0% found this document useful (0 votes)
21 views22 pages

Brain Tumor Classification From MRI Using Image

The article presents a novel methodology for brain tumor classification from MRI images using image enhancement techniques and convolutional neural networks (CNNs). The proposed model achieved a classification accuracy of 97.84% for glioma, meningioma, and pituitary tumors, demonstrating superior performance compared to existing methods. This approach aims to assist physicians in making accurate diagnoses and optimizing patient treatment plans.

Uploaded by

Trust Mary
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)
21 views22 pages

Brain Tumor Classification From MRI Using Image

The article presents a novel methodology for brain tumor classification from MRI images using image enhancement techniques and convolutional neural networks (CNNs). The proposed model achieved a classification accuracy of 97.84% for glioma, meningioma, and pituitary tumors, demonstrating superior performance compared to existing methods. This approach aims to assist physicians in making accurate diagnoses and optimizing patient treatment plans.

Uploaded by

Trust Mary
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/ 22

brain

sciences
Article
Brain Tumor Classification from MRI Using Image
Enhancement and Convolutional Neural Network Techniques
Zahid Rasheed 1 , Yong-Kui Ma 1 , Inam Ullah 2, * , Yazeed Yasin Ghadi 3 , Muhammad Zubair Khan 4 ,
Muhammad Abbas Khan 5 , Akmalbek Abdusalomov 6 , Fayez Alqahtani 7 and Ahmed M. Shehata 8

1 School of Electronics and Information Engineering, Harbin Institute of Technology, Harbin 150001, China
2 Department of Computer Engineering, Gachon University, Sujeong-gu, Seongnam-si 13120, Republic of Korea
3 Department of Computer Science, Al Ain University, Abu Dhabi P.O. Box 112612, United Arab Emirates
4 Faculty of Basic Sciences, Balochistan University of Information Technology Engineering and Management
Sciences, Quetta 87300, Pakistan
5 Department of Electrical Engineering, Balochistan University of Information Technology, Engineering and
Management Sciences, Quetta 87300, Pakistan
6 Department of Artificial Intelligence, Tashkent State University of Economics, Tashkent 100066, Uzbekistan;
[email protected]
7 Software Engineering Department, College of Computer and Information Sciences, King Saud University,
Riyadh 12372, Saudi Arabia
8 Computer Science and Engineering Department, Faculty of Electronic Engineering, Menoufia University,
Menofia 32511, Egypt
* Correspondence: [email protected]

Abstract: The independent detection and classification of brain malignancies using magnetic reso-
nance imaging (MRI) can present challenges and the potential for error due to the intricate nature and
time-consuming process involved. The complexity of the brain tumor identification process primarily
stems from the need for a comprehensive evaluation spanning multiple modules. The advancement
of deep learning (DL) has facilitated the emergence of automated medical image processing and
Citation: Rasheed, Z.; Ma, Y.-K.; diagnostics solutions, thereby offering a potential resolution to this issue. Convolutional neural
Ullah, I.; Ghadi, Y.Y.; Khan, M.Z.; networks (CNNs) represent a prominent methodology in visual learning and image categorization.
Khan, M.A.; Abdusalomov, A.; The present study introduces a novel methodology integrating image enhancement techniques, specif-
Alqahtani, F.; Shehata, A.M. Brain ically, Gaussian-blur-based sharpening and Adaptive Histogram Equalization using CLAHE, with
Tumor Classification from MRI Using
the proposed model. This approach aims to effectively classify different categories of brain tumors,
Image Enhancement and
including glioma, meningioma, and pituitary tumor, as well as cases without tumors. The algorithm
Convolutional Neural Network
underwent comprehensive testing using benchmarked data from the published literature, and the
Techniques. Brain Sci. 2023, 13, 1320.
https://doi.org/10.3390/
results were compared with pre-trained models, including VGG16, ResNet50, VGG19, InceptionV3,
brainsci13091320 and MobileNetV2. The experimental findings of the proposed method demonstrated a noteworthy
classification accuracy of 97.84%, a precision success rate of 97.85%, a recall rate of 97.85%, and an
Academic Editors: Carson K. Leung,
F1-score of 97.90%. The results presented in this study showcase the exceptional accuracy of the
Daichi Sone and Iman Beheshti
proposed methodology in accurately classifying the most commonly occurring brain tumor types.
Received: 31 July 2023 The technique exhibited commendable generalization properties, rendering it a valuable asset in
Revised: 10 September 2023 medicine for aiding physicians in making precise and proficient brain diagnoses.
Accepted: 12 September 2023
Published: 14 September 2023 Keywords: deep learning; brain tumor; magnetic resonance imaging; classification; neural network;
pre-trained models; healthcare

Copyright: © 2023 by the authors.


