Measurement: Sensors 31 (2024) 101026

Contents lists available at ScienceDirect

Measurement: Sensors
journal homepage: www.sciencedirect.com/journal/measurement-sensors

MRI brain tumor detection using deep learning and machine

learning approaches
Shenbagarajan Anantharajan a, *, Shenbagalakshmi Gunasekaran a, Thavasi Subramanian a,
Venkatesh R b
a
Mepco Schlenk Engineering College, Sivakasi - 626005, Virudhunagar, Tamilnadu, India
b
Ramco Institute of Technology, Rajapalayam - 626117, Virudhunagar, Tamilnadu, India

A R T I C L E I N F O A B S T R A C T

Keywords: The development of aberrant brain cells, some of which may become cancerous, is known as a brain tumour. The
Brain tumour quality of life and life expectancy of patients are enhanced by early and timely illness identification and treat­
Magnetic resonance imaging (MRI) ment plans. Magnetic Resonance Imaging (MRI) scans are the most common approach for finding brain tumors.
Deep learning(DL)
However, the ability of radiologists and other clinical experts to identify, segment, and remove contaminated
Machine learning(ML)
tumour regions from MRI images is a critical factor in a process that is iterative and labor-intensive and relies on
Adaptive contrast enhancement algorithm
(ACEA) those individuals’ abilities in these areas. Concepts for image processing may envision the diverse human organ
Gray-level co-occurrence matrix (GLCM) anatomical structures. It is difficult to find abnormal brain regions using simple imaging methods. Over the last
Ensemble deep neural support vector machine several years, interest in the emerging machine learning field of “Deep Learning (DL)" has grown significantly. It
(EDN-SVM) was extensively used in numerous applications and shown to be an effective Machine Learning (ML) technique
for many of the challenging issues. This research suggests a novel MRI brain tumour detection method based on
DL and ML. Initially the MRI images are collected and preprocessed using Adaptive Contrast Enhancement Al­
gorithm (ACEA) and median filter. Fuzzy c-means based segmentation is done to segment the preprocessed
images. The features like energy, mean, entropy and contrast are extracted using Gray-level co-occurrence matrix
(GLCM). The abnormal tissues are classified using the proposed Ensemble Deep Neural Support Vector Machine
(EDN-SVM) classifier. The numerical findings reveal a better accuracy (97.93 %), sensitivity (92 %), and spec­
ificity (98 %) in recognizing aberrant and normal tissue from brain MRIimages, which supports the effectiveness
of the approach that was recommended.

1. Introduction detect the hyperactivity of tumour cells in the body [2]. Consequently,
MRI scans are considered to be the most effective diagnostic tracers for
The human brain is regarded to be one of the most essential organs detecting breaches in the BBB. Fig. 1(A) and (B) depicts the images of
since it is responsible for a large number of the body’s regulatory pro­ healthy brain and tumour brain.
cesses, including memory, emotions, vision, motor skills, responses, and There are between 7 and 11 cases of brain tumors per 100,000 people
breathing. In the event that a tumour begins to form inside the brain, in various age groups per year. It is estimated that 227,000 people die
these functions will be significantly disrupted [1,3]. This tumour is each year as a result of this dreadful illness. In addition, about 7.7
either a primary brain tumour (BT), which develops from inside the million survivors are adjusting to life with a disability [4]. As well as
brain itself and represents the development of brain tissues themselves, saving lives, an early diagnosis of a brain tumour may help prevent
or it is a metastatic BT, which develops in another part of the body and disability. The brain, the body’s most delicate organ, will be subjected to
eventually spreads to the brain. When compared to tumors that originate less modification and surgery if it is detected early. To begin with, a
in any other organ of the human body, those that occur in the brain radiologist will need to take a picture of the affected area in order to do a
provide a significant diagnostic challenge. Because the brain has the manual diagnosis [5]. After that, an experienced physician is consulted
“Blood-Brain Barrier (BBB)”, ordinary radioactive markers are unable to for the purpose of image analysis and the formulation of a treatment

* Corresponding author.
E-mail addresses: [email protected] (S. Anantharajan), [email protected] (S. Gunasekaran), [email protected] (T. Subramanian), venkey88me@
gmail.com (V. R).

https://doi.org/10.1016/j.measen.2024.101026
Received 4 January 2023; Received in revised form 3 March 2023; Accepted 2 January 2024
Available online 7 January 2024
2665-9174/© 2024 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
S. Anantharajan et al. Measurement: Sensors 31 (2024) 101026

