0% found this document useful (0 votes)
56 views20 pages

Optimized Brain Tumor Detection A Dual-Module Approach For MRI Image Enhancement and Tumor Classification

Uploaded by

darshank11012
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
56 views20 pages

Optimized Brain Tumor Detection A Dual-Module Approach For MRI Image Enhancement and Tumor Classification

Uploaded by

darshank11012
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 20

Received 6 February 2024, accepted 11 March 2024, date of publication 20 March 2024, date of current version 26 March 2024.

Digital Object Identifier 10.1109/ACCESS.2024.3379136

Optimized Brain Tumor Detection:


A Dual-Module Approach for MRI
Image Enhancement and Tumor
Classification
ABDULLAH A. ASIRI 1 , TOUFIQUE AHMED SOOMRO 2 , (Senior Member, IEEE),
AHMED ALI SHAH 3 , (Senior Member, IEEE), GANNA POGREBNA4,5,6 ,
MUHAMMAD IRFAN 7 , AND SAEED ALQAHTANI 1
1 Radiological Sciences Department, College of Applied Medical Sciences, Najran University, Najran 61441, Saudi Arabia
2 Department of Electronic Engineering, The University of Larkano (UOL), Larkana 77111, Pakistan
3 Electrical Engineering Department, Sukkur IBA University, Sukkur 65200, Pakistan
4 Artificial Intelligence and Cyber Futures Institute, Charles Sturt University, Bathurst, NSW 2795, Australia
5 The University of Sydney Business School, The University of Sydney, Darlington, NSW 2006, Australia
6 The Alan Turing Institute, NW1 2DB London, U.K.
7 Electrical Engineering Department, College of Engineering, Najran University, Najran 61441, Saudi Arabia

Corresponding author: Toufique Ahmed Soomro ([email protected])


This work was supported by the Deanship of Scientific Research, Najran University, Kingdom of Saudi Arabia, for funding this work
through the National Research Priorities and Najran Area Funding Program, under Grant NU/NRP/MRC/12/26.

ABSTRACT Neurological and brain-related cancers are one of the main causes of death worldwide.
A commonly used tool in diagnosing these conditions is Magnetic Resonance Imaging (MRI), yet the manual
evaluation of MRI images by medical experts presents difficulties due to time constraints and variability.
This research introduces a novel, two-module computerized method aimed at increasing the speed and
accuracy of brain tumor detection. The first module, termed the Image Enhancement Technique, utilizes a trio
of machine learning and imaging strategies—adaptive Wiener filtering, neural networks, and independent
component analysis—to normalize images and combat issues such as noise and varying low region contrast.
The second module uses Support Vector Machines to validate the output of the first module and perform
tumor segmentation and classification. Applied to various types of brain tumors, including meningiomas and
pituitary tumors, our method exhibited significant improvements in contrast and classification efficiency.
It achieved an average sensitivity and specificity of 0.991, accuracy of 0.989, and a Dice score (DSC) of
0.981. Furthermore, the processing time of our method, averaging at 0.43 seconds, was markedly lower
compared to existing methods. These results underscore the superior performance of our approach over
current state-of-the-art methods in terms of sensitivity, specificity, precision, and DSC. Future enhancements
will seek to increase the robustness of the tumor classification method by employing a standardized approach
across a suite of classifiers.

INDEX TERMS Magnetic resonance imaging (MRI), image enhancement technique, brain tumor
segmentation, neural networks, brain tumor classification.

I. INTRODUCTION There are approximately 200 different types of brain tumors


Brain abnormalities, commonly referred to as tumors in that can occur in various regions of the human brain. These
medical terminology, are classified as malignant or benign. tumors can have a significant and often life-changing impact
on individuals’ lives. Numerous studies provide strong
The associate editor coordinating the review of this manuscript and scientific evidence of increasing brain tumor incidence and
approving it for publication was Yi Zhang . its association with human mortality [1]. According to the
2024 The Authors. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
42868 VOLUME 12, 2024
A. A. Asiri et al.: Optimized Brain Tumor Detection: A Dual-Module Approach

American Cancer Society, brain tumor is one of the serious networks to normalize the image contrast. Then, we used an
diseases in which the brain tissues are increased irregularly independent component analysis to even out the contrast and
and affects the brain function. The research was conducted obtain a well-contrasted image. This enhancement technique
by the National Brain Tumor Foundation, and they reported is validated on the CE-MRI database and its impact on
that the number of people who have lost their lives to brain the classification of brain tumors has also been observed.
tumors has increased by 300% over the past three decades. The novelty of the proposed method lies in its innovative
Treatment is necessary, otherwise brain tumors could lead to application of contrast enhancement techniques, including
death if left untreated [2]. The physical appearance of the the use of the adaptive Wiener filter, neural networks,
brain and the complexity of brain tumors pose challenges and independent component analysis,it aims to produce
for health care to diagnose and recommend early treatment clearer, more consistent, and well-contrasted MRI images,
to the patient. Early detection of brain tumors and proper thereby improving the accuracy of tumor segmentation and
treatment play an important role in improving the survival classification. The proposed method contributes to the field
rate of patients. Because biopsy of a brain tumor is not such of brain MRI image analysis in several significant ways:
a simple task as biopsy of other parts of the body, as it must 1) Contrast Enhancement Techniques: The method
necessarily be performed with surgical intervention. SO. This introduces innovative implementations of contrast
requires the most effective methods to diagnose brain tumor enhancement techniques, including the use of the
without surgery and many imaging modalities are used, but adaptive Wiener filter, neural networks, and indepen-
magnetic resonance imaging (MRI) is the best option and it dent component analysis, to address the challenge of
is commonly used to diagnose brain tumors [3], [4]. low-contrast MRI images.
Brain MRI imaging utilizes a range of techniques to 2) Validation on CE-MRI Image Database: The
acquire data, which is subsequently processed to form input method’s effectiveness is validated on the CE-MRI
vectors for classification purposes, and MRI imaging filters image database, providing evidence of its practical
are used to identify brain tumors and help the radiologist rec- application and impact on the accurate classification
ommend treatment. The radiologist has two types of choices of brain tumors.
for identifying brain tumors. First, they distinguish between 3) Innovation in MRI Enhancement: The proposed
typical or normal brain MRI images, and second, they classify method’s novelty lies in its innovative application
brain MRI images into different brain tumor types. Brain MRI of contrast enhancement techniques, offering a fresh
images are the most important symptomatic neuroimaging approach to addressing the specific challenges of MRI
testing device that identifies baseline variations from norms images.
in the brain. This is particularly crucial for the diagnosis and 4) Improved Accuracy: By producing clearer, more con-
monitoring of brain tumors, as MRI scans can accurately sistent, and well-contrasted MRI images, the method
detect and characterize tumors, assess their growth over time, ultimately aims to enhance the accuracy of tumor
and aid in treatment planning [5], [6]. Brain MRI images segmentation and classification, contributing to more
are used to diagnose brain abnormalities, including those that reliable clinical diagnoses.
may indicate the presence of a brain tumor [9]. MRI provides The proposed method’s contributions include addressing the
detailed images of the brain and is one of the most common challenges of low-contrast MRI images, noise suppression,
tests used to identify brain tumors. MR images provide the resolution enhancement, coherence improvement, and valida-
soft tissues of the human body, and these images are mainly tion on a relevant database. These contributions collectively
used to detect brain structure and function [7], [8]. In this improve the accuracy of brain tumor segmentation and
research work, we have proposed enhancement method to classification, which is valuable in clinical and medical
solve the contrast problem of brain MRI images, and this imaging contexts.
leads to give accurate identification of brain tumor.
The main objective of this research work is to understand II. RELATED WORK
brain MRI images and to solve the problem of low contrast The early detection of brain tumors is crucial for effec-
variation of these images which leads to segmenting the tive treatment, and various methods have been explored
tumor and classifying normal and abnormal images. Brain by researchers [10] to achieve this goal. Selvapandian
MRI image analysis is challenging due to issues like and Manivannan [11] used the non-subsampled contourlet
noise, resolution, contrast, motion artifacts, variability, and transform (NSCT) to enhance brain images and extract
complex brain structures. Our proposed method is based on surface features, followed by adaptive neural techniques for
new implementations of contrast enhancement techniques classification and glioma brain tumor area segmentation.
and its impact on segmentation and classification. Image While their approach addressed important aspects of the
enhancement techniques solve the three problems of brain problem, specific performance metrics for their methodology
MRI images, and the first problem is noise suppression, and were not provided, leaving room for further evaluation.
the second problem is the resolution of low contrast and In a different approach, Sharif et al. [12] introduced
varying contrast and the last step is based on coherence of the a technique for skull section removal from brain images
image contrast. We used the adaptive Wiener filter with neural using brain surface extraction (BSE) and particle swarm