Licensee MDPI, Basel, Switzerland.
1. Introduction
This article is an open access article
distributed under the terms and The development of a brain tumor can occur when there is an abnormal proliferation
conditions of the Creative Commons of cells within the brain tissues. Tumors have been identified by the World Health Organi-
Attribution (CC BY) license (https:// zation (WHO) as the second most significant contributor to global mortality [1,2]. Brain
creativecommons.org/licenses/by/ tumors can be categorized into two main types: benign and malignant. In most instances,
4.0/). benign tumors are not considered a substantial risk to an individual’s health. It is primarily

Brain Sci. 2023, 13, 1320. https://doi.org/10.3390/brainsci13091320 https://www.mdpi.com/journal/brainsci


Brain Sci. 2023, 13, 1320 2 of 22

due to their comparatively slower growth rate than malignant tumors, lack of ability to
infiltrate adjacent tissues or cells, and inability to metastasize. Their recurrence is generally
uncommon after the surgical removal of benign tumors.
Compared to benign tumors, malignant tumors can infiltrate adjacent tissues and
organs, and if not promptly and effectively managed, they can result in significant physio-
logical dysfunction. Detecting brain tumors in their earliest stages is crucial for optimizing
the survival rate of patients. Gliomas, meningioma, and pituitary tumors are the three most
frequently diagnosed types of brain tumors. Glioma is a neoplasm originating from the glial
cells that encompass and provide support to neurons. The cellular composition of these
structures includes astrocytes, oligodendrocytes, and ependymal cells. A pituitary tumor
is formed within the pituitary gland. A meningioma is a tumor originating within the
meninges, the three layers of tissue between the skull and the brain. According to the cited
source, it has been established that meningiomas are classified as benign tumors, while
gliomas are categorized as malignant tumors. Additionally, pituitary tumors have been
identified as benign. The dissimilarity above represents the most notable differentiation
among these three cancer variants [3–5].
Various symptoms can be produced by benign and malignant brain tumors, depend-
ing on factors such as their size, location, and growth rate. The symptoms of primary
brain tumors may exhibit variability among individual patients. Glioma has the potential
to induce various symptoms, including aphasia, visual impairments or loss, cognitive
impairments, difficulties with walking or balance, and other associated manifestations. A
meningioma is often associated with mild symptoms, including visual disturbances and
morning migraines. Pituitary tumors can exert pressure on the optic nerve, leading to
symptoms such as migraines, vision disorders, and diplopia [6,7].
Hence, it is imperative to distinguish among these diverse tumor classifications to
precisely diagnose a patient and determine the optimal course of treatment. The expertise of
radiologists significantly influences the speed at which they can detect brain malignancies.
Although magnetic resonance imaging (MRI) presents challenges due to its dependence
on human interpretation and the complexity of processing large volumes of data, it is
commonly employed to categorize different forms of cancer. Biopsies are commonly
employed in identifying and managing brain lesions, although their utilization before
definitive brain surgery is infrequent. Developing a comprehensive diagnostic instrument
for detecting and classifying tumors based on MR images is imperative [8]. The imple-
mentation of this approach will effectively mitigate the occurrence of excessive operations
and uphold the impartiality of the diagnostic procedure. The healthcare industry has
been significantly influenced by recent technological advancements, particularly in the
fields of artificial intelligence (AI) and machine learning (ML) [9–12]. Solutions to various
healthcare challenges, such as imaging, have been successfully identified [13–18]. Various
machine-learning techniques have been developed to provide radiologists with unusual in-
sights into the recognition and classification of MR images. Medical imaging techniques are
widely recognized as highly effective and widely utilized modalities for cancer detection.
These methodologies facilitate the identification and detection of malignant neoplasms.
The methodology holds significance due to its non-invasive nature, as it does not require
invasive procedures [19,20].
MRI and other imaging modalities are commonly employed in medical interventions
because they produce distinct visual representations of brain tissue, facilitating the identifi-
cation and categorization of diverse brain malignancies. Brain tumors exhibit various sizes,
dimensions, and densities [21]. Moreover, it is worth noting that tumors can exhibit similar
appearances, even when they possess distinct pathogenic characteristics. A substantial
quantity of images within the database posed a significant challenge in classifying MR
images utilizing specialized neural networks. Due to the ability to generate MR images
in multiple planes, there is a potential for increased database sizes. In order to obtain the
desired classification outcome, it is necessary to preprocess MR images before integrating
them into different networks. The Convolutional Neural Network (CNN) is employed to
Brain Sci. 2023, 13, 1320 3 of 22