strategy. Unfortunately, the research that investigated the accuracy of methodology as well as a deep neural network technique to categorise
manually diagnosing brain tumors reported a discrepancy amongst the an MRI as “tumour detected” or “tumour not detected.” The study [13]
experts who reviewed the data. According to reports, the level of showed the promise of DL in MRI scans as a non-invasive method for
agreement amongst specialists for the manual diagnosis of a BT is be­ simultaneous and automated tumors segmentation, identification, and
tween 90 and 95%. The degree of disagreement amongst the specialists grading of LGG in clinical settings. The research [14] presents a faster
is further reduced when it comes to mixed types of tumour, mixed gli­ and more accurate method for detecting human brain cancers by
oma, and medulloblastoma, falling to 77 % and 58 %, respectively [6]. combining the “Template-based K-means (TK)” algorithm with pixels in
Digital image processing and other advancements in medical imag­ the image and “Principal Components Analysis (PCA)”.
ing have contributed to the widespread use of computer-aided diagnosis The “Watershed Dynamic Angle Projection - Convolution Neural
(CAD) in recent years. The MRI technique is favoured for use in diag­ Network (WDAPP-CNN)” is able as a nation method for tumors identi­
nostic systems like these since it does not pose a threat from ionising fication in this research [15]. The tumors area was successfully
radiation and is able to reliably identify blood flow in veins [7]. The use segmented using the watershed technique. The research [16] suggested
of large medical image datasets, such as Brain MRI scans, for the iden­ technique guarantees to be very effective and exact for detecting, clas­
tification of BT may be aided by the use of ML and DL algorithms. sifying, and segmenting brain tumors. Automated segmentation is per­
Creating a ML and DL model is a multistep process that involves training formed on image data using a CNN-based method, which employs very
using a significant quantity of medical imaging data [8]. This is neces­ small kernel sizes of 33. The author of [17] focused early identification
sary in order to get the correct prediction or insight from the model, of benign brain tumors. Segmentation is required in the early stages of
which is necessary in order to make an appropriate clinical decision. In brain tumors identification. Algorithms usually for segmentation have
this study, we investigate the identification of brain tumors using DL and several limitations, including the inability to handle noisy data and the
ML techniques. inability to identify subtle intensity variations in the image. The study of
The major contributions of this study are as follows. [18] presented a comprehensive and entirely automated MRI brain tu­
mors identification and segmentation approach employing the
• To preprocess the MRI images, Adaptive Contrast Enhancement Al­ “Gaussian mixture model”, “Fuzzy C-Means”, “active contour”, “wavelet
gorithm (ACEA) and median filter is used. transform”, and “entropy segmentation” techniques as an effective
• Fuzzy c-means based segmentation is done to segment the pre­ clinical-aided tool. The two key components of the suggested approach
processed images. are tumors auto-detection and segmentation as well as skull removal.
• The features are extracted using GLCM approach. The research [19] suggested approach tries to distinguish between BT
• The abnormal tissues are classified using the proposed EDN-SVM and normal brains. Brain magnetic resonance imaging is used to
classifier. research various forms of brain malignancies. Support vector machines
and various wavelet transformations are used to identify and categorise
The study is structured in such a manner that Section 2 provides MRIbrain cancers. The Study [20] proposed hybrid K-means Galatic
related work and problem statement, Section 3 outlines the recom­ Swarm Optimization (GSO) technique is adopted as a practical solution
mended approach, Section 4 depicts findings and discussion, and Section to the image segmentation issue, which is considered as a classification
5 draws a conclusion to the research along with the future work that will model. Study developed a Fuzzy C-Means clustering technique, which
be done. was followed by conventional detectors and CNN to remove brain tu­
mors from 2D MRI. The experiment utilised real-time dataset with
2. Related WORKS various tumors sizes, locations, forms, and image brightness. The author
of [21] presented a comprehensive assessment of the literature on cur­
The author of [9] suggested method identifies the kind of tumors rent approaches to segmenting BT from brain MRI data. The author of
present in the BT MRI image and marks the tumour region. Alex Net [26] provided a thorough critique of the research and discoveries made
model and the Faster R–CNN algorithm’s Region Proposal Network in the recent past to identify and categorise brain tumors using MRI
(RPN) are utilised as the basic models for classifying various tumour scans. Researchers that specialize in deep learning and are interested in
kinds. The study [10] employed a Deep NN classifier, a component of the using their knowledge for the identification and classification of brain
deep learning designs, to divide 66 brain MRI scans into four categories, tumors may particularly benefit from this work. According to the study
including “normal”, “glioblastoma”, “sarcoma”, and “metastatic bron­ [27] an automated approach is offered to distinguish between malignant
chogenic carcinoma tumors”. The author of [11] constructed brain MRI and non-cancerous brain MRI scans by Using three benchmark datasets,
images were utilised to create a Convolutional Neural Network (CNN) to the suggested technique is verified, with average results of 97.1 % ac­
identify a tumour. In Ref. [12], the author used a CNN-based curacy, 0.98 area under the curve, 91.9 % sensitivity, and 98.0 %

Fig. 1. Image of (A) Healthy brain (B) Brain with a tumour.