VOLUME 12, 2024 42869


A. A. Asiri et al.: Optimized Brain Tumor Detection: A Dual-Module Approach

optimization (PSO) for segmentation. This method achieved overfitting and boost the sample size, with VGG16 achieving
an impressive maximum accuracy of 99% when evaluated on an outstanding accuracy of 98.69%, the highest among the
complex brain datasets, indicating its potential for accurate compared methods. Another study [21] used a two-stage
brain tumor detection. However, it’s important to note that strategy to distinguish healthy brains from those with tumors,
this method may not have addressed other types of brain employing CNNs for preprocessing and Error-Correcting
tumors or potential performance issues in different scenarios. Output Codes SVM (ECOCSVM) for classification in the
Kumar et al. [13] proposed the Weighted Correlation Feature second stage. Impressively, AlexNet achieved the highest
Selection Based Iterative Bayesian Multivariate Deep Neu- accuracy at 99.55%. The evaluation was conducted using the
ral Learning (WCFS-IBMDNL) method, which combines BraTS and RIDER databases.
feature selection with Iterative Bayesian Multivariate Deep Abiwinanda [22] utilized a CNN to detect common brain
Neural Network (IBMDNN) classification to identify brain tumors, achieving high training and validation accuracy rates
tumors. While this method offers an integrated approach, of 98.51% and 84.19%, respectively, with a dataset of 3064 T-
it exhibited high False Alarm Rates (FARs), suggesting that 1 weighted CE-MRI images. Pashaei et al. [23] proposed
there is room for improvement in terms of accuracy and a CNN-based model for identifying meningioma, glioma,
reducing false positives. and pituitary tumors, achieving an accuracy of 93.68%
Ozyurt et al. [14] took a mixed approach, combining and comparing their method with other models using ten-
eutrosophy and Convolutional Neural Network (CNN) with fold cross-validation. In Gumaei et al. [24] study, involving
Neutrosophy to categorize malignant and benign tumor preprocessing, feature extraction, and classification, they
regions in brain images. Their method achieved an impressive achieved an accuracy of 94.23% with Principle Component
average success rate of 95.62% through the use of CNN Analysis Normalized GIST (PCA-NGIST) and Regularized
features and SVM for classification. However, the choice Extreme learning machine (RELM), though they did not
of CNN features may not be universally optimal for all conduct a comparative analysis with previously reported
types of brain tumors, and further exploration of feature techniques.
selection could enhance the method’s versatility. In another Phaye et al. [25] introduced DCNet++ and DCNet as
study, Raju et al. [15] utilized a computerized method techniques for brain tumor classification, attaining accuracy
based on a multi-class Support Vector Neural Network values of 95.03% and 93.04%, respectively, using MRI
(SVNN) trained with the Harmony-Crow Search (HCS) images. However, a drawback was the absence of a dataset
optimization technique for tumor identification in MRI representing healthy individuals for classification, limiting
images. However, this method solely relied on the BRATS the context of the results. Badža and Barjaktarovic [26]
dataset for evaluation, potentially limiting its generalizability presented a CNN-based method with 22 layers designed
to a broader range of cases. Researchers and practitioners for the classification of pituitary, meningioma, and glioma
should carefully consider the specific requirements of their tumors, achieving an impressive accuracy of 96.56% through
brain tumor detection tasks when choosing among these k-fold cross-validation. Sharma et al. [27] adopted a hybrid
methods, considering factors such as the type of tumor, approach that combined Artificial Neural Networks (ANN)
available datasets, and the desired balance between accuracy and K-means clustering for tumor classification. Their
and computational complexity. method featured efficient tumor region identification through
Yin et al. [16] explored background elimination, fea- GLCM-based feature extraction and k-means algorithm
ture extraction, and classification phases using multilayer integration, which was a notable strength. Rani et al. [28]
perceptron networks, incorporating techniques like the employed SVM for brain tumor classification, complemented
Whale Optimization Algorithm and chaos theory for feature by Otsu thresholding. However, comparative studies revealed
selection and classification. However, their algorithm did the potential for accuracy improvement in their method.
not significantly enhance accuracy. In contrast, Krol and To meet the demand for enhanced brain tumor detection
Gimi [17] employed deep learning with a substantial MRI algorithms, Molina-Torres [29] employed a kernel SVM
image dataset, using dual neural networks (fully connected approach, specifically the Gaussian Radial Basis (GRB)
and CNNs) to classify brain images with tumors, achieving kernel, focusing on metrics such as specificity, precision, and
an impressive 91.43% accuracy via 5-fold cross-validation. accuracy, which provided valuable insights into algorithm
Similarly, Mittal et al. [18] presented an MRI-based brain performance.
tumor identification method that automated segmentation In conclusion of analysing the existing work, the body
with a Growing Convolution Neural Network (GCNN) of research reviewed in this section represents a significant
and Stationary Wavelet Transform (SWT). This approach advancement in the field of brain tumor detection and clas-
outperformed other methods, including CNN, SVM, and sification. These studies have explored various techniques,
KNN, in terms of accuracy [18], [19]. ranging from deep learning and CNNs to hybrid approaches
Rehman et al. [20] used VGGNet, GoogLeNet, and and SVMs, achieving promising results in terms of accuracy
AlexNet CNNs to distinguish between three types of and classification performance. However, it is important to
brain tumors. They employed data augmentation to prevent acknowledge the existing gaps and limitations, including

42870 VOLUME 12, 2024


A. A. Asiri et al.: Optimized Brain Tumor Detection: A Dual-Module Approach

the need for more comprehensive datasets, comparative Wiener filter on brain MRI imaging depends on the three
analyses, and improvements in the consistency of evaluation parameters, the local window of each pixel of brain MRI
metrics. These insights provide a solid foundation for the images and the noise variance [31]. This is additive noise
proposed methods and approaches that will be discussed and the adaptive Wiener filtering works on the local window
in the subsequent sections, aiming to further refine and around each pixel in the image to get a denoised brain MRI
enhance the state-of-the-art in brain tumor detection using image. The step-by-step representation of adaptive Wiener
MRI images. filtering is defined below.
1) The Adaptive Wiener filter operates on a local window
III. PROPOSED METHOD around each pixel (x,y) in the image. The filtered output
Brain MRI images are used in our proposed method as at (x,y) is computed using Equation 1:
it is a non-invasive process and it provides us with the
multidimensional analysis as compared to other modalities g (x, y) = w (x, y) ∗ [f (x, y) − m (x, y)] + m (x, y) .
such as CT scan and X-ray images. The main purpose (1)
of brain MRI images is to identify the brain tumor and
classify its precise region. Before performing segmentation where w (x, y) is the Wiener filter, m (x, y) is the local
and classifications of brain MRI images, it is necessary to mean of the input image in the window around (x,y),
improve image contrast as well as reduce noise. In this and ∗ denotes convolution.
research work, we propose image enhancement techniques as 2) The wiener filter is process the images based on the
well as the analysis of its impact on the segmentation of the Equation
abnormal region. The proposed methodology is illustrated in
[σf2 (x, y) − σv2 (x, y)]
the Figure 1 w(x, y) = . (2)
The main theme of the implementation of the proposed σf2 (x, y)
method to identify the abnormal region as a tumor from
brain MRI images. In the proposed technique as shown where σf2 (x, y) is the local variance of the input image
in Figure 1. It is a computerized method of brain tumor in the window around (x,y), and σv2 (x, y) is the variance
detection, and it depends on the noise-free trained image of the noise.
of brain MRI images and then it is processed to enhance The Wiener filter adjusts the strength of the filtering based
the image using independent component analysis. The image on the local signal-to-noise ratio (SNR) of the input image.
enhancement technique for brain MRI images is implemented When the SNR is high, the Wiener filter is close to 1, and
in this research work and its impact is validated on the the filtering is weak, preserving the image features. When the
post-processing steps that lead to obtaining the brain tumor. SNR is low, the Wiener filter is close to 0, and the filtering is
The different steps of image enhancement techniques for strong, reducing the noise. Overall, the Adaptive Wiener filter
brain tumors as well as brain tumor detection are elaborated can effectively enhance the quality of brain MRI imaging by
in the following sections. removing noise while preserving important image features as
shown in the Figure 3.
A. STEP 01: BRAIN MRI IMAGES PROCESSING
Processing imaging data is a crucial step, particularly in han- C. STEP 03: TRAINING OF BRAIN MRI IMAGING BY USING
dling medical images like brain MRI images. Analyzing these RBF NEURAL NETWORK FILTERING
images constitutes a key stage in computerized methods, We employed Radial Basis Function (RBF) neural network
especially in the context of brain image analysis. Our work filtering as a means to enhance the quality of medical
led to the development of the CE-MRI brain image database, images, particularly those of MRI. RBF neural networks
encompassing the axial, sagittal, and coronal planes. Each are a type of artificial neural network commonly adopted
image presents unique challenges and reveals distinct details. for image filtering, classification, and pattern recognition
As illustrated in Figure 2, the images displayed therein are tasks. In our study, however, this technique was leveraged
susceptible to noise interference. Consequently, there is a to enhance the quality of images and provide pre-classified
requirement for a dependable filter to effectively eliminate depictions of brain tumor anomalies [32]. The training of
this interfering factor. The process for noise suppression is RBF neural network filtering for brain MRI imaging involves
addressed in the subsequent step. the following steps:
1) The first step is data pre-processing and it is an
B. STEP 02: ADAPTIVE WIENER FILTERING TO REDUCE important step to pre-process brain MRI images to
NOISE remove artifacts as well as noise that can affect the
Adaptive Wiener filtering is the type of signal processing performance of the RBF neural network. This is
approach that can be used on medical images [30], and we achieved in the previous steps because we removed
have used Wiener filtering on brain MRI imaging to improve noise using adaptive Wiener filtering.
overall image quality by reducing the noise and retaining 2) The next step is an important step in the functionality
image detail. The mathematical representation of the adaptive of RBF neural networks which is feature extraction.

VOLUME 12, 2024 42871


A. A. Asiri et al.: Optimized Brain Tumor Detection: A Dual-Module Approach

FIGURE 1. The proposed model.

The RBF filtering process, as outlined above, is predicated


on two parameters. These involve the creation of the image
through minimizing its variation and the utilization of RBF
to derive a denoised image by reducing the error function on
its variation minimization. This methodology serves to lower
the noise level and deliver a desirable output image.

1) MINIMIZATION OF THE VARIATION


FIGURE 2. Brain MRI images of three planes namely axial plane as shown Variation minimization is mainly used to denoise images and
in Fig(b), sagittal plane as shown in Fig(b) and coronal plane as shown in is one of the most effective methods [33], [34]. In this paper,
Fig(c).
we have used the variation minimization method based on the
RBF neural network. The variation minimization variation
was introduced by Rudin and Osher [32], and their algorithm
is based on the equation 3, which is implemented in this
research work for brain MRI images, as well as RBF neural
network.
Z
f (u) = |D| + λ∥µO −µ∥2 ddxy. (3)

R
Here, |D| represents the variation model of the image

FIGURE 3. Adaptive wiener filtering output of Brain MRI images of three (retinal image or MRI brain image) µ. If the Rimage µ
planes axial plane, sagittal plane and coronal plane. is regular, then Equation 3 becomes simply |∇µ|dx.