solve this problem, benefiting from several advantages, such as reduced preprocessing
and feature engineering requirements. A network with lower complexity necessitates a
reduced allocation of resources for implementation and training compared to one with
higher complexity. Resource limitations hinder the utilization of the system for medical
diagnostics or on mobile platforms. The method must be relevant to brain disorders for
daily regular clinical diagnosis.
The main contributions to this investigation are delineated as follows:
• This study presents a novel methodology integrating Gaussian-blur-based sharpening
and Contrast-Limited Adaptive Histogram Equalization (CLAHE) with the proposed
model to facilitate more precise diagnostic procedures for identifying glioma, menin-
gioma, pituitary tumors, and cases without malignancies.
• This investigation aims to demonstrate the superiority of the proposed methodology
above existing methodologies while highlighting its ability to achieve comparable re-
sults with fewer resources. Additionally, an assessment is conducted on the network’s
potential for integration into clinical research endeavors.
• The results obtained from this analysis demonstrate that the novel strategy surpasses
previous methodologies, as indicated by its ability to attain the highest levels of
accuracy on benchmark datasets. Further, we evaluate the prediction capabilities of
this strategy by comparing it to pre-trained models and other established strategies.
The subsequent sections of this work delineate the literature review in Section 2.
Section 3 explores the dataset, methodology, optimization techniques, and pre-trained
models. Section 4 presents the findings obtained from the conducted experiments. Section 5
involves a discussion. Lastly, Section 6 provides a conclusive summary.

2. Literature Review
It is challenging to distinguish between various varieties of brain tumors. The au-
thors [22] examined the clinical applications of DL in radiography and outlined the pro-
cesses necessary for a DL project in this discipline. They also discussed the potential clinical
applications of DL in various medical disciplines. In a few radiology applications, DL
has demonstrated promising results, but the technology is not yet developed enough to
replace the diagnostic occupation of a radiologist [23]. There is a possibility that DL algo-
rithms and radiologists will collaborate to enhance diagnostic effectiveness and efficiency.
Numerous studies have investigated the capability of MRI to identify and classify brain
tumors utilizing a variety of research methodologies. Afshar et al. developed a modified
version of the CapsNet architecture for categorizing the primary brain tumor consisting of
3064 images using tumor boundaries as supplementary inputs to increase effort, surpass
previous techniques, and achieve a classification rate of 90.89% [24]. Gumaei et al. proposed
a brain tumor classification method using hybrid feature extraction techniques and RELM.
The authors preprocessed brain images using min–max normalization, extracted features
using the hybrid method, classified them using RELM, and achieved a maximum accuracy
of 94.23% [25].
Kaplan et al. proposed brain tumor classification models using nLBP and αLBP feature
extraction methods. These models accurately classified the most common brain tumor
types, including glioma, meningioma, and pituitary tumors, and achieved a high accuracy
of 95.56% using the nLBPD = 1 feature extraction method and KNN model [19]. Rezaei et al.
developed an integrated approach for segmenting and classifying brain tumors in MRI
images. The methods included noise removal, SVM-based segmentation, feature extraction,
and selection using DE. Tumor slices were classified using KNN, WSVM, and HIK-SVM
classifiers. Combined with MODE-based ensemble techniques, these classifiers achieved
a 92.46% accuracy rate [26]. Fouad et al. developed a brain tumor classification method
using HDWT-HOG feature descriptors and the WOA for feature reduction. The approach
utilized the Bagging ensemble techniques and achieved an average accuracy of 96.4% with
Bagging, and, when used, Boosting attained 95.8% [27].
Brain Sci. 2023, 13, 1320 4 of 22