2
S. Anantharajan et al. Measurement: Sensors 31 (2024) 101026

specificity. Compared to current techniques, it can be utilised to detect applied, and a GLCM matrix is used to extract the features of the images.
the tumour more precisely and with less processing time. The study [28] The EDN-SVM approach is then used to classify the images of healthy
proposes a two-step Dragonfly algorithm (DA) clustering method to and tumorous brain tissue.
precisely extract starting contour points. At the preprocessing stage, the
brain is removed from the skull. Then, tumour edges are extracted using 3.1. Dataset collection
the two-step DA, and these extracted edges are utilised as a starting
contour for the MRI sequence. We utilize a dataset that may be found on the Kaggle open data
website in order to evaluate the performance of the suggested archi­
2.1. Problem statement tectural design. This dataset includes 255 T1-mode MRI images. It in­
cludes 98 MRI slices taken from healthy brain tissue and 155 MRI slices
Brain tumors have the potential to generate consequences such as taken from tumorous brain tissue. Because each of these images had a
physical impairments, which would then need patients to undergo very unique dimension, we needed to adjust it so that it would fit inside the
intensive therapy, which is often rather painful, in order to cure or lessen parameters of our image requirements. A portion of the dataset that we
the caused disabilities. In addition, the negative effects that brain tumors used for our investigation is shown in Fig. 3(A). It depicts that the width
have on the functioning of the brain might vary depending on the size of and height of the images vary from one another. When looking at MRI
the tumour, where it is located, and what kind it is. Because a tumour brain scans with various heights and widths, it might be difficult to
might exert pressure on the region of the brain that regulates the body’s appropriately classify healthy brain tissue and tumorous brain tissue. As
mobility, the patient can become immobile as a result of this. If it is a consequence of this, before moving on to the preprocessing stage, we
diagnosed sooner, it may be possible to prevent disability from occur­ resize the images such that their width and height are the same. The
ring. There are a number of obstacles that need to be overcome in order scaled versions of the images from the dataset are shown in Fig. 3(B).
to correctly categorise brain tumors. These obstacles include the fact
that brain tumors exhibit a high degree of variation with regard to their 3.2. Preprocessing
size, shape, and intensity, and that tumors of different pathological types
may have similar outward appearances. 3.2.1. Adaptive contrast enhancement algorithm
MRI image contrast is crucial for tumour identification since this
3. Proposed methodology technique relies heavily on image brightness. The two typical histo­
grams for MRI scans are high contrast and low contrast. A piecewise
This section provides a comprehensive discussion of the identifica­ linear histogram modification is often used to improve intensity
tion of MRI brain tumors utilizing both DL and ML techniques. The contrast. Because the transformation slope varies from one set of data to
progression of the recommended technique is shown in Fig. 2. In the another, it is difficult to establish universal minimum and maximum
beginning, MRI brain tumour data were obtained and preprocessed with values that apply to all images. Here, we use an automated and dynamic
the help of ACEA and the median filter in order to get rid of the noise. In method to extract the parameters from each image. A brain tumour (T),
order to segment the MRI brain images, a fuzzy c-means technique is healthy brain tissue (B), and vessels (V) can all be distinguished in an

Fig. 2. Workflow of the proposed technique.

3
S. Anantharajan et al. Measurement: Sensors 31 (2024) 101026

Fig. 3. Sample dataset images of (A) varying size (B) same size.

MRI scan of the head. To begin, we take the training data and choose used to re-estimate the intensity histograms of brain images (Fig. 4).
some random voxels to represent each of the three classes. It is hy­ ⎧
pothesized what the intensity probability distribution function (PDF) ⎨ (Ix − Pmin ) (255 − 1)+1 if (Pmin ≤ Ix ≤ Pmax )
⎪
looks like for each category. (Pmax − Pmin )
⎪
⎩ (1)
In addition to averaging their values (μT Z , μB Z , μV Z ), and calculating 0 if (Ix < Pmin ∪ Ix > Pmax )
their standard derivations (σT, σ B , σV, ), we additionally calculate the Pmin : μtumor − 3σ , Pmax : μvessel + 3σ
highest PDF intensity value for the brain class, NZ . The mean value have
the μT Z < μB Z < μV Z characteristic. While mean and derived values 3.2.2. Median filter
may vary from image to image, the normal brain always has a higher When applied to MRI scans of the brain, this nonlinear technique
mean value than the tumour and a lower mean value than the blood effectively eliminates unwanted background noise. Edge preservation
vessels. One may use this to divide things up into three categories and with this approach is common practise. Salt and pepper noise may be
we name it Curve Pattern Z. First, we determine the intensity PDF of a eliminated with great success. Median filter is similar to mean filter in
brand-new image, using the segmented brain volume as input. We find that it iteratively processes an image, but it replaces each pixel’s value
the volume of the new brain that corresponds to the maximum proba­ with the middle value of its neighbours rather than the mean. After
bility density NV of the probability density function (PDF). Healthy tis­ sorting all the neighbouring pixel values into numerical order, the me­
sues dominate volume, therefore this assumption is reasonable. We may dian pixel value from the neighbouring pixel values is used to replace the
determine the mean intensity values of the other two classes by pixel under consideration. When it comes to decreasing noise without
adjusting the three curves in Curve Pattern Z. Tumour mean value and diminishing image quality, the median filter is much superior.
vessel values are μT = NV − NZ + μT Z and μV = NV − NZ + μV X . All three }
classes employ conventional derivations. Set the lower limit as Pmin = c (i, j) = median(o,b)∈Gij {k{o, b) (2)
μT − 3σ T and the upper limit as Pmax = μV − 3σ V . The formula for in­
tensity transform of range 0–255 is given in equation (1). The formula is where Gij denotes the sets of coordinates centred on the coordinates (i,j)
within the rectangle sub image frame, and median is the median value
inside the window. To demonstrate how effective the noise reduction
approach is, we apply it to the noisy picture and the result is presented in
Fig. 5.

3.3. Fuzzy C-means (FCM) segmentation

The process of image segmentation benefits greatly from the use of

fuzzy clustering. An efficient approach that may be used in fuzzy clus­
tering is known as the fuzzy c-means algorithm. The FCM is a clustering
approach that makes it possible for a single pixel to belong to many
clusters at the same time. The FCM method makes an effort to divide a
finite collection of pixels into a collection of “C" fuzzy clusters by
applying certain supplied criteria to the decision-making process. The
FCM method seeks to achieve the lowest possible value for the objective
function shown below
f f
( ) ∑ ∑ ∑
m
Y O, f1 , f2 , …, ff = Yx = = onxy s2xy (3)
x=1 x=1 y=1

sxy is the Euclidean distance between the xth centroids and yth data
Fig. 4. Transformation histogram. point. nF [1,∞] is a weighting function. oxy is between 0 and 1. An

4
S. Anantharajan et al. Measurement: Sensors 31 (2024) 101026

1. Initialize O = [oxy ] matrix, O(O) .

2. At k-step: calculate the centers vectors F(g) = [f y ] with O(g) .
∑
M
onxy ix
f y = x=1 (6)
∑M
onxy
x=1

Update O(g) , O(g+1) .