Rudin et al. [32] assumed that noise that corrupts the image
Features are extracted based on the image intensity and can be distinguished from noiseless images based on the
Rq
spatial information of the data in this proposed method. size of total variation, which is defined as µ2x + µ2y dxdy,
3) The RBF neural network is trained based on brain MRI 
where  represents the image dimensions ux and uy with
images as well as feature extraction in previous steps. corresponding partial differentiation. The Euler-Lagrange
The training image gives corresponding outputs the Equation [33] is used to minimize Equation 3, and Equation 4
filtered brain MRI image. The training process is based is obtained.
on learning the mapping of input features to desired    
output images. ∂  ux + ∂ q
uy

4) After training the brain MRI images, the validation ∂x ∂y
q
µ2 + µ2x y µ2 + µ2 x y
of the RBF neural network is performed and the
validation gives the information about the validated
data according to the desired output. − λ |µ − µo | = 0. (4)

42872 VOLUME 12, 2024


A. A. Asiri et al.: Optimized Brain Tumor Detection: A Dual-Module Approach

represented by h. This function is mathematically represented


by Equation 8.
 
− ∥x − ci ∥
Ii (x) = exp . (8)
2σ 2 i
The center and spread width of the i-th node are denoted
by ci and σi , respectively. All parameters of the RBF neural
network are initialized before training. The weight can be
initialized with small random values or zero, and the spread
FIGURE 4. RBF Model of Proposed Method for MRI Brain Image. initial value can be selected as the average of the nearest
neighbour distances among the initial centers. After training,
The Lagrange multiplier, denoted by λ, is given in Equation 5. the RBF image can be represented by the equation given
below as Equation 9
(µo )x µx
Z q Z
1 1
λ= µx + µ y −
2 2 µ = N (µo , hi , ci ) . (9)
2σ 2 2σ 2
q
µx 2 + µ2 y
The noisy image is denoted by µ. The gray level output of
(µo )y µy each pixel can be calculated using Equation 10 as µO (x, y).
Z
1
+ . (5)
2σ 2
q
µx 2 + µ2 y Xh 
−∥Mo (x, y) − c∥i
!
N (x, y) = hi exp . (10)
As mentioned earlier, the denoised image is obtained by 2σ 2 i
i=1
minimizing the error function, which is represented by
Mo (x, y) represents the grey level values of the neighbor-
Equation 6.
ing pixels corresponding to the pixel values of µO (x, y). The
    window size of the pixels can be 3 by 3, 5 by 5, etc. and it is
∂  µx ∂ µy illustrated in Figure 4.
E (x, y) = + q 
∂x ∂x
q
µx 2 + µy 2 µx 2 + µy 2 3) TRAINING OF RBF NEURAL NETWORK IMAGE
− λ |µ − µo | . (6) The aim is to minimize the error during the training of
the image Mo (x, y) using the RBF neural network. The
Equation 6 represents the main task of obtaining an image network is trained for a maximum number of iterations until
where µ is locally constant, with ux and uy being the the error is less than the convergence error threshold. The
dimensions of the image. qTo achieve this, it is assumed that parameters for the RBF neural network are determined using
ε > 0 and the expression µ2 x + µ2 y + ε is used. Equation 11, while the training parameters are dependent on
Equation 12.
2) PROPOSED NOISE REDUCTION MODEL USING RBF
NEURAL NETWORK
hi (s + 1) = hi (s) − η1 1hi (s). (11)
In addition, the denoising of color retinal fundus images and ci (s + 1) = ci (s) − η2 1ci (s). (12)
MRI brain images is achieved by minimizing Equation 6 The RBF neural network parameters, such as weights
using a proposed RBF neural network. The RBF neural and centers, are updated at each iteration using the update
network is composed of three layers: the input layer, the RBF rules given in Equation 13, where hi (s + 1) and ci (s + 1)
layer, and the output layer. The RBF layer is made up of unity represent the updated values of the parameters at the next
value scalar weights and the input vector, with the entire input iteration, while hi (s) and ci (s) are the current values, and
vector being fed to each neuron in the RBF network. The η1 and η2 are positive learning rates. The update is performed
output layer consists of a vector of m outputs that are linearly by minimizing the error function of Equation 6 to obtain the
combined to form the output image. The RBF model for parameter variations 1hi and 1ci
MRI brain images is depicted in Figure 4 and mathematically X ∂E (x, y)
represented in Equation 22. 1hi = . (13)
x,y
∂hi
h
X ∂E (x, y)
X
y = f (I ) = hi Ii (x) . (7)
1ci = . (14)
i=1
x,y
∂ci
The output image, denoted by f (I ), is determined by the ∂E (x, y)
radial basis function I (x) of the i-th hidden node and the ∂hi
corresponding hidden-to-output weight hi . The RBF function " # " #
is based on the distance between the input vector and a pre- ∂ ux ∂ uy
∂x ∂y
q q
defined vector, which is explained in the previous section and µ2 x +µ2 y µ2 x +µ2 y
= +
depicted in Figure 4. The total number of hidden nodes is ∂hi ∂hi
VOLUME 12, 2024 42873
A. A. Asiri et al.: Optimized Brain Tumor Detection: A Dual-Module Approach

vector S are random vectors with zero mean and unit


variance. The unknown matrix A is the mixing matrix of the
independent components, assumed to be in a square pattern.
Additionally, the ICA model assumes that the independent
component vector S has a non-Gaussian distribution. The
estimation of non-Gaussianity is a crucial parameter in the
ICA model, as it is required for the calculation of independent
components. The analysis is based on the random vector X ,
and the main goal is to estimate the unknown components
A and S using X . If the unknown mixing matrix A can be
estimated, its inverse can be computed as W , which enables
the calculation of the independent component vector S. This
FIGURE 5. Illustration of training of rbf neural network for brain MRI
images. process is expressed as follows:

S = A−1 X = WX . (18)
∂λ ∂µ
− (µ − µo ) − λ . (15)
∂hi ∂hi To estimate W , the FASTICA algorithm [35] can be
∂E (x, y) utilized, which initializes random vectors and utilizes a
∂ci fixed point algorithm to obtain a single independent compo-
" # " # nent. For multiple independent components, the FASTICA
∂ ux ∂ uy algorithm must be iterated n times with weight vectors
∂x ∂y
q q
µ2 x +µ2 y µ2 x +µ2 y w1 , w2 , w3 . . . . . . . . . wn . However, there is a high probability
= +
∂ci ∂ci of correlation between the vector values, which may result
∂λ ∂µ in different random vectors converging at the same maxima.
− (µ − µo ) − λ . (16) To prevent this, it is crucial to de-correlate or orthogonalize
∂ci ∂ci
the weight vector outputs wT1 X , wT2 X , wT3 X . . . ..wTn X using
The process of training is essential to achieve the desired
the Gram-Schmidt-Orthogonalization method [36] after each
outcome of the image. For instance, in order to denoise
iteration.
brain MRI images, training is conducted and the results are
ICA comprises two architectures designed for deriving
presented in Figure 5.
feature vectors. The chosen architecture, ICA1, is specif-
ically applied to identify components in medical images,
D. STEP 04:INDEPENDENT COMPONENT ANALYSIS FOR
particularly in the context of brain MRI images. Its imple-
BRAIN MRI IMAGES
mentation involves evaluating ICA’s effectiveness in contrast
In the methodology section, we elaborate on our utilization enhancement. Notably, ICA1 demonstrates excellent contrast
of Independent Component Analysis (ICA) as the computa- image quality and performs well in segmenting abnormal
tional technique chosen to disentangle intricate, multivariate regions, such as brain tumors. The following section offers
signals into self-standing, non-Gaussian components. Our a detailed explanation of the application of ICA1 on brain
primary application of ICA is within the realm of brain MRI MRI images, elaborating on its capabilities in enhancing
image analysis, where it serves the purpose of distinguishing contrast and accurately delineating abnormal regions within
and isolating the autonomous components residing within the images.
these MRI images. Consequently, the outcome comprises a
collection of normalized images that conspicuously exhibit 1) ICA ARCHITECTURE 1 (ICA1) ON BRAIN MRI IMAGES
the extracted independent component brain MRI images,
The ICA1 model involves preparing the image database in
characterized by their normalized and coherent contrast.
a matrix format, where each row vector corresponds to an
Equation 17 demonstrates how the Independent Component
individual image or ICA independent component. This model
Analysis (ICA) approach can be further developed from
is illustrated in Fig. 6, where images are regarded as random
its mixed model. The equation illustrates that the observed
variables and pixels are treated as trials. As depicted in
mixed signals (X) can be decomposed into a matrix of
the figure, the data matrix X , also known as the mixture
independent source signals (S) that are linearly combined
matrix, is a combination of n independent ICA components or
through a mixing matrix (A). By utilizing this equation, ICA
images. The coefficient matrix, denoted by W , is calculated
can effectively separate and extract the independent source
using the FASTICA algorithm, with W being required to be in
signals from the observed mixed signals.
a square pattern. The source matrix, denoted by S, comprises
X = AS. (17) n independent basis images. In the ICA1 approach, the data
matrix has n training samples with image column vectors of
The ICA model is represented by Equation 17. It assumes length m. The dimension of the data matrix is represented
that the mixture vector X and independent component as m × n with column data matrix X = x1 , x2 , . . . , xn . The

