mRMR +SVM 3.5829
Logistic Regression 3.0018
Proposed 0.1003

process. Also, the basic behavior of ROA tends to improve the overall speed of the weight-
tuning as well the classification processes.
Table 10 presents the average runtime of the proposed and the existing methods in
classifying the brain tumors based on the tumor grades. The time taken by the proposed
technique for training is computationally less compared to the existing methods. This is due to
the hybridization of an efficient ROA in the DNN over different runs. Since the ROA is simple
to implement, the run time of the method has been reduced. Moreover, the stability of the
solutions over different runs have been maintained with the help of the minimization of the
weight parameter over different iterations. For every iteration, the weight has been tuned in the
hidden layers of DNN making it computationally efficient and stable.

5 Conclusion

Multi-grade brain tumor classification plays a significant role in the medical field to identify
the growth and seriousness of the tumors in the brain. The MRI images can be used to identify
the tumors and to present a detailed study on it. This paper presents a solution for multi-grade
brain tumor classification with higher accuracy. The images acquired from the radiopaedia and
BTS datasets are used in the proposed approach and those images are subjected to pre-
processing initially. Then, the piFCM clustering method is used for image segmentation to
reduce the uncertainty. After segmentation, the feature extraction process is proposed to derive
the most important intensity, texture and shape features from the images. After extraction, the
major features are selected using the MRFO algorithm based on the foraging strategies of
manta rays. Finally the images are subjected to classification where the technique proposed the
HDNN-AROA technique to classify the images based on the grades. The technique classified
the images into HGT and LGT based on the tumor growth. The proposed technique showed a
higher accuracy rate for both the datasets when compared with the other techniques. The
analysis of the proposed technique proved that the technique is more efficient than any other
existing techniques and the precision, recall, accuracy and DSC for radiopaedia dataset are
found to be 99.6%, 99.38%, 99.5% and 99.49% and for BTS dataset are found to be 99.39%,
99.2%, 99.27% and 99.28% respectively.
Though the technique played well in classifying the grades of brain tumors, there are some
limitations in the proposed model. The adaptation of ROA into the DNN classifier is a
challenging task as the algorithm requires further knowledge in optimal setting of its param-
eters. When the hidden units are increased, the algorithm is seen to stick with the local
minimum due to the high dimensionality. Since the learning model is deep with many neurons,
the ROA algorithm is required to run for multiple iterations and hence might degrade the
8056 Multimedia Tools and Applications (2022) 81:8021–8057

overall performance for huge datasets. Therefore, the future scope relies on building a more
reliable model for classifying the brain tumors with the choice of a more efficient optimization
algorithm that can work well even with larger datasets while functioning with minimum
parameters.