1
oxy = ( )n−2 1 (7)
∑
f
‖ix − f y‖
g=1 ‖ x g‖
i − f

⃦ ⃦
If ⃦O(g+1) − O(g) ⃦ < ε then STOP; otherwise return to step 2.
D. Feature extraction using Gray level Co-occurrence matrix (GLCM).
The GLCM technique is used in order to extract the image features in
this approach. A statistical approach known as the co-occurrence matrix
is used to a particular brain image in order to extract the second order
statistical textural qualities. When using GLCM, the number of gray
levels must always be equal to the number of rows and columns. The
following equations are used to extract the characteristics that are based
on the first order of the histogram. In the approach that was suggested,
shape, textural, and statistical characteristics are taken from each clus­
ter, and then these features are supplied to the classifier so that it can
locate the tumors in the MRI image that was provided. The following are
the extracted features:
∑1 mk−
mk− ∑1
Fig. 5. Noise removed image using median filter method. Entropy = Txy log Txy (8)
x=0 y=0

iterative optimization of the objective function is used to accomplish A random variable’s entropy may be thought of as a measurement of
fuzzy portioning of known data samples. its degree of uncertainty. When all of the components of the co-
occurrence matrix are the same, its value will be at its highest possible
1
oxy = ( )(2/n− 1)
(4) level.
f
∑ sxy
sgy ∑1 mk−
mk− ∑1
g=1 (x, y)t(x, y) − μi μj
(9)
x=0 x=0
Correlation =
∑
m σi σj
onxy iy
y=1
fxy = ∑ (5) The correlation metric determines how closely connected the refer­
m
onxy ence pixel is to its neighbouring pixel.
y=1
∑1 mk−
mk− ∑1
The iteration will end when maxxy {|o(g+1)
xy − o(g)
xy |} < ε, where ε is a
Energy = T2xy (10)
termination condition between 0 and 1, and g is the iteration steps. The
x=0 y=0

method approaches a local minimum or a saddle point of yn . Fig. 6 de­ Energy is the metric used to define the sum of squared components.
picts the MRI brain image after segmentation. Here are the individual This determines how homogenous the mixture is. When pixels are sub­
components that make up the algorithm: stantially comparable to one another, the energy value will be high.

Fig. 6. MRI brain image after FCM approach.

5
S. Anantharajan et al. Measurement: Sensors 31 (2024) 101026

∑1
mk− ∑1 mk−
mk− ∑1 discrepancies between the model’s output and the instance’s target
Contrast = m2 T(x, y)2 (11) value, a range of permissible deviations is denoted by F, and the total
number of training samples is denoted by h. F is a parameter controlling
m=0 x=0 =0

The contrast of an image is the difference in brightness between the the magnitude to which such deviations are allowed, and h is the total
reference pixel and its neighbouring pixel. number of training samples. The terms “primal goal” and “primal vari­
ables” refer, respectively, to the variables in Equation (15). By intro­
∑1
mk−
Mean (μ) = x.t(x) (12) ducing the Lagrange multipliers α, α(∗) and solving for the saddle point’s
x=0 coordinates, we may rewrite the original goal and constraints as follows.
The average amount of brightness of an image or texture is described [
∑ h
( ∗ )
∑h
by its mean. Around the mean, the variance characterises the range of − ε
∗
(αx + αx )+ αx − αx jx
max
values in terms of their relative intensity. (17)
x=1
x=1
]
( )
α(∗) − 1 ∑ ( α∗ − α ) α∗ − α ( I − I )
h
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
√mk− 1 x y x y
√∑ 2 x,y=1 x y

Standard Deviation(σ) = √ (x − μ)2 .t(x) (13)

x=0
conditional on the following:
∑1
( ) mk− ∑
h
( )
Variance σ2 (x − μ).t(x) (14) 0 ≤ ∝(∗)
x ≤ F and αx − α∗x = 0 (18)
x=0 x=1

Here, ∝(∗)
x stands in for αx and αx , whereas α
∗ (∗)
represents the vectors
3.4. Classification using ensemble deep Neural Support Vector Machine that include all αx values. The term “dual objective” refers to Equation
∗

(EDN-SVM) (17). “Bias constraint” refers to the second constraint in Equation (18).
Finding the α and α(∗) that maximises the dual goal is the first step in a
n this paper, we introduce the Neural Support Vector Machine linear regression SVM’s determination of its output.
(NSVM), a hybrid machine learning algorithm consisting of both neural
∑
h
( )
networks and SVMs. The output of the NSVM is given by support vector c(i) = α∗x − αx (Ix .I) + v (19)
machines that take a small central feature layer as their input. This x=1