42874 VOLUME 12, 2024


A. A. Asiri et al.: Optimized Brain Tumor Detection: A Dual-Module Approach

of SI correspond to the base images of ICA1. The final


independent images are obtained by projecting the vertically
centered data matrix on the independent eigenvectors of
the base images, as given by Z = yve yT WI . The number
of independent components in Z depends on the user’s
requirements. In the case of brain MRI images, which are
in JPEG format and contain three independent components,
ICA1 yields three independent components. This process is
explained in greater detail below.
The utilization of Independent Component Analysis (ICA)
is a crucial step in our approach, wherein we apply ICA
individually to each component of the brain MRI images
generated by the RBF neural network. This process requires
FIGURE 6. ICA Architecture 1 model.
a deep comprehension of the intrinsic imaging and biological
characteristics of these images before any normalization steps
are taken. It’s noteworthy that the three components of the
main idea behind implementing ICA1 is to transpose the data
brain MRI images are linearly independent. However, a closer
matrix X into a mixture row data matrix Y = X T . The Y
examination reveals distinct attributes: the first component
matrix is represented as Y = y1 , y2 , . . . . . . ym . The following
(ICA1) contains both significant luminance information and
steps are followed to implement ICA1.
noise, the second component exhibits reduced noise, while
the third component introduces shadows and additional
a: CENTERING PROCESS
noise. The primary objective of our study centers on
In the first step, the mean
P vector is estimated from the
identifying a suitable component within the brain MRI
data matrix as µI = m1 j = 1m yj . This mean vector is
images that showcases well-contrasted regions against a
subtracted from each column vector of the data matrix such
more uniform background. Consequently, we employ ICA
that (yj − µI ) → yj → yv . Here, yv represents the vertically
on each non-uniform background output image to enhance
centered row data matrix, and each column vector of the data
the contrast levels in various regions of the contrast-enhanced
matrix has been subtracted by the mean vector to generate a
MRI (CE-MRI) brain images when juxtaposed against their
zero-mean image row data matrix.
respective backgrounds. It’s essential to bear in mind that the
CE-MRI database encompasses three distinct imaging planes,
b: WHITENING PROCESS
each offering unique information. Furthermore, tumors
The subsequent step involves the whitening of the vertically may manifest in various locations within each imaging
centered data matrix using Principal Component Analysis plane.
(PCA). To perform this, orthonormal eigenvectors V = ICA1 architecture is employed to investigate its efficacy
m
v1 , v2 , v3 . . . ..vn of the covariance matrix I = m1 yj yj T
P P
in enhancing the contrast of various brain MRI regions.
j=1 The findings of this study are displayed in Fig. 7, which
are computed, with the largest positive eigenvalues (p) as
demonstrates that the second component of ICA1 ( basis
y1 ≥ y2 ≥ . . . . . . .. ≥ yp . The whitening process can be
on measurement of contrast and signal to noise ratio also)
expressed mathematically as:
produced superior results in terms of improving the contrast
1
H = V D− 2 . (19) of different regions against their background compared to
the first and third components. The primary contribution of
where D is the diagonal matrix of the largest eigenvalues. this research is the validation and implementation of ICA1
architecture to produce a normalised and enhanced contrast
c: TRANSFORMATION PROCESS brain MRI image, taking into account imaging properties
After performing the whitening process, the centered row such as biological, structural, and imaging properties. This
data matrix yv is transformed using the orthonormal eigen- implementation could aid in accurately segmenting brain
vectors V as follows: tumors.
y = H T yv .
e (20)
IV. POST-PROCESSING:BRAIN TUMOR CLASSIFICATION
These pre-processing steps play a crucial role in the Our proposed method focuses on the impact of enhancement
application of ICA, as they prepare the data matrix for the techniques based on ICA1 analysis on the classification
subsequent analysis. Once these steps are performed, ICA is process. The postprocessing step involves classifying brain
applied to the whitened data matrix, denoted by H , to obtain tumors using Support Vector Machine (SVM), a supervised
a matrix SI consisting of independent images arranged row- learning method that utilizes statistical learning theory to
wise. This is achieved by multiplying the whitened data classify data [37]. The first essential step in this process is
matrix e
y with a coefficient matrix WI , where the row vectors data labeling, which entails representing the training dataset

VOLUME 12, 2024 42875


A. A. Asiri et al.: Optimized Brain Tumor Detection: A Dual-Module Approach

FIGURE 7. Comparsion of Enhancement techniques with ICA1 enhanced


image. Fig(a) is first component of ICA1. Fig(b) is second component of
ICA1 Fig (c) is third component of ICA1.

as D = {|x, y| , |x → datasample, y → classlabel|}. SVM’s


primary objective is to calculate functions represented by f ,
such as f (x) = y, for all image data or pixels to achieve
accurate brain tumor classification. The mapping function’s
primary goal is to establish a relationship between the
labeled samples and classified data, enabling accurate brain
tumor detection. The decision function is utilized to classify
the tumor and non-tumor regions, and its mathematical
representation is known as the feed-forward process of SVM
classification, as illustrated below:
X 
N
D (m) = αi yi K (d i mi ) + t . (21) FIGURE 8. Figure shows the steps involved in the SVM-based
classification process for detecting brain tumors. In Figure (a), the original
i=1
brain image is converted to a binary image. Next, in Figure (b), the binary
The equation shown in 21 represents the alpha coefficient image is skeletonized. The resulting image is then divided into zones and
appended areas to form an image matrix. The generated feature vector is
(αi ) of support vector class labels or feature vectors. The based on Euler numbers and related parameters, such as pixel
SVM vector is denoted as yi , and the input vector is denoted distributions with x and y planes, resulting in about 100 feature vectors
for each image. The SVM classifier is then trained using these feature
as d i . The kernel function with a bias t is represented by vectors in matrix form, as shown in Figure(c), to classify tumor regions.
K (d i mi ). In SVM-based brain tumor image classification, the Finally, the decision function of the SVM is used to detect the tumor
process can be broken down into three steps. The first step region, as shown in Figure(d).

involves selecting the feature vector through feature vector


extraction. The second step involves training the data, and the
third step involves tumor classification and identification of V. DATABASE AND PARAMETERS MEASUREMENT
the tumor region. he efficacy of the suggested method is assessed through
The feature vector is created by combining data in an array, three distinct criteria: the enhancement of the brain image,
allowing for object classification on a per-feature vector the segmentation of the tumor region, and the classification
basis. In brain tumor images, the image is first converted to a of brain tumors. The enhancement outcomes are evaluated
binary image, as seen in Figure 8(b), and then skeletonized, based on two factors: contrast and peak signal-to-noise
as shown in Figure 8(a). The image is then divided into ratio. The segmentation outcomes are appraised using five
zones and appended areas to form the image matrix. The metrics: mean, standard deviation, entropy, kurtosis, and
generated feature is based on Euler numbers and related skewness. The performance of the brain tumor classification
parameters, such as pixel distributions with the x and y planes. is determined using sensitivity, specificity, and accuracy
Approximately 100 feature vectors are generated to classify measures.
the tumor region in the image. For our research, we processed
310 brain tumor images, and SVM training was conducted A. MEASURING PARAMETERS: ENHANCEMENT OF THE
by combining the feature vectors in matrix form to classify BRAIN IMAGE
tumor regions. The support vectors of the SVM method are The enhancement results are assessed for performance by
the closest collection of data points to the decision surface, quantifying five parameters: peak signal-to-noise ratio and
making it challenging to classify brain tumors. These points contrast. Each of these parameters is further explained below.
are determined based on a location optimization process to
decide the surface to locate the anomalous region. SVM 1) SIGNAL TO NOISE RATIO OF THE BRAIN MRI IMAGE
maximizes the margin around the hyperplane separation, and The Peak Signal to Noise Ratio (PSNR) is a parameter used
the decision function is based on the training sample subset to assess the quality of a brain image compared to a noisy
to identify the tumor region. The output of the SVM process image. It calculates the ratio between the maximum possible
is presented in Figure 8. signal level and the noise level present in the MRI image of

42876 VOLUME 12, 2024


A. A. Asiri et al.: Optimized Brain Tumor Detection: A Dual-Module Approach

the brain. The PSNR calculation formula, as introduced by L. this can be represented in the equation 25.
Jong and Sen in 1980, is represented by the Equation 22. v
u  m−1 n−1
u 1 XX
STD(σ ) = t (I (i, j) − M )2 . (25)
  u
R
PSNR = 20log10 . (22) m×n
σ i=0 j=0

In this equation, the value of the PSNR is obtained by taking 3) ENTROPY


the logarithm in base 10 of the ratio between the maximum Entropy is a metric that quantifies the level of randomness
value of the signal R and the standard deviation of the noise σ . or unpredictability in an image’s texture. It provides a
The standard deviation σ represents the variability of the measure of the amount of information needed to describe the
image intensities produced by the noise. A higher value of distribution of pixel intensities in the image. Higher entropy
σ represents more noise in the image, while a lower value values indicate a more complex and diverse texture, while
represents less noise. The maximum image value (R) related lower entropy values indicate a more regular and uniform
to the highest possible intensity level that can be represented texture. Mathematically, entropy can be represented in the
in the image. It typically ranges from 0 to 255 in an 8-bit Equation 26.
grayscale image, where 0 represents black and 255 represents m−1 n−1
white. XX
Entropy = − f (i, j) log2 (i, j) . (26)
i=0 j=0
2) IMAGE CONTRAST
From the perspective of brain MRI images, image contrast 4) KURTOSIS
(IC) suggests the difference in intensity level between a In the context of brain MRI images, kurtosis is a parameter
particular pixel and its surrounding neighboring pixels in that provides insight into the probability distribution of
the image. Mathematically, image contrast (IC) can be random variables in the image. This helps us understand
represented by: the shape of the distribution and the presence of outliers
or outliers. The kurtosis is denoted Kurt(I) for image I,
m−1 n−1 2
XX indicating that it applies specifically to the features of that
Contrast = (i − j) f (i, j) . (23) particular image. Mathematically, kurtosis can be represented
i=0 j=0 using Equation 27:
((f (i, j) − M )4 )
 P
B. MEASURING PARAMETERS: SEGMENTATION OF BRAIN 1
kurt (I ) = . (27)
TUMOR m×n ST D4
The segmentation results are assessed for performance By calculating the kurtosis of brain MRI images, we can
by quantifying five parameters: mean, standard deviation, better understand the distribution of intensities, which can be
entropy, kurtosis, and skewness. Each of these parameters is useful in tasks such as identifying abnormalities, detecting
further explained below. subtle changes, or characterizing specific tissue properties.
It helps in understanding the statistical properties of image
1) MEAN data and can help in various medical image analysis
The average parameter is calculated by multiplying the pixel applications.
values of the image by the total number of pixels in the image.
The calculation of the average is represented by the following C. SKEWNESS
equation: In the context of brain MRI images, skewness is a parameter
that provides information about the similarity or dissimilarity
  m−1 n−1 of pixel distributions in the image. Skewness measures
1 XX
Mean = I (i, j). (24) the asymmetry of the distribution and helps us understand
m×n
i=0 j=0 whether pixel values are more concentrated on one side or
the other.
where ‘m’ and ‘n’ represent the number of pixels correspond-
(f (i, j) − M )3
 P
ing to the dimensions of the image. 1
SK (I ) = . (28)
m×n ST D3
2) STANDARD DEVIATION By calculating the asymmetry of brain MRI images, we can
The standard deviation (STD) is obtained by taking the assess the symmetry or asymmetry of the pixel distribution,
square root of the variance or the central mean of the image. which can be useful in various applications such as tissue
It gives insight into how pixels are distributed relative to the characterization, lesion detection and abnormality identifica-
average and reveals information about non-uniformity in the tion. Skewness helps us understand the statistical properties
image. A higher average value suggests greater saturation and of image data and can aid in the analysis and interpretation of
increased contrast at the edges of the image. Mathematically, brain MRI images.