Ayadi et al. presented brain tumor classification techniques using normalization,


dense speeded-up robust features, and the histogram of gradient approaches to enhance
the image quality and generate a discriminative feature. In addition, they used SVM for
classification and achieved a 90.27% accuracy on the benchmarked dataset [28]. Srujan et al.
built a DL system with sixteen layers of CNN to classify the tumor types by leveraging
activation functions like ReLU and Adam optimizer, and the system achieved a 95.36%
accuracy [29]. Tejaswini et al. proposed a CNN model to detect meningioma, glioma,
and pituitary brain tumors with an average training accuracy of 92.79% and validation
accuracy of 87.16%; in addition, the tumor region segmentation was performed using Otsu
thresholding, Fuzzy c-means, and watershed techniques [30]. Huang et al. developed a
CNNBCN to classify brain tumors. The network structure was generated using a random
graph algorithm, achieving an accuracy of 95.49% [31].
Ghassemi et al. suggested a DL framework for brain tumor classification. The authors
used pre-trained networks as GAN discriminators to extract robust features and learn MR
image structures. By replacing the fully connected layers and incorporating techniques
like data augmentation and dropout, the method achieved a 95.6% accuracy using fivefold
cross-validation [32]. Deepak et al. combined the CNN feature with SVM for the medical
image classification of brain tumors. The automated system achieved an accuracy of 95.82%
evaluated on the fivefold cross-validation procedure, outperforming the state-of-the-art
method [33]. Noreen et al. adapted fine-tuned pre-trained networks, such as InceptionV3
and Xception, for identifying brain tumors. The models were integrated with various
ML methods, namely Softmax, SVM, Random Forest, and KNN, and achieved a 94.34%
accuracy with the InceptionV3 ensemble [34]. Shaik et al. addressed the challenging task
of brain tumor classification in medical image analysis. The authors introduced a multi-
level attention mechanism, MANet, which combined spatial and cross-channel attention
to prioritize tumors and maintain cross-channel temporal dependencies. The method
achieved a 96.51% accuracy for primary brain tumor classification [35].
Ahmad et al. proposed a deep generative neural network for brain tumor classification.
The method combined variational auto encoders and generative adversarial networks
to generate realistic brain tumor MRI images and achieved an accuracy of 96.25% [36].
Alanazi et al. proposed a deep transfer learning model for the early diagnosis of brain
tumor subtypes. The method involved constructing isolated CNN models and adjusting the
weights of a 22-layer CNN model using transfer learning. The developed model obtained
95.75- and 96.89-percent accuracies on MRI images [37]. Almalki et al. used an ML approach
with MRI to promptly diagnose brain tumor severity (glioma, meningioma, pituitary, and
no tumor). They extracted Gaussian and nonlinear scale features, capturing small details by
breaking MRIs into 8 × 8-pixel images. The strongest features were selected and segmented
into 400 Gaussian and 400 nonlinear scale features, and they were hybridized with each
MRI. They obtained a 95.33% accuracy using the SVM classifier [38]. Kumar et al. compared
three CNN models (AlexNet, ResNet50, and InceptionV3) to classify the primary tumor
types and employed data augmentation techniques. The results showed that AlexNet
achieved an accuracy of 96.2%, surpassing the other models [39].
Swati et al. employed a pre-trained deep CNN model and proposed a block-wise fine-
tuning technique using transfer learning. This approach was evaluated using a standard-
ized dataset consisting of T1-weighted images. Using minimal preprocessing techniques
and excluding handcrafted features, the strategy demonstrated an accuracy of 94.82% with
VGG19, VGG16 achieved 94.65%, and AlexNet achieved 89.95% when evaluated using
a fivefold cross-validation methodology [40]. Ekong et al. integrated depth-wise separa-
ble convolutions with Bayesian techniques to precisely classify and predict brain cancers.
The recommended technique demonstrated superior performance compared to existing
methods in terms of an accuracy of 94.32% [41].
Asiri et al. enhanced computer-aided systems and facilitated physician learning using