feature layer is in turn the output of a number of neural networks,

A linear relationship between Ix and Ij is assumed in the provided
trained through backpropagation of the derivatives of the dual objec­
SVR model. Obviously, we need a nonlinear version of the SVR model.
tives of the SVMs with respect to the feature-node values. The NSVM
To achieve this, we may utilize a map to convert the training patterns Ix
aims to overcome the problems of the standard SVM. First, the NSVM
into a higher-dimensional feature space, and then run the data through
adds more layers to the SVM, making it “deeper”. Furthermore, the
the standard SVR method. However, this strategy may quickly become
neural networks can learn arbitrary features, making the kernel func­
computationally impossible. It is sufficient to know that g(Ix , I): =
tions much more flexible. Finally, by combining multiple SVMs with a
ψ (Ix ).ψ (I), in order to get the regression estimate and the dual goal,
shared feature layer into one learning architecture, the NSVM extends an
respectively. In order to make SVR nonlinear, this kernel function is
SVM’s generalization capability to multiple outputs.
often used. Several types of kernel functions, including as radial basis
We introduced a novel hybrid approach termed as Ensemble Deep
functions, polynomial functions, and sigmoidal functions, have exten­
Neural Support Vector Machine (EDN-SVM) for classification. This al­
sive use.
gorithm is comprised of both NNs and SVMs. SVMs accepted the
dimensionally reduced central features to provide the output of the
3.4.2. Ensemble deep Neural Support Vector Machine (EDN-SVM)
EDN-SVM. To train the EDN-SVM, significant feature layer is utilised. As
There are four parts to the EDN-SVM, as shown in Fig. 7: (1) an input
a result of being trained by backpropagation of the derivatives of the
layer with S nodes, (2) a central feature layer with s nodes, (3) a sum of s
dual goals of the SVMs with regard to the values of feature-node, a
two-layer NNs (MLPs) M each one capture the full input layer as input
number of NNs produce this feature layer as their output. The short­
and produce one of the attribute values, and (4) a main support vector
comings of the conventional SVM are targeted for improvement by the
regression model N that takes the entire feature layer as input and de­
EDN-SVM. To begin, the EDN-SVM is able to “deepen” the SVM by
termines the output node value. After receiving a pattern x in dimension
adding more layers to it. In addition, the NN is able to learn completely
S, the EDN-SVM sends it on to the NNs, where it is transformed into a
arbitrary characteristics, which makes the kernel functions far more
vector that becomes the basis for the feature layer’s values. As a short­
adaptable. In conclusion, the EDN-SVM is able to expand the general­
hand for the mapping that the NNs produce, we use φ(I|θ), and we
ization power of an SVM to multiple outputs. This is accomplished by
represent the NNs’ weights as a vector. The value of the output node is
merging several SVMs with a shared feature layer into a single learning
calculated using the support vector machine M, which takes as input the
architecture.
representation in the feature layer. In order to arrive at its conclusion,
the regression EDN-SVM uses:
3.4.1. Support vector regression (SVR)
The training for linear ξ insensitive SVR consists of finding the ∑
h
( )
optimal solution to the following restricted optimization problem: C(i) = α∗x − αx g(φ(Ix |θ), φ(I|θ)) + v (20)
x=1
[ ⃦ )]
min 1⃦
⃦ 2 ∑h
( This is the primary SVM kernel function, which is denoted by g(‘; ’).
⃦e‖ + F ∗
ξi + ξx (15)
e, ξ(∗) , v 2⃦ x=1 As shown in Fig. 7, the EDN-SVM estimator has the following architec­
tural layout. Three separate neural networks, each extracting a different
conditional on the following: feature, make up the feature layer in this example.
jx − e.Ix − v ≤ ∈ +ξx , e.Ix + v − jx ≤∈ +ξ∗x , and ξx , ξ∗x ≥ 0 (16)
3.4.3. Modified objectives
In this equation, e denotes the weight vector, v means the bias value, (Ix , Finding an acceptable value for the variable c requires the system to
jx ) refers to the training sample and its respective target value, ξx and ξ∗x learn a representation of the input MRI brain data in ‘a’ that codes the
are the “slack variables” that enable the model to accommodate properties that are most relevant for computing the desired output. The

6
S. Anantharajan et al. Measurement: Sensors 31 (2024) 101026

Fig. 7. Architecture of END-SVM estimator.

fundamental goal of SVR is modified by the EDN-SVM, which does this conditional on the following:
by exchanging the training samples Ix for their representation in the ∑h ( )
feature layer. As a consequence of this, the weight vector of the NNs θ is 0 ≤ ∝i , ∝∗x ≤ F, and x=1
α∗x − αx = 0 (23)
included as an extra primary variable. This results in the following
primary goal for an EDN-SVM when using a linear SVR model N: 3.4.4. Training procedure
[ ⃦ ] The training of the EDN-SVM estimator has two primary objectives:
min 1⃦ ∑h
( )
⃦ 2
⃦e‖ + F ξi + ξ∗x (21) (1) Determine the value of α(∗) that maximises Equations (22), and (16)
e, v, θ 2⃦ x=1 Determine the values of the weights of the NNs θ in such a way that each
The new ‘dual objective’ for the EDN-SVM when the SVR system N network Mα contributes to the Ix that minimises Equation (22). Both of
utilises a kernel function K(‘; ’) is, in accordance, as follows: these objectives cannot be accomplished in isolation; rather, in order to
adjust, it is necessary to also alter the NNs, and vice versa. Whenever the
min max (∗)
∑h
( ∗ ) ∑ h
( ∗ ) model is shown a training pattern Ix , the END-SVM employs a gradient
(∗) E(∝ , θ) = − ε αx + αx + αx − αx jx
ascent method to make adjustments to each αx and ∝x in the direction
θ ∝ (∗)
x=1 x=1
of a local maximum of E+ (.):
1 ∑ h
( ∗ ) ( ( ⃒ ))
− α − αx (α∗y − αy G φ(Ix |θ), φ Iy ⃒θ (22)
2 x,y=1 x

Fig. 8. Tumour identification after EDN-SVM approach.