VOLUME 12, 2024 42877


A. A. Asiri et al.: Optimized Brain Tumor Detection: A Dual-Module Approach

D. MEASURING PARAMETERS: CLASSIFICATION OF


BRAIN TUMOR
To assess the effectiveness of classification, we used a
combination of training and test data. To ensure a compre-
hensive evaluation of the database, we used cross-validation,
a widely adopted technique for validating the performance
of classification models. The evaluation of the classification
model involved the use of the following parameters.
FIGURE 9. CE-MRI images illustrate different types of brain tumors as
1) SENSITIVITY follows: Figure (a) depicts a meningioma brain tumor. Figure (b) shows a
glioma brain tumor. Figure (c) shows a pituitary brain tumor.
In the context of brain MRI, sensitivity refers to the
measurement of precisely identified pixels belonging to
positive classes. It quantifies the ability of a classification 4) DICE SCORE
model to correctly detect positive instances. Mathematically, In the context of brain MRI, the Dice Score (DSC) is a
sensitivity is calculated as the ratio of the total number of parameter that provides information about the overlap ratio
true positive pixels to the total number of pixels identified between the predicted output and the actual ground truth
as positive, as defined in the equation 29. Sensitivity is also values. It quantifies the similarity between predicted and
commonly referred to as True Positive Rate (TPR). The ground truth regions of interest. The Dice score is normalized
equation 29 represents the sensitivity calculation: by considering the true positive values compared to the mean
TP of the predicted and ground truth values. Mathematically, the
Sensitivity = (29) Dice score is represented as equation 32.
TP + FN
where as: TP refers to the number of true positive pixels, 2 × TP
DSC = (32)
indicating the number of correctly identified pixels belonging (2 × TP + FN + FP)
to the positive class. FN represents the number of false By evaluating the Dice score in brain MRI, we can assess
negative pixels, indicating the number of pixels belonging to the degree of overlap and similarity between the predicted
the positive class that were incorrectly identified as negative. output and the ground truth values. The Dice score provides
a valuable measure of model performance and helps assess
2) SPECIFICITY
the accuracy and quality of segmentation or registration tasks
In the context of brain MRI images, specificity refers to in MRI analysis of the brain
the measurement of precisely identified pixels belonging to
negative proportions or classes. It quantifies the ability of E. DATABASES
a classification model to correctly detect negative instances. We are making use of the Contrast-Enhanced Magnetic
Mathematically, specificity is calculated as the ratio of the Resonance Imaging (CE-MRI) image dataset, which was
total number of true negative pixels to the total number of established by Nanfang Hospital in Guangzhou, China, and
pixels identified as negative, as defined in the equation 30. the General Hospital at Tianjin Medical University, China,
Specificity is also commonly referred to as true negative rate during the period spanning from 2005 to 2010. The term
(TNR). The equation 30 represents the specificity calculation: ‘‘CE-MRI brain dataset’’ typically refers to a compilation
TN of images generated through contrast-enhanced magnetic
Specificity = (30)
TN + FP resonance imaging techniques, which are crucial for exam-
where as: TN refers to the number of true negative pixels, ining biological structures. In CE-MRI procedures, contrast
indicating the number of correctly identified pixels belonging agents are introduced during scans to enhance the visibility
to the negative class. FP represents the number of false of specific tissues or detect abnormal conditions within
positive pixels, indicating the number of pixels belonging to MRI brain images. This dataset comprises 3064 images
the negative class that were incorrectly identified as positive. collected from 233 individuals, including 930 pituitary
tumors, 708 meningiomas, and 1426 gliomas. These images
3) ACCURACY have a resolution of 512 by 512 pixels, with each pixel
In the context of brain MRI, accuracy is a parameter measuring 0.49 × 0.49, mm2 , and a 1mm gap between
that provides information about the accuracy of predictions slices. The identification of tumors within these images was
regarding true pixels. The most effective approach to performed manually by three highly skilled radiologists.
evaluating the performance of a model’s predictions is to Figure 9 showcases several examples from this CE-MRI
count the number of pixels detected accurately. Accuracy image dataset.
measures the proportion of correct predictions of true pixels Key considerations regarding CE-MRI datasets and their
and is mathematically defined as equation 31. applications encompass:
TP + TN 1) Different imaging techniques, including contrast-
Accuracy(AC) = . (31) enhanced T1-weighted sequences, are utilized to
TP + FP + TN + FN
42878 VOLUME 12, 2024
A. A. Asiri et al.: Optimized Brain Tumor Detection: A Dual-Module Approach

TABLE 1. Impact of image enhancement technique on ce-mri image.

TABLE 2. The analysis of the performance of different image enhancement methods.

acquire CE-MRI datasets. These enhanced T1- VI. RESULTS AND DISCUSSION
weighted sequences facilitate the observation and A. ANALYSIS ON ENHANCEMENT TECHNIQUE ON BRAIN
quantification of the distribution of contrast agents MRI IMAGES
within tissues. The Results and Discussion section of the research paper
2) CE-MRI datasets are frequently compiled for research presents an analysis of peak signal-to-noise ratio (PSNR)
objectives as part of clinical studies. Notably, public and contrast in brain MRI images. This section provides a
repositories and research institutions like The Cancer detailed review of PSNR values and contrast analysis based
Imaging Archive (TCIA) and the Alzheimer’s Dis- on CE-MRI image database and analysis is given below.
ease Neuroimaging Initiative (ADNI) maintain MRI
databases that encompass CE-MRI data, in addition to
other imaging modalities. 1) IMPACT OF ENHANCEMENT TECHNIQUE
3) CE-MRI datasets typically comprise 3D volumetric The results of the analysis of our proposed method on
image sequences obtained at various time intervals fol- the CE-MRI image database, with and without the use
lowing the injection of a contrast agent. These datasets of enhancement techniques, are presented in table 1. The
may encompass baseline anatomical sequences, sup- table shows the results obtained by applying our method to
plementary pre-contrast images, and dynamic post- different types of brain MRI images in the CE-MRI image
contrast series. database, in particular the Meningiomas Tumor Images,
4) CE-MRI assesses brain tumors and defines their Gliomas Tumor Images and Pituitary Tumor Image. After
boundaries in brain tumor imaging. Contrast-enhanced careful analysis, he validated that our method significantly
sequences are instrumental in pinpointing tumor- improves the contrast of every type of brain MRI image
afflicted areas within MRI brain images, thereby in the CE-MRI image database. The contrast enhancement
assisting in diagnosis, treatment strategy formulation, amounts to 40 indicating a substantial improvement in the
and therapy effectiveness evaluation. visibility and distinction of various structures or regions
5) The provided dataset contains 3064 contrast-enhanced in the images as well as leading to accurate segmentation
T1-weighted images collected from 233 individuals, and classification of brain tumors. Additionally, the analysis
encompassing three distinct categories of brain can- reveals a 3dB improvement in the peak signal-to-noise ratio
cers: meningiomas (comprising 708 slices), gliomas (PSNR) metric, signifying the improved quality and fidelity
(comprising 1426 slices), and pituitary tumors (com- of the processed images. These improvements validate the
prising 930 slices) [38]. In our approach, we allocated capability of our proposed technique for the analysis of
70% of the images for the training phase, which medical images, not only limited to brain abnormalities
includes preprocessing steps such as brain image observed in the CE-MRI image database, but also extendable
processing and enhancement (ICA-1). The remaining to other types of medical images. For example, our method
30% of the dataset contains testing the classifier, which has potential applicability to retinal fundus images, allowing
involves post-processing modules utilizing an SVM- better analysis of eye abnormalities. Therefore, the results
based classifier. It is important to note that images highlight the efficiency and versatility of our technique in
from the same subjects of each brain tumor type are improving and analyzing medical images, especially in the
included in both the training and testing process of the context of brain abnormalities in the CE-MRI brain MRI
method. database.

VOLUME 12, 2024 42879


A. A. Asiri et al.: Optimized Brain Tumor Detection: A Dual-Module Approach

TABLE 3. Performance analysis of segmentation model of brain tumor.

FIGURE 10. Brain tumor segmentation results of the proposed method of different position of Axial plane Brain MRI
Image. The first column represents the original images of the database. The second column represents the ground
truth and third column represents the output of the image.

B. COMPARATIVES ANALYSIS OF IMAGE ENHANCEMENT methods consistently outperform other techniques in terms
WITH EXISTING IMAGE ENHANCEMENT TECHNIQUES of PSNR and contrast. The proposed method achieves the
The different image enhancement techniques: Histogram highest PSNR and contrast values for each type of Brain
Equalization (HE), Contrast Limiting Adaptive Histogram MR image, indicating superior enhancement and better
Equalization (CLAHE), Brightness Bi-Histogram Equal- visibility of structures in brain MRI images. This comparative
ization (BBHE) are compared to our proposed method. analysis highlights the effectiveness of the proposed method
By analyzing the performance of different image enhance- in improving image quality, especially in terms of PSNR
ment techniques, it can be observed that the proposed and Contrast. This implies that the proposed method has the

42880 VOLUME 12, 2024


A. A. Asiri et al.: Optimized Brain Tumor Detection: A Dual-Module Approach

FIGURE 11. Brain tumor segmentation results of the proposed method of different position of Coronal
plane Brain MRI Image. The first column represents the original images of the database. The second
column represents the ground truth and third column represents the output of the image.