References

1. Aboelenein NM, Songhao P, Koubaa A, Noor A, Afifi A (2020) HTTU-net: hybrid two track U-net for
automatic brain tumor segmentation. IEEE Access 8:101406–101415
2. Ali MB, Gu IY, Berger MS, Pallud J, Southwell D, Widhalm G, Roux A, Vecchio TG, Jakola AS (2020)
Domain mapping and deep learning from multiple MRI clinical datasets for prediction of molecular
subtypes in low grade gliomas. Brain Sci 10(7):463
3. Ayadi W, Elhamzi W, Charfi I, Atri M (2021) Deep CNN for brain tumor classification. Neural Process Lett
53:1–30
4. Baliarsingh SK, Vipsita S (2020) Chaotic emperor penguin optimised extreme learning machine for
microarray cancer classification. IET Syst Biol 14(2):85–95
5. Baliarsingh SK, Vipsita S, Gandomi AH, Panda A, Bakshi S, Ramasubbareddy S (2020) Analysis of high-
dimensional genomic data using MapReduce based probabilistic neural network. Comput Methods Prog
Biomed 195:105625
6. Baliarsingh SK, Muhammad K, Bakshi S (2021) SARA: a memetic algorithm for high-dimensional
biomedical data. Appl Soft Comput 101:107009
7. CabezasM, Valverde S, González-Villà S, Clérigues A, SalemM, Kushibar K, Bernal J, Oliver A, Lladó X
(2018) Survival prediction using ensemble tumor segmentation and transfer learning. arXiv preprint arXiv:
1810.04274
8. Caver E, Chang L, Zong W, Dai Z, Wen N (2018) Automatic brain tumor segmentation using a U-net
neural network. In: Pre-conference proceedings of the 7th MICCAI BraTS challenge, 63-73
9. Deepak S, Ameer PM (2019) Brain tumor classification using deep CNN features via transfer learning.
Comput Biol Med 111:103345
10. Dequidt P, Bourdon P, Ahmed OB, Tremblais B, Guillevin C, Naudin M, Fernandez-Maloigne C, Guillevin
R (2019) Recent advances in glioma grade classification using machine and deep learning on MR data. In:
2019 fifth international conference on advances in biomedical engineering (ICABME), IEEE, 1-4
11. Díaz-Pernas FJ, Martínez-Zarzuela M, Antón-Rodríguez M, González-Ortega D (2021) A deep learning
approach for brain tumor classification and segmentation using a multiscale convolutional neural network.
Healthcare 9(2):153
12. Hsieh KL-C, Lo C-M, Hsiao C-J(2017)Computer-aided grading of gliomas based on local and global MRI
features. Comput Methods Prog Biomed 139:31–38
13. Hu A, Razmjooy N (2021) Brain tumor diagnosis based on metaheuristics and deep learning. Int J Imaging
Syst Technol 31(2):657–669
14. Jefferson B, Shanmugasundaram RS (n.d.) Brain Tumor Classification in 3d-Mri Using Features From
Radiomics and 3d-Cnn Combined With Knn Classifier International Journal of Electrical Engineering and
Technology (IJEET) 12
15. Joardar BK, Jayakodi NK, Doppa JR, Li H, Pande PP, Chakrabarty K (2020) GRAMARCH: a gpu-reram
based heterogeneous architecture for neural image segmentation. In: 2020 Design, Automation & Test in
Europe Conference & Exhibition (DATE), IEEE, 228-233
16. Jones TL, Byrnes TJ, Yang G, Howe FA, Bell BA, Barrick TR (2015) Brain tumor classification using the
diffusion tensor image segmentation (D-SEG) technique. Neuro-oncology 17(3):466–476
17. Kumar RL, Kakarla J, Isunuri BV, Singh M (2021)Multi-class brain tumor classification using residual
network and global average pooling. Multimed Tools Appl 80:–34
18. Kurc T, Bakas S, Ren X, Bagari A, Momeni A, Huang Y, Zhang L, Kumar A, Thibault M, Qi Q, Wang Q
(2020) Segmentation and classification in digital pathology for glioma research: challenges and deep
learning approaches. Front Neurosci 14:27
19. Maekawa T, Hori M, Murata K, Feiweier T, Kamiya K, Andica C, Hagiwara A, Fujita S, Koshino S, Akashi
T, Kamagata K (2020) Differentiation of high-grade and low-grade intra-axial brain tumors by time-
dependent diffusion MRI. Magn Reson Imaging 72:34–41
20. Meier R, Bauer S, Slotboom J, Wiest R, Reyes M (2013) A hybrid model for multimodal brain tumor
segmentation. Multimodal Brain Tumor Segmentation 31:31–37
Multimedia Tools and Applications (2022) 81:8021–8057 8057