7
S. Anantharajan et al. Measurement: Sensors 31 (2024) 101026

( )
∂e + (.) h (
∑ ) 2 We maintain all α(∗) coefficients between 0 and F. It is our goal to
∝(∗) modify the weights of each NN Mα in such a way that its output results in
(∗)
x ← ∝x + λ with E+ (.) = E(.) − T1 ∝∗y − ∝y
∂∝(∗)
the lowest possible value of E+ (.). We will refer to the output of Mα given
x y=1

− T2 ∝∗x ∝x (24) training pattern Ix as a azx from now on. After that, we will attempt to
reduce azx by making use of gradient descent.
where λ is a metaparameter that controls the learning rate, and T1 and T2
are parameters that are decided upon in advance. With the use of the ∂E+ (.) ( ∗ )∑h ( ) ∂G( φ(I |θ), φ( I ⃒⃒θ))
(27)
∗ x y
= − ∝ − ∝ x ∝ − ∝ y
derivative of E+ (.) in Equation (24), we can make sure that ∝x is
x y
(∗) ∂azx y=1
∂azx
modified such that the bias requirement holds true and that the pair
Given an input pattern Ix , we may use backpropagation to reduce the
(∝x , ∝x ) has at least one value equal to zero. The following are the two
(∗)
error of Mα by adjusting its weights according to Equation (27). Algo­
expressions that represent the derivatives of our gradient ascent rithm 1 provides the whole training procedure. The results demonstrate
algorithm: that training the neural networks and the primary SVM are interleaved
for a certain number of epochs. The average error is used to learn the
SVR model N’s bias value.
Algorithm 1. EDN-SVM

)
∑h ( ) ( ∑h ( )
∂E+ (.) ( ⃒ )
= − ε +jx − ∝∗y − ∝y G φ(Ix |θ),φ Iy ⃒θ − 2T1 ∝∗y − ∝y The END-NSVM has a computational time and space complexity that
∂∝∗x y=1 y=1
scales linearly with the number of inputs and also scales linearly with
− T2 ∝∗x the number of feature-extracting neural networks used. In the same way
(25) as SVMs do, its complexity is proportional to the amount of instances
used for training.
and

∑h ( ) (
)
∑h ( ) 4. Results and discussion
∂E+ (.) ( ⃒ )
= − ε − jx + ∝∗y − ∝y G φ(Ix |θ),φ Iy ⃒θ +2T1 ∝∗y − ∝y
∂∝∗x y=1 y=1 In this part, we will explore the identification of brain tumors by MRI
− T2 ∝∗x utilizing both deep learning and machine learning techniques. Here we
(26) used the Python 3.7.16 version for implementation. The four basic
matrices that are used in performance prediction are referred to as “True
Positive (tp )", “True Negative (tn )", “False Positive (fp ), and “False
Negative (fn )." In this context, cases are said to be true positives if the
tumour can be accurately anticipated. Instances that may have been
fairly anticipated to be negative, known as true negatives, are examined.
Instances of cases that were meant to be successfully predicted but
turned out to be inaccurate are examples of false positives. False nega­
tives are situations that are meant to be mistakenly detected but are, in
reality, it is properly predicted one. In addition, the accomplishments of
the suggested methodology are compared with the accomplishments of
existing methods such as “Convolutional neural network (CNN)",
“Random Forest Classifier (RFC)", “Artificial neural network (ANN)",
and " Region-based Convolutional Neural Networks (R–CNN)" in terms
of “Accuracy”, “Bit error rate”, “Computational time”, “Peak Signal
Ratio”, “Jaccard coefficient”, “Sensitivity” and “Specificity”. MRI brain
tumour images may be predicted with the use of the EDN-SVM classifier
model that was described before, as shown in Fig. 8.

4.1. Accuracy

Fig. 9. Accuracy comparison. Accuracy is measured by comparing the total number of MRI brain

8
S. Anantharajan et al. Measurement: Sensors 31 (2024) 101026

Fig. 12. Psnr comparison.

Fig. 10. Computational time comparison.

4.3. Jaccard coefficient(JC)

images to the number of total images and calculating the percentage of
healthy and tumorous brain tissue that can be reliably predicted. The The Jaccard co-efficient is a metric that determines the proportion of
greatest consideration has been given to determining the optimal value similarities that exist between manually segmented ground truth images
for the evaluation parameter of the estimating classifier. The accuracy of and the segmented output. If image A is the result of applying any
the recommended approach is compared to that of the standard methods segmentation method to it, and image B is the image that represents the
in Fig. 9. By computing the proportion of healthy and tumorous brain ground truth, then the Jaccard coefficient may be written as
tissue, the recommended technique for classifying MRI brain pictures
has a higher degree of accuracy than the traditional approaches, as is Jaccard coefficient =
A∩B
(29)
clearly evident in the figure. The proposed method has a high accuracy A∪B
of 97.93 % compared with the existing methods. Fig. 11 depicts the comparison of JC of the suggested method with
tp + tn the conventional methods. Comparing the recommended way of MRI
Accuracy = (28) brain images to more conventional approaches, the figure clearly dem­
tp + tn + fp + fn
onstrates that the suggested method has a higher JC.
4.2. Computational time
4.4. Peal signal to noise ration (PSNR)
The amount of time needed to carry out a computing operation is
referred to as the “computation time,” although it is also often termed PSNR is useful for checking the noise level in an image in order to
the “running time.” Fig. 10 displays a comparison of the amount of time identify any decline in image quality that may have occurred. Fig. 12
required to compute using the recommended technique with the time depicts the PSNR comparison of the suggested method with the con­
required by the conventional methods. In contrast to conventional ap­ ventional methods. It shows the proposed method has a higher PSNR
proaches for categorising MRI images, the recommended method EDN- ratio when compared to the traditional methods. The present study’s
SVM computes the detection in less time, as seen by the figure. suggested strategy outperforms existing techniques with a PSNR score of
52.98 %.