TABLE 4. Comparison of performance segmentation model of brain tumor detection.

potential to contribute significantly to the analysis of medical parameters. These parameters are analyzed to assess the
images, especially those related to brain abnormalities in the effectiveness of the method in the precise identification and
CE-MRI image database. delineation of brain tumors. Statistical measures, as presented
in the table 3, are used to analyze the performance of
C. ANALYSIS OF SEGMENTATION MODULE the proposed method. These parameters provide valuable
PERFORMANCE information about the quality of the segmentation by
The performance of the brain tumor segmentation method highlighting the contrast obtained in the brain region and
on the CE-MRI database is analyzed by analyzing different the normalization of the image, which facilitates an accurate

VOLUME 12, 2024 42881


A. A. Asiri et al.: Optimized Brain Tumor Detection: A Dual-Module Approach

FIGURE 12. Brain tumor segmentation results of the proposed method of different position of Sagittal plane
Brain MRI Image. The first column represents the original images of the database. The second column
represents the ground truth and third column represents the output of the image.

TABLE 5. Impact of image enhancement technique on classification model.

segmentation of the tumor. The table 3 shows the capability of brain tumor detection, providing valuable evidence of its
the proposed methods as it leads to an accurate segmentation capabilities in the field.
of the brain tumor and visualizes the analysis is shown in
Figure 10, Figure 11 and Figure 12. Visualizations of tumor D. COMPARISON OF SEGMENTATION MODULE
detections clearly show the module’s ability to accurately PERFORMANCE BASED ON DIFFERENT CLASSIFIER
identify and locate brain tumors. These results underscore The Mnay researchers used different classifiers to segment
the competence of the module and its potential to aid in the brain tumors and analyze the corresponding images

42882 VOLUME 12, 2024


A. A. Asiri et al.: Optimized Brain Tumor Detection: A Dual-Module Approach

TABLE 6. Comparison of performance classification model of brain tumor detection with different classifier.

FIGURE 13. Confusion Matrix of random selected Images sets of brain different brain tumor types.

through statistical analysis. Most of these researchers have yield significantly higher sensitivity, specificity, accuracy,
validated their method based on statistical measures such as and DSC values when image enhancement techniques are
mean, standard deviation (STD), intensity correlation (IC), applied compared to when they are not. Sensitivity reflects the
and peak signal-to-noise ratio (PSNR) to assess the brain model’s ability to scan and detect tumor regions uniformly,
tumor in brain MRI images. Performance comparison of the while specificity indicates its accuracy in identifying abnor-
most used classifiers including K-Nearest Neighbor (KNN), mal region pixels.
Self-organizing Map (SOM), Genetic Algorithm (GA), Additionally, DSC and accuracy offer insights into the
Graph Convolutional Neural Network (GCNN), Kernel- overall performance of classifying abnormal regions across
Based Support Vector Machine (SVM) and our proposed all types of tumors. Specifically, the ‘‘Meningiomas Tumor
method which is presented in the Table 4. The results show Images’’ category demonstrates a substantial improvement in
that our proposed method outperforms existing techniques, sensitivity (Sen), increasing from 0.67 without enhancement
demonstrating its effectiveness in detecting brain tumors to 0.99 with enhancement. Similar enhancements in perfor-
and accurately classifying tumor regions. Classification mance are observed in the ‘‘Gliomas Tumor Images’’ and
performance will be validated in the next section. ‘‘Pituitary Tumor Image’’ categories. The model consistently
achieves higher sensitivity, specificity, accuracy, and DSC
when applied image enhancement techniques.
E. IMPACT OF IMAGE ENHANCEMENT TECHNIQUE ON These noteworthy improvements in sensitivity, specificity,
CLASSIFICATION MODULE PERFORMANCE accuracy, and DSC strongly indicate that image enhancement
The evaluation of image enhancement methods’ impact on techniques positively impact the classification model’s ability
brain MRI image classification using the CE-MRI database to detect and categorize tumors in MRI images.
involves assessing key metrics, including sensitivity (Sen),
specificity (Spec), accuracy (Acc), and Dice Similarity
Coefficient (DSC). This section aims to validate image F. COMPARISON OF CLASSIFICATION MODULE
classification performance with and without enhancement PERFORMANCE
methods. These performance metrics gauge the effectiveness Many researchers use the different classifiers for the detection
of our proposed classification methods, as presented in of brain tumors and their analysis is based on the statistical
Table 5. The analysis reveals that our methods consistently measurements of various parameters. The table 6 shows the

VOLUME 12, 2024 42883


A. A. Asiri et al.: Optimized Brain Tumor Detection: A Dual-Module Approach

TABLE 7. Performance of exiting mr imaging segmentation methods.

comparison of different classifiers such as K-Nearest Neigh- improve training data using machine learning techniques. Our
bor (KNN), Self-organizing Map (SOM), Genetic Algorithm objective is to surpass the performance of current methods
(GA), Graph Convolutional Neural Network (GCNN) and in future implementations. By emphasizing innovation and
Kernel-Based SVM, is compared to our proposed method leveraging advanced machine learning techniques, we aim to
classifier known as ICA-NN-SVM. It is clearly observed that make significant strides in brain tumor classification.
our proposed method outperforms the existing methods in
terms of sensitivity, specificity and accuracy as well as we VII. CONCLUSION
also compared the computation time, it is also validated that In this study, we introduced a pioneering two-module
our proposed method exhibits the faster processing time as methodology for the analysis of brain tumors using MRI
compared to existing methods.The effectiveness of the our images. Our approach uniquely addresses the challenges
proposed method in accurately detecting abnormal tumor of image quality enhancement and effective classification,
regions is verified through statistical parameter analysis. aiming to revolutionize the field of brain tumor analysis. The
Essential metrics such as sensitivity, specificity, and accuracy initial module termed the ‘Image Enhancement Technique
are derived from the confusion matrix. The confusion matrix for Brain MRI Images’, focuses on refining image quality
of random selected images of our proposed method is shown by employing a combination of machine learning and
in the Figure 13. imaging techniques. Specifically, adaptive Wiener filter-
ing, neural networks, and independent component analysis
G. COMPARATIVE ANALYSIS WITH EXISTING WORK work collaboratively to enhance and normalize images.
To assess the effectiveness of our proposed method, we per- This pre-processing step optimally prepares the images for
formed an extensive comparison with recent techniques subsequent segmentation and classification tasks.
developed between 2019 and present. The results of the The second module of our approach provides robust
performance evaluation are presented in table 7. It is analyzed validation for the image enhancement technique and performs
that despite the numerous methods based on CNN or deep two critical functions: segmentation and classification. The
learning since 2019, their performance remains relatively segmentation process effectively isolates and highlights
low due to lack of data as well as proper training. This tumor regions within the brain images, while the classifica-
can be attributed to the limited innovation in treatment tion accurately categorizes tumors into specific types using
methods within these approaches. In contrast, our proposed support vector machines (SVM).
method ICA-NN-SVM. This helps improve performance by Evaluation of our method using a diverse CE-MRI database
improving consistency between different brain MRI image demonstrated considerable advancements in contrast and effi-
regions, leading to more accurate classification. Our future ciency for differentiating tumor types. Notably, our approach
work will mainly focus on designing new methods to achieved outstanding performance metrics, including an