artificially generated medical imaging data. A deep learning technique, a Generative
Adversarial Network (GAN), was employed, wherein a generator and a discriminator
94.82% with VGG19, VGG16 achieved 94.65%, and AlexNet achieved 89.95% when eva
uated using a fivefold cross-validation methodology [40]. Ekong et al. integrated depth
wise separable convolutions with Bayesian techniques to precisely classify and predi
Brain Sci. 2023, 13, 1320 5 of 22
brain cancers. The recommended technique demonstrated superior performance com
pared to existing methods in terms of an accuracy of 94.32% [41].
Asiri et al. enhanced computer-aided systems and facilitated physician learning u
engage in a competitive process to generate precise MRI data. The proposed methodology
ing artificially generated medical imaging data. A deep learning technique, a Generativ
demonstrated a notable level of precision, with an accuracy rate of 96%. The evaluation
ofAdversarial
this approach Network (GAN), using
was conducted was employed, wherein a generator
a dataset comprising MRI scans and a discriminator
collected from en
gage in
various a competitive
Chinese process to the
hospitals throughout generate precise MRI
period spanning fromdata.
2005 toThe
2020proposed methodolog
[42]. Shilaskar
etdemonstrated
al. proposed aasystem
notablecomprising
level of precision,
three mainwith an accuracy
components: rate of 96%. HOG
preprocessing, The evaluation
for o
this approach was conducted using a dataset comprising MRI scans collected from variou
feature extraction, and classification. The results indicated varying levels of accuracy
when employing
Chinese multiple
hospitals machinethe
throughout learning classifiers,
period spanning including
from 2005SVM,toGradient Boosting,
2020 [42]. Shilaskar et a
KNN, XG Boost, and Logistic Regression, with the XG Boost classifier attaining
proposed a system comprising three main components: preprocessing, HOG for featur the highest
accuracy rate of 92.02% [43].
extraction, and classification. The results indicated varying levels of accuracy when em
3.ploying
Materials multiple machine learning classifiers, including SVM, Gradient Boosting, KNN
and Methods
XGThis
Boost, and Logistic Regression,
section presents the with thewhich
proposed method, XG Boost classifier
consists of two attaining the highest a
primary compo-
curacy
nents: ratepreprocessing
image of 92.02% [43].and model training. The flowchart illustrating the suggested
system is presented in Figure 1. To enhance the quality of the image, the preprocessing
3. Materials
stage andGaussian-blur-based
incorporated Methods sharpening and Adaptive Histogram Equalization
techniques using CLAHE. Subsequently,
This section presents the proposed labeledmethod,
images were
whichresized whileof
consists maintaining
two primarythe compo
aspect ratio, normalized, and divided into three sets, as shown in Figure 2. Furthermore, the
nents: image preprocessing and model training. The flowchart illustrating the suggeste
model underwent training using 5-fold cross-validation [44] using the Adam optimizer and
system is presented in Figure 1. To enhance the quality of the image, the preprocessin
incorporated the ReduceLROnPlateau callbacks to dynamically regulate the learning rate
stage incorporated
throughout the trainingGaussian-blur-based sharpening
process. The effectiveness and Adaptive
of the proposed model wasHistogram
evaluatedEqualiz
tion metrics
using techniques using
such as CLAHE.
accuracy, Subsequently,
precision, labeled images were resized while main
recall, and F1-score.
taining the aspect
This study ratio,a normalized,
employed and divided
publicly accessible intoMsoud
MRI dataset three sets, as shown
[45], obtained in Figure
from
the Kaggle repository.
Furthermore, This dataset
the model combines
underwent three publicly
training accessible
using 5-fold datasets, including
cross-validation [44] using th
Figshare [46], SARTAJ
Adam optimizer and[47], and BR35H the
incorporated [48].ReduceLROnPlateau
It consists of 7023 MRIs of the human
callbacks brain
to dynamically reg
provided in grayscale and jpg format. The dataset includes primary types of brain
ulate the learning rate throughout the training process. The effectiveness of the propose tumors,
namely glioma, meningioma, pituitary tumors, and images without tumors.
model was evaluated using metrics such as accuracy, precision, recall, and F1-score.