Fig. 11. Jaccard coefficient. Fig. 13. Sensitivity comparison.

9
S. Anantharajan et al. Measurement: Sensors 31 (2024) 101026

large number of trees. While R–CNN cannot be done in real time since it
takes around 47 s for each test image.

6. Conclusion

For the purpose of detecting brain tumors, the suggested EDN-SVM

method proposes a novel method of image classification by establish­
ing a direct link between all levels and ensuring that data is freely shared
among them. Extensive simulations are conducted to test the effective­
ness and viability of the suggested model. It has correctly identified the
tumour image with a 97.93 % accuracy rate. Several benefits are shown
by the data, suggesting that this model combination is worthwhile. To
begin, the model automatically extracted the salient characteristics,
making feature extraction far less time-consuming and arduous than it
would have been with more conventional classifiers. Second, the sug­
gested EDN-SVM model included the best features of deep NN and SVM,
the two most well-known and widely-used classifiers for image recog­
nition and classification. The following are the directions that the work
going forward should take: (a) The algorithms that were created need to
be included in the software that doctors use, and (b) the methodologies
Fig. 14. Specificity comparison. and procedures that were proposed in this research can only be used to
grayscale photographs. Color images may be utilised to study the same
( )
2552 difficulties and also work with 3D brain scans to obtain more effective
PSNR = 10 log (30)
MSE brain tumour segmentation in future research.

4.5. Sensitivity Ethical compliance

The term “sensitivity” refers to the likelihood of a positive test on the No.
assumption that it is positive. Another term for this is “true positive
rate.” The sensitivity comparison between the proposed approach and Declaration of competing interest
the conventional methods is shown in Fig. 13. It reveals that when
compared to the sensitivity of the earlier approaches, the sensitivity of The authors declare that they have no known competing financial
MRI images to classifying by the recommended approach EDN-SVM is interests or personal relationships that could have appeared to influence
superior. As compared to the current approaches, the suggested method the work reported in this paper.
has a high sensitivity of 92 %.
TP Data availability
Sensitivity = × 100 (31)
(TP + FN)
No data was used for the research described in the article.

4.6. Specificity
Acknowledgments
Specificity, also known as the true negative rate, is the likelihood of a
There is no funding
negative test result under the assumption that the result is in fact
negative. A comparison of the specificity of the suggested approach with
that of the conventional methods is shown in Fig. 14. The classification References
of MRI images reveals that the suggested method of EDN-SVM has a high
[1] E. Schulz, S.J. Gershman, The algorithmic architecture of exploration in the human
degree of specificity when compared to the existing techniques. Evalu­ brain, Curr. Opin. Neurobiol. 55 (2019) 7–14.
ating the suggested method to the existing systems, it has a high speci­ [2] P.J. van Lonkhuizen, K.M. Klaver, J.S. Wefel, M.M. Sitskoorn, S.B. Schagen,
ficity of 98 %. K. Gehring, Interventions for cognitive problems in adults with brain cancer: a
narrative review, Eur. J. Cancer Care 28 (3) (2019) e13088.
TN [3] A. Del Dosso, J.P. Urenda, T. Nguyen, G. Quadrato, Upgrading the physiological
Specificity = × 100 (32) relevance of human brain organoids, Neuron 107 (6) (2020) 1014–1028.
TN + FP [4] S.L. Fernandes, U.J. Tanik, V. Rajinikanth, K.A. Karthik, A reliable framework for
accurate brain image examination and treatment planning based on early diagnosis
5. Discussion support for clinicians, Neural Comput. Appl. 32 (20) (2020) 15897–15908.
[5] Z.U. Rehman, S.S. Naqvi, T.M. Khan, M.A. Khan, T. Bashir, Fully automated multi-
parametric brain tumour segmentation using superpixel based classification,
The comparison of the suggested method to the existing models is Expert Syst. Appl. 118 (2019) 598–613.
shown in Figs. 9–14. In order to conduct this inquiry, conventional [6] Z.U. Rehman, M.S. Zia, G.R. Bojja, M. Yaqub, F. Jinchao, K. Arshid, Texture based
localization of a brain tumor from MR-images by using a machine learning
methods such as CNN [22], RFC [23], ANN [24], and R–CNN [25] were approach, Med. Hypotheses 141 (2020) 109705.
used. The constraints of the approach that has previously been employed [7] C. Kv, G.G. King, Brain tumour classification: a comprehensive systematic review
make the proposed strategy superior in terms of performance. The dis­ on various constraints, in: Computer Methods in Biomechanics and Biomedical
Engineering: Imaging & Visualization, 2022, pp. 1–13.
advantages of the present approaches include the ones that are listed
[8] K. Rezaei, H. Agahi, A. Mahmoodzadeh, Multi-objective differential evolution-
below. The ANN is a model of hardware dependency, and it exhibits based ensemble method for brain tumour diagnosis, IET Image Process. 13 (9)
behaviour that cannot be described. In order to categorise tumour im­ (2019) 1421–1430.
ages, CNN requires a vast amount of training data. The fundamental [9] R. Ezhilarasi, P. Varalakshmi, Tumor detection in the brain using faster R-CNN. In
2018 2nd international conference on I-smac (IoT in social, mobile, analytics and
drawback of RFC is that it may be rendered useless and unreasonably cloud)(I-smac) I-smac (IoT in social, mobile, analytics and cloud)(I-smac), 2018, in:
sluggish for use in real-time prediction when used with an excessively 2nd International Conference on, IEEE, 2018, August, pp. 388–392.