42884 VOLUME 12, 2024


A. A. Asiri et al.: Optimized Brain Tumor Detection: A Dual-Module Approach

average sensitivity, specificity, accuracy, and Dice score [12] M. Sharif, J. Amin, M. Raza, M. Yasmin, and S. C. Satapathy,
(DSC) of 0.99, 0.99, 0.989, and 0.981, respectively. More- ‘‘An integrated design of particle swarm optimization (PSO) with fusion
of features for detection of brain tumor,’’ Pattern Recognit. Lett., vol. 129,
over, the speed of our method sets it apart from existing pp. 150–157, Jan. 2020.
approaches, with an average processing time of 0.43 seconds, [13] A. Kumar, M. Ramachandran, A. H. Gandomi, R. Patan, S. Lukasik,
highlighting a unique blend of accuracy and computational and R. K. Soundarapandian, ‘‘A deep neural network based classifier
for brain tumor diagnosis,’’ Appl. Soft Comput., vol. 82, Sep. 2019,
efficiency. Art. no. 105528.
It is evident from our results that our method outperforms [14] F. Özyurt, E. Sert, E. Avci, and E. Dogantekin, ‘‘Brain tumor detec-
contemporary methods in sensitivity, specificity, accuracy, tion based on convolutional neural network with neutrosophic expert
maximum fuzzy sure entropy,’’ Measurement, vol. 147, Dec. 2019,
and DSC. This underscores the potential of our two-module Art. no. 106830.
approach to impact brain tumor analysis through image [15] A. R. Raju, P. Suresh, and R. R. Rao, ‘‘Bayesian HCS-based multi-SVNN:
enhancement and classification significantly. Looking ahead, A classification approach for brain tumor segmentation and classification
using Bayesian fuzzy clustering,’’ Biocybern. Biomed. Eng., vol. 38, no. 3,
there are promising avenues for further research based on pp. 646–660, 2018.
our groundbreaking methodology. Integrating more advanced [16] B. Yin, C. Wang, and F. Abza, ‘‘New brain tumor classification method
artificial intelligence techniques, such as deep learning based on an improved version of whale optimization algorithm,’’ Biomed.
models, holds the potential further to improve the accuracy Signal Process. Control, vol. 56, Feb. 2020, Art. no. 101728.
[17] A. Krol and B. Gimi, ‘‘Medical imaging 2017: Biomedical applica-
and efficiency of brain tumor classification. Additionally, tions in molecular, structural, and functional imaging,’’ in Society of
exploring multi-modal imaging data and incorporating a Photo-Optical Instrumentation Engineers. Bellingham, WA, USA: SPIE,
wider variety of tumor types and demographic ranges in the 2017.
[18] M. Mittal, L. M. Goyal, S. Kaur, I. Kaur, A. Verma, and D. J. Hemanth,
dataset could enhance the generalizability of the method. ‘‘Deep learning based enhanced tumor segmentation approach
In conclusion, our study lays the groundwork for a for MR brain images,’’ Appl. Soft Comput., vol. 78, pp. 346–354,
future where brain tumor detection and classification are May 2019.
[19] M. A. Mohammed, M. K. Abd Ghani, R. I. Hamed, D. A. Ibrahim, and
significantly more precise, efficient, and widely applicable. M. K. Abdullah, ‘‘Artificial neural networks for automatic segmentation
By continuously refining and innovating our methodology, and identification of nasopharyngeal carcinoma,’’ J. Comput. Sci., vol. 21,
we envision contributing to improved global health outcomes pp. 263–274, Jul. 2017.
in the field of brain tumor analysis [20] A. Rehman, S. Naz, M. I. Razzak, F. Akram, and M. Imran, ‘‘A deep
learning-based framework for automatic brain tumors classification using
transfer learning,’’ Circuits, Syst., Signal Process., vol. 39, pp. 757–775,
REFERENCES Sep. 2020.
[1] Z. Khazaei, E. Goodarzi, V. Borhaninejad, F. Iranmanesh, H. Mirshekar- [21] M. K. Abd-Ellah, A. I. Awad, A. A. M. Khalaf, and H. F. A. Hamed,
pour, B. Mirzaei, H. Naemi, S. M. Bechashk, I. Darvishi, R. E. Sarabi, ‘‘Two-phase multi-model automatic brain tumour diagnosis system
and A. Naghibzadeh-Tahami, ‘‘The association between incidence and from magnetic resonance images using convolutional neural networks,’’
mortality of brain cancer and human development index (HDI): An EURASIP J. Image Video Process., vol. 2018, no. 1, p. 97, 2018.
ecological study,’’ BMC Public Health, vol. 20, p. 1696, Nov. 2020. [22] N. Abiwinanda, M. Hanif, S. T. Hesaputra, A. Handayani, and
[2] T. S. Armstrong, M. Z. Cohen, J. Weinberg, and M. R. Gilbert, T. R. Mengko, ‘‘Brain tumor classification using convolutional neural
‘‘Imaging techniques in neuro-oncology,’’ Seminars Oncol. Nursing, network,’’ in Proc. World Congr. Medical Phys. Biomed. Eng., vol. 1.
vol. 20, pp. 231–239, Nov. 2004. Prague, Czech Republic: Springer, 2019, pp. 183–189.
[3] H. W. Kim, L. Van Assche, R. B. Jennings, W. B. Wince, C. J. Jensen, [23] A. Pashaei, H. Sajedi, and N. Jazayeri, ‘‘Brain tumor classification via
W. G. Rehwald, D. C. Wendell, L. Bhatti, D. M. Spatz, M. A. Parker, convolutional neural network and extreme learning machines,’’ in Proc.
and E. R. Jenista, ‘‘Relationship of T2-weighted MRI myocardial 8th Int. Conf. Comput. Knowl. Eng. (ICCKE), Oct. 2018, pp. 314–319.
hyperintensity and the ischemic area-at-risk,’’ Circulation Res., vol. 117, [24] A. Gumaei, M. M. Hassan, M. R. Hassan, A. Alelaiwi, and G.
no. 3, pp. 254–265, 2015. Fortino, ‘‘A hybrid feature extraction method with regularized extreme
[4] W. Deng, Q. Shi, K. Luo, Y. Yang, and N. Ning, ‘‘Brain tumor segmentation learning machine for brain tumor classification,’’ IEEE Access, vol. 7,
based on improved convolutional neural network in combination with non- pp. 36266–36273, 2019.
quantifiable local texture feature,’’ J. Med. Syst., vol. 43, no. 6, p. 152, [25] S. Samarth R Phaye, A. Sikka, A. Dhall, and D. Bathula, ‘‘Dense
2019. and diverse capsule networks: Making the capsules learn better,’’ 2018,
[5] A. Lundervold and A. Lundervold, ‘‘An overview of deep learning in arXiv:1805.04001.
medical imaging focusing on MRI,’’ Zeitschrift für Medizinische Physik, [26] M. M. Badža and M. Č. Barjaktarović, ‘‘Classification of brain tumors from
vol. 29, no. 2, pp. 102–127, 2019. MRI images using a convolutional neural network,’’ Appl. Sci., vol. 10,
[6] Y. Jiang, J. Hou, X. Xiao, and H. Deng, ‘‘A brain tumor segmentation new no. 6, p. 1999, 2020.
method based on statistical thresholding and multiscale CNN,’’ in Proc. [27] M. Sharma, G. N. Purohit, and S. Mukherjee, ‘‘Information retrieves from
Int. Conf. Intell. Comput., 2018, pp. 235–245. brain MRI images for tumor detection using hybrid technique K-means
[7] M. Hssayeni, ‘‘Computed tomography images for intracranial hemorrhage and artificial neural network (KMANN),’’ in Networking Communication
detection and segmentation,’’ Intracranial Hemorrhage Segmentation and Data Knowledge Engineering (Lecture Notes on Data Engineering and
Using Deep Convolutional Model. Data, vol. 5, no. 1, p. 14, Mar. 2020. Communications Technologies), vol. 4. G. Perez, K. Mishra, S. Tiwari,
[8] T. A. Soomro and J. Gao, ‘‘Neural network based denoised methods for and M. Trivedi, Eds. Singapore: Springer, 2018, doi: 10.1007/978-981-10-
retinal fundus images and MRI brain images,’’ in Proc. Int. Joint Conf. 4600-1_14.
Neural Netw. (IJCNN), Vancouver, BC, Canada, Jul. 2016, pp. 1151–1157. [28] R. R. Monika and A. Kamboj, ‘‘Brain tumor classification for MR imaging
[9] E. F. Badran, E. G. Mahmoud, and N. Hamdy, ‘‘An algorithm for detecting using support vector machine,’’ in Progress in Advanced Computing and
brain tumors in MRI images,’’ in Proc. Int. Conf. Comput. Eng. Syst., Intelligent Engineering (Advances in Intelligent Systems and Computing),
Nov. 2010, pp. 368–373. vol. 714, C. Panigrahi, A. Pujari, S. Misra, B. Pati, and K. C. Li, Eds.
[10] T. A. Soomro, L. Zheng, A. J. Afifi, A. Ali, S. Soomro, M. Yin, and Singapore: Springer, 2019, doi: 10.1007/978-981-13-0224-4_16.
J. Gao, ‘‘Image segmentation for MR brain tumor detection using machine [29] R. Torres-Molina, C. Bustamante-Orellana, A. Riofrío-Valdivieso,
learning: A review,’’ IEEE Rev. Biomed. Eng., vol. 16, pp. 70–90, 2023. F. Quinga-Socasi, R. Guachi, and L. Guachi-Guachi, ‘‘Brain tumor
[11] A. Selvapandian and K. Manivannan, ‘‘Fusion based glioma brain classification using principal component analysis and kernel support
tumor detection and segmentation using ANFIS classification,’’ Comput. vector machine,’’ in Proc. 20th Int. Conf. Intell. Data Eng. Automated
Methods Programs Biomed., vol. 166, pp. 33–38, Nov. 2018. Learn. Manchester, U.K.: Springer, 2019, pp. 89–96.