Figure
Figure Flow
1. 1. chart
Flow of the
chart of suggested scheme.
the suggested scheme.
This study employed a publicly accessible MRI dataset Msoud [45], obtained from
the Kaggle repository. This dataset combines three publicly accessible datasets, including
Figshare [46], SARTAJ [47], and BR35H [48]. It consists of 7023 MRIs of the human brain
provided
Brain Sci. 2023, 13, 1320 in grayscale and jpg format. The dataset includes primary types of brain
6 oftumors,
22
namely glioma, meningioma, pituitary tumors, and images without tumors.

Figureof
Figure 2. Illustration Illustration
2. the of the distribution
distribution of imagesof images
amongamong various class
various class labels throughout
labels the training,
throughout the train-
validation, and testing dataset splits. The bar graph displays the distribution of images across
ing, validation, and testing dataset splits. The bar graph displays the distribution of images across
different classes, with the training set at 64%, the validation set at 16%, and the testing set at 20%.
different classes, with the training set at 64%, the validation set at 16%, and the testing set at 20%.
3.1. Preprocessing
3.1. Preprocessing We implemented a preprocessing framework to improve image quality by integrating
sharpening and Contrast-Limited Adaptive Histogram Equalization (CLAHE) approaches.
We implemented
The processa preprocessing framework
of sharpening commenced to improve
by implementing image quality
a Gaussian by integrat-
blur through the
utilization of a specific technique. The utilization of a 5 × 5
ing sharpening and Contrast-Limited Adaptive Histogram Equalization (CLAHE) ap-kernel was suitable in the
process of attenuating high-frequency noise. The resultant enhanced image was determined
proaches. The using
process of sharpening commenced by implementing a Gaussian blur
the formula:
through the utilization of a specific technique. The utilization of a 5 × 5 kernel was suitable
Sharpened Image = 1.5 × Original Image − 0.5 × Blurred Image (1)
in the process of attenuating high-frequency noise. The resultant enhanced image was de-
termined usingSubsequently,
the formula:
the image underwent a conversion process to grayscale, facilitating a precise
enhancement of contrast. To achieve this, CLAHE was utilized, characterized by an 8 × 8-tile
Sharpened Image = 1.5 × Original Image − 0.5 × Blurred Image
grid and a clip limit of 2.0. Distinct from global histogram equalization, CLAHE adopts a (1)
localized strategy by partitioning the image into discrete tiles and performing individual
Subsequently, the image underwent
equalizations, encapsulated by a conversion process to grayscale, facilitating a pre-
cise enhancement of contrast. To achieve this, CLAHE was utilized, characterized by an
Hlocal (i ) = CLAHE( Htile (i )) (2)
8×8-tile grid and a clip limit of 2.0. Distinct from global histogram equalization, CLAHE
adopts a localized Instrategy
order to ensure accordance with
by partitioning theimage
the specifications
into ofdiscrete
the subsequent
tiles deep
and learning
performing
framework, the enhanced grayscale image was transformed into the RGB color space [49,50].
individual equalizations, encapsulated by
Figure 3 illustrates the several stages of enhancing picture quality, from the initial image to the
CLAHE-enhanced image. This
H local ) = CLAHE
(idepiction showcases
( H tile (ithe
)) effectiveness of our preprocessing (2)
method and its notable impact on improving the overall quality of the image.
In order to ensure accordance with the specifications of the subsequent deep learning
framework, the enhanced grayscale image was transformed into the RGB color space
[49,50]. Figure 3 illustrates the several stages of enhancing picture quality, from the initial
image to the CLAHE-enhanced image. This depiction showcases the effectiveness of our
Brain
BrainSci. 2023,13,
Sci. 2023, 13,1320
x FOR PEER REVIEW 77 of
of 22
23

Figure 3.
Figure 3. Sequential
Sequentialimage
imageimprovement
improvement asas part
part of of
thethe preprocessing
preprocessing framework.
framework. TheThe stages
stages pro-
progress
gress from the unaltered original image through Gaussian blurring for noise suppression, sharpen-
from the unaltered original image through Gaussian blurring for noise suppression, sharpening the
ing the emphasized edge definition to the final enhancement using CLAHE.
emphasized edge definition to the final enhancement using CLAHE.

3.2. Proposed Architecture


3.2. Architecture
Figure 44depicts
Figure depictsthe theproposed
proposed model,
model, which
which acquires
acquires MRI MRIdatadata
withwith
input input dimen-
dimensions
sions
of 224 ×of 224 and reveals its operational characteristics. The model consists of multiple multi-
× 224 and reveals its operational characteristics. The model consists of server
ple server
blocks. blocks. A convolutional
A convolutional layer [51] waslayer [51] wasin
employed employed
the initialinstage,
the initial stage,ofconsisting
consisting 16 filters.
of 16 filter
Each filters.
wasEach filter was
employed withemployed
a kernel size × 3 and
withofa3kernel size of 3 ×size