10
S. Anantharajan et al. Measurement: Sensors 31 (2024) 101026

[10] H. Mohsen, E.S.A. El-Dahshan, E.S.M. El-Horbaty, A.B.M. Salem, Classification [19] C. Sheela, G. Suganthi, Brain tumor segmentation with radius contraction and
using deep learning neural networks for brain tumors, Future Computing and expansion based initial contour detection for active contour model, Multimed.
Informatics Journal 3 (1) (2018) 68–71. Tool. Appl. 79 (33) (2020) 23793–23819.
[11] M. Siar, M. Teshnehlab, Brain tumor detection using deep neural network and [20] S.J. Nanda, I. Gulati, R. Chauhan, R. Modi, U. Dhaked, A K-means-galactic swarm
machine learning algorithm, in: 2019 9th International Conference on Computer optimization-based clustering algorithm with Otsu’s entropy for brain tumor
and Knowledge Engineering (ICCKE), IEEE, 2019, October, pp. 363–368. detection, Appl. Artif. Intell. 33 (2) (2019) 152–170.
[12] C.L. Choudhury, C. Mahanty, R. Kumar, B.K. Mishra, Brain tumor detection and [21] A. Wadhwa, A. Bhardwaj, V.S. Verma, A review on brain tumor segmentation of
classification using convolutional neural network and deep neural network, in: MRI images, Magn. Reson. Imag. 61 (2019) 247–259.
2020 International Conference on Computer Science, Engineering and Applications [22] M.L. Martini, E.K. Oermann, Intraoperative brain tumour identification with deep
(ICCSEA), IEEE, 2020, March, pp. 1–4. learning, Nat. Rev. Clin. Oncol. 17 (4) (2020) 200–201.
[13] M.A. Naser, M.J. Deen, Brain tumor segmentation and grading of lower-grade [23] M. Soltaninejad, G. Yang, T. Lambrou, N. Allinson, T.L. Jones, T.R. Barrick, F.
glioma using deep learning in MRI images, Comput. Biol. Med. 121 (2020) 103758. A. Howe, X. Ye, Supervised learning based multimodal MRI brain tumour
[14] M.K. Islam, M.S. Ali, M.S. Miah, M.M. Rahman, M.S. Alam, M.A. Hossain, Brain segmentation using texture features from supervoxels, Comput. Methods Progr.
tumor detection in MR image using superpixels, principal component analysis and Biomed. 157 (2018) 69–84.
template based K-means clustering algorithm, Machine Learning with Applications [24] N. Arunkumar, M.A. Mohammed, S.A. Mostafa, D.A. Ibrahim, J.J. Rodrigues, V.H.
5 (2021) 100044. C. de Albuquerque, Fully automatic model-based segmentation and classification
[15] T.A. Jemimma, Y.J. Vetharaj, Watershed algorithm based DAPP features for brain approach for MRI brain tumor using artificial neural networks, Concurrency
tumor segmentation and classification, in: 2018 International Conference on Smart Comput. Pract. Ex. 32 (1) (2020) e4962.
Systems and Inventive Technology (ICSSIT), IEEE, 2018, December, pp. 155–158. [25] K. Salçin, Detection and classification of brain tumours from MRI images using
[16] G. Hemanth, M. Janardhan, L. Sujihelen, Design and implementing brain tumor faster R-CNN, Tehnički glasnik 13 (4) (2019) 337–342.
detection using machine learning approach, in: 2019 3rd International Conference [26] M. Nazir, S. Shakil, K. Khurshid, Role of deep learning in brain tumor detection and
on Trends in Electronics and Informatics (ICOEI), IEEE, 2019, April, classification (2015 to 2020): a review, Comput. Med. Imag. Graph. 91 (2021)
pp. 1289–1294. 101940.
[17] S.K. Chandra, M.K. Bajpai, Effective algorithm for benign brain tumor detection [27] J. Amin, M. Sharif, M. Yasmin, S.L. Fernandes, A distinctive approach in brain
using fractional calculus, in: TENCON 2018-2018 IEEE Region 10 Conference, tumor detection and classification using MRI, Pattern Recogn. Lett. 139 (2020)
IEEE, 2018, October, pp. 2408–2413. 118–127.
[18] M. Gurbină, M. Lascu, D. Lascu, Tumor detection and classification of MRI brain [28] H.A. Khalil, S. Darwish, Y.M. Ibrahim, O.F. Hassan, 3D-MRI brain tumor detection
image using different wavelet transforms and support vector machines, in: 2019 model using modified version of level set segmentation based on dragonfly
42nd International Conference on Telecommunications and Signal Processing algorithm, Symmetry 12 (8) (2020) 1256.
(TSP), IEEE, 2019, July, pp. 505–508.

11