VOLUME 12, 2024 42885


A. A. Asiri et al.: Optimized Brain Tumor Detection: A Dual-Module Approach

[30] T. A. Soomro, N. A. Jandan, A. Ali, M. Irfan, S. Rahman, [50] J. Amin, M. Sharif, M. Raza, T. Saba, R. Sial, and S. A. Shad, ‘‘Brain tumor
W. A. Aldhabaan, A. S. Khairallah, and I. Abuallut, ‘‘Impact of detection: A long short-term memory (LSTM)-based learning model,’’
retinal vessel image coherence on retinal blood vessel segmentation,’’ Neural Comput. Appl., vol. 32, no. 20, pp. 15965–15973, Oct. 2020.
Electronics, vol. 12, no. 2, p. 396, 2023. [51] M. S. Alam, M. M. Rahman, M. A. Hossain, M. K. Islam, K. M. Ahmed,
[31] Y. E. Almalki, N. A. Jandan, T. A. Soomro, A. Ali, P. Kumar, K. T. Ahmed, B. C. Singh, and M. S. Miah, ‘‘Automatic human brain tumor
M. Irfan, M. U. Keerio, S. Rahman, A. Alqahtani, S. M. Alqhtani, detection in MRI image using template-based K means and improved fuzzy
M. A. M. Hakami, S. A. Saeed, W. A. Aldhabaan, and A. S. Khairallah, c means clustering algorithm,’’ Big Data Cognit. Comput., vol. 3, no. 2,
‘‘Enhancement of medical images through an iterative mccann retinex p. 27, May 2019.
algorithm: A case of detecting brain tumor and retinal vessel segmenta- [52] S. Janardhanaprabhu and V. Malathi, ‘‘Brain tumor detection using depth-
tion,’’ Appl. Sci., vol. 12, no. 16, p. 8243, 2022. first search tree segmentation,’’ J. Med. Syst., vol. 43, no. 8, pp. 1–12,
[32] L. I. Rudin and S. Osher, ‘‘Nonlinear total variation based noise removal Aug. 2019.
algorithms,’’ Physica D, Nonlinear Phenomena, vol. 60, no. 1, 1992, [53] P. Rajan and C. Sundar, ‘‘Brain tumor detection and segmentation by
Art. no. 259268. intensity adjustment,’’ J. Med. Syst., vol. 43, no. 8, pp. 1–13, 2019.
[54] S. Pereira, A. Pinto, V. Alves, and C. A. Silva, ‘‘Brain tumor segmentation
[33] K. Debakla and M. Benyettou, ‘‘Image restoration using multilayer neural
using convolutional neural networks in MRI images,’’ IEEE Trans. Med.
networks with minimization of total variation approach,’’ IJCSI Int.
Imag., vol. 35, no. 5, pp. 1240–1251, May 2016.
J. Comput. Sci. Issues, vol. 383, no. 1, pp. 424–432, 2014.
[55] N. Noreen, S. Palaniappan, A. Qayyum, I. Ahmad, M. Imran, and
[34] J. Sen, ‘‘Digital image enhancement and noise filtering by use of local M. Shoaib, ‘‘A deep learning model based on concatenation
statistics,’’ IEEE Trans. Pattern Anal. Mach. Intell., vol. PAMI-2, no. 2, approach for the diagnosis of brain tumor,’’ IEEE Access, vol. 8,
pp. 165–168, Mar. 1980. pp. 55135–55144, 2020.
[35] A. Hyvarinen, ‘‘Fast and robust fixed-point algorithms for independent [56] M. Sharif, J. Amin, M. Raza, M. A. Anjum, H. Afzal, and S. A. Shad,
component analysis,’’ IEEE Trans. Neural Netw., vol. 10, no. 3, ‘‘Brain tumor detection based on extreme learning,’’ Neural Comput. Appl.,
pp. 626–634, May 1999. col. 43, p. 282, Jul. 2020.
[36] Y. Matsuda and K. Yamaguchi, ‘‘Gram-schmidt orthonormalization to [57] P. Saxena, A. Maheshwari, and S. Maheshwari, ‘‘Predictive modeling of
the adaptive ICA function for fixing the permutation ambiguity,’’ in brain tumor: A deep learning approach,’’ Innovations in Computational
Proc. Int. Conf. Neural Inf. Process. Cham, Switzerland: Springer, 2016, Intelligence and Computer Vision (Advances in Intelligent Systems and
pp. 152–159. Computing), vol. 1189, M. K. Sharma, V. S. Dhaka, T. Perumal, N. Dey,
[37] R. Vankdothu and M. A. Hameed, ‘‘Brain tumor segmentation of MR and J. M. R. S. Tavares, Eds. Singapore: Springer, 2021, doi: 10.1007/978-
images using SVM and fuzzy classifier in machine learning,’’ Meas., 981-15-6067-5_30.
Sensors, vol. 24, Dec. 2022, Art. no. 100440. [58] E. M. Senan, M. E. Jadhav, T. H. Rassem, A. S. Aljaloud,
[38] C. Jun, ‘‘Brain tumor dataset,’’ Figshare Dataset, Jun. 2017, doi: B. A. Mohammed, and Z. G. Al-Mekhlafi, ‘‘Early diagnosis of
10.6084/m9.figshare.1512427.v5. brain tumour mri images using hybrid techniques between deep and
[39] H. A. Vrooman, C. A. Cocosco, F. van der Lijn, R. Stokking, machine learning,’’ Comput. Math. Methods Med., vol. 2022, May 2022,
M. A. Ikram, M. W. Vernooij, Monique M. B. Breteler, and W. J. Niessen, Art. no. 8330833.
‘‘Multi-spectral brain tissue segmentation using automatically trained k-
nearest-Neighbor classification,’’ Neuroimage, vol. 37, no. 1, pp. 71–81,
2007. ABDULLAH A. ASIRI received the Ph.D. degree
[40] T. Logeswari and M. Karnan, ‘‘An improved implementation of brain in medical imaging from the Centre for Advanced
tumor detection using segmentation based on hierarchical self organizing Imaging, The University of Queensland, Australia.
map,’’ Int. J. Comput. Theory Eng., vol. 2, no. 4, pp. 591–595, He has more than 13 years of academic expe-
2010. rience (since 2010) in teaching and research.
[41] A. Kharrat, K. Gasmi, M. B. Messaoud, N. Benamrane, and M. Abid, Currently, he is an Associate Professor with
‘‘A hybrid approach for automatic classification of brain MRI using genetic the Radiological Sciences Department, Najran
algorithm and support vector machine,’’ Leonardo J. Sci., vol. 17, no. 1, University, Saudi Arabia. His research interests
pp. 71–82, 2010. include neuroimaging and diagnostic radiology,
[42] A. K. Mandle, S. P. Sahu, and G. Gupta, ‘‘Brain tumor segmentation and especially MRI. He was the Chairperson of the
classification in MRI using clustering and kernel-based SVM,’’ Biomed. Radiological Sciences Department, from 2019 to 2023. He acted as the
Pharmacol. J., vol. 15, no. 2, pp. 699–716, Jun. 2022. Vice Dean for the Academic Affairs with the College of Applied Medical
[43] N. B. Bahadure, A. K. Ray, and H. P. Thethi, ‘‘Image analysis for MRI Sciences, for one year. He is currently the Director of Scholarships with
based brain tumor detection and feature extraction using biologically Najran University. He is also the Founder of MRI Saudi and a member of
inspired BWT and SVM,’’ Int. J. Biomed. Imag., vol. 2017, Mar. 2017,
several national and international radiology societies.
Art. no. 9749108.
[44] P. Mohamed Shakeel, T. E. El. Tobely, H. Al-Feel, G. Manogaran,
and S. Baskar, ‘‘Neural network based brain tumor detection using TOUFIQUE AHMED SOOMRO (Senior Mem-
wireless infrared imaging sensor,’’ IEEE Access, vol. 7, pp. 5577–5588, ber, IEEE) received the Ph.D. degree in AI and
2019. image processing from Charles Sturt University,
[45] P. Kumar Mallick, S. H. Ryu, S. K. Satapathy, S. Mishra, G. N. Nguyen, and Australia, in 2018. Currently, he is an Associate
P. Tiwari, ‘‘Brain MRI image classification for cancer detection using deep Professor and the Head of the Department of Elec-
wavelet autoencoder-based deep neural network,’’ IEEE Access, vol. 7, tronic Engineering, The University of Larkano,
pp. 46278–46287, 2019.
Pakistan. He is an accomplished academic with
[46] C. Han, L. Rundo, R. Araki, Y. Nagano, Y. Furukawa, G. Mauri,
expertise in engineering and artificial intelligence
H. Nakayama, and H. Hayashi, ‘‘Combining noise-to-image and image-to-
(AI). He has made significant contributions to
image GANs: Brain MR image augmentation for tumor detection,’’ IEEE
Access, vol. 7, pp. 156966–156977, 2019.
the field, with 61 research publications in AI for
[47] M. Li, L. Kuang, S. Xu, and Z. Sha, ‘‘Brain tumor detection based on
medical imaging. He has collaborated on projects related to AI for biomed-
multimodal information fusion and convolutional neural network,’’ IEEE ical imaging with the Ministry of Education, Saudi Arabia, and Najran
Access, vol. 7, pp. 180134–180146, 2019. University. Additionally, he was honored as a Young Research Professor
[48] T. Hossain, F. S. Shishir, M. Ashraf, M. A. Al Nasim, and F. Muhammad with Guangdong University of Technology, China, from 2019 to 2020. With
Shah, ‘‘Brain tumor detection using convolutional neural network,’’ in his extensive knowledge and research expertise, he continues to advance the
Proc. 1st Int. Conf. Adv. Sci., Eng. Robot. Technol. (ICASERT), May 2019, field of AI and image processing. As an Educator, he inspires students and
pp. 1–6. contributes to the growth of electronic engineering. His accomplishments
[49] S. N. Qasem, A. Nazar, A. Qamar, S. Shamshirband, and A. Karim, demonstrate his commitment to research excellence and his contributions to
‘‘A learning based brain tumor detection system,’’ Comput., Mater. the scientific community. His research interests include image enhancement,
Continua, vol. 59, no. 3, pp. 713–727, 2019. segmentation, classification, and analysis for medical images.

42886 VOLUME 12, 2024


A. A. Asiri et al.: Optimized Brain Tumor Detection: A Dual-Module Approach

AHMED ALI SHAH (Senior Member, IEEE) MUHAMMAD IRFAN received the Ph.D. degree
received the Bachelor of Engineering (B.E.) in electrical and electronic engineering from Uni-
degree in electronic engineering from Mehran versiti Teknologi PETRONAS, Malaysia, in 2016.
University of Engineering and Technology, He has two years of industry experience (October
Jamshoro, in 2010. In 2012, he received the 2009–October 2011) and six years of academic
prestigious Higher Education Commission (HEC) experience (since January 2017) in teaching and
Pakistan fully funded HRDI-UET/USTPs Scheme research. Currently, he is an Associate Profes-
Scholarship, for the M.S. studies leading to sor with the Electrical Engineering Department,
the Ph.D. studies from Hanyang University’s Najran University, Saudi Arabia. He has authored
Education Research Industry Cluster Ansan more than 200 research papers in reputed journals,
(E.R.I.C.A ) Campus, South Korea. He has around eight years of diverse books, and conference proceedings (Google Scholar Citations of 2400 and
experience in academia, industry, and research. Over the past few years, H-index of 22). His research interests include automation and process
he has participated in several industrial projects and received various control, energy efficiency, condition monitoring, vibration analysis, artificial
technical training and certifications. He has hands-on experience in the intelligence, the Internet of Things (IoT), big data analytics, smart cities, and
soft fabrication of photonic-based sensing and imaging nanoplatforms, smart healthcare.
specifically his focus was to improve the limit of detection values (LODs),
that in turn enhances sensitivity and minimizes safety risks. His research
interests include surface-enhanced Raman scattering (SERS)-based sensing
and imaging, anisotropic electric field responsive nanostructures, and smart
sensors. To date, his research work has been disseminated to international
audiences, in the form of patents, book chapters, SCI research journals
articles, and peer-reviewed international conference proceedings.

GANNA POGREBNA is a pioneer of behavioural


data science. She is currently the Executive Direc-
tor of the Artificial Intelligence and Cyber Futures
Institute, Charles Sturt University; an Honorary
Professor with The University of Sydney; and the
Lead for Behavioural Data Science with The Alan
Turing Institute. Her expertise blends behavioral
science, AI, computer science, data analytics,
and business innovation. Her work aids diverse
groups, including businesses, governments, and
individuals, in understanding decision-making processes and optimiz-
ing behavior for enhanced outcomes. Her recent projects span smart SAEED ALQAHTANI received the Ph.D. degree in
technologies, cybersecurity, human–computer interactions, and innovative medical imaging from the University of Dundee.
business models. Notably, her work in cybersecurity, particularly her He has more than eight years of academic
CyberDoSpeRT scale, has gained international recognition. His contri- experience (since 2013) in teaching and research.
butions have earned her prestigious awards, including the Leverhulme Currently, he is an Assistant Professor with the
Research Fellowship and the TechWomen100 Award. She is also recognized Radiological Sciences Department, Najran Uni-
as a Leading Data Scientist by the AI Time Journal. She actively versity, Saudi Arabia. His ResearchGate Citations
disseminates her knowledge through her Data Driven YouTube channel is 45, R-score is 27.7, and an H-index is three.
(https://www.youtube.com/@datadrivenchannel/videos) and Inclusion AI His research interests include diagnostic radiology,
blog (https://inclusionai.org/), contributing to The Oxford Handbook of especially MRI and prostate MRI imaging. He is a
Ethics of AIand co-editing the upcoming Handbook of Behavioural Data member of SINAPSE (Scottish Imaging Network).
Science (Cambridge).

VOLUME 12, 2024 42887

You might also like