21. Menze B, Isensee F, Wiest R, Wiestler B, Maier-Hein K, Reyes M, Bakas S (2020) Analyzing magnetic
resonance imaging data from glioma patients using deep learning. Comput Med Imaging Graph 88:101828
22. Minaee S, Boykov Y, Porikli F, Plaza A, Kehtarnavaz N, Terzopoulos D (2020) Image segmentation using
deep learning: a survey. arXiv preprint arXiv:2001.05566. Transactions on Pattern Analysis and Machine
Intelligence
23. Moazzeni AR, Khamehchi E (2020) Rain optimization algorithm (ROA): a new metaheuristic method for
drilling optimization solutions. J Pet Sci Eng 195:107512
24. Mohanaiah P, Sathyanarayana P, GuruKumar L (2013) Image texture feature extraction using GLCM
approach. Int J Sci Res Publ 3(5):1–5
25. Mzoughi H, Njeh I, Wali A, Slima MB, BenHamida A, Mhiri C, Mahfoudhe KB (2020) Deep multi-scale
3D convolutional neural network (CNN) for MRI gliomas brain tumor classification. J Digit Imaging 33:
903–915
26. Narmatha C, Eljack SM, Tuka AARM, Manimurugan S, Mustafa M (2020) A hybrid fuzzy brain-storm
optimization algorithm for the classification of brain tumor MRI images. Journal of Ambient Intelligence
and Humanized Computing:1–9
27. Pang S, Du A, Orgun MA, Yu Z, Wang Y, Wang Y, Liu G (2020) CTumorGAN: a unified framework for
automatic computed tomography tumor segmentation. Eur J Nucl Med Mol Imaging 47(10):2248–2268
28. Prabhu LA, Jayachandran A (2018) Mixture model segmentation system for parasagittal meningioma brain
tumor classification based on hybrid feature vector. J Med Syst 42(12):1–6
29. Ryall S, Tabori U, Hawkins C (2020) Pediatric low-grade glioma in the era of molecular diagnostics. Acta
Neuropathol Commun 8(1):1–22
30. Salehi H, Vahidi J, Abdeljawad T, Khan A, Rad SYB (2020) A SAR image Despeckling method based on
an extended adaptive wiener filter and extended guided filter. Remote Sens 12(15):2371
31. Sasank VV, Venkateswarlu S (2021) Brain tumor classification using modified kernel based softplus
extreme learning machine. Multimedia Tools Appl 80(9):13513–13534
32. Shaukat F, Raja G, Ashraf R, Khalid S, Ahmad M, Ali A (2019) Artificial neural network based
classification of lung nodules in CT images using intensity, shape and texture features. J Ambient Intell
Humaniz Comput 10(10):4135–4149
33. Singh R, Goel A, Raghuvanshi DK (2020) Computer-aided diagnostic network for brain tumor classifica-
tion employing modulated Gabor filter banks. Vis Comput 37(8):2157–2171
34. Soltaninejad M, Yang G, Lambrou T, Allinson N, Jones TL, Barrick TR, Howe FA, Ye X (2017)
Automated brain tumour detection and segmentation using superpixel-based extremely randomized trees
in FLAIR MRI. Int J Comput Assist Radiol Surg 12(2):183–203
35. Udendhran R, Balamurugan M, Suresh A, Varatharajan R (2020) Enhancing image processing architecture
using deep learning for embedded vision systems. Microprocess Microsyst 76:103094
36. Zacharaki EI, Wang S, Chawla S, Yoo DS, Wolf R, Melhem ER, Davatzikos C (2009) Classification of
brain tumor type and grade using MRI texture and shape in a machine learning scheme. Magn Reson Med
62(6):1609–1618
37. Zhang HM, Dong B (2020) A review on deep learning in medical image reconstruction. J Oper Res
Soc China:1–30
38. Zhao W, Zhang Z, Wang L (2020) Manta ray foraging optimization: an effective bio-inspired optimizer for
engineering applications. Eng Appl Artif Intell 87:103300
39. Zhou Z, He Z, Jia Y (2020) AFPNet: a 3D fully convolutional neural network with atrous-convolution
feature pyramid for brain tumor segmentation via MRI images. Neurocomputing 402:235–244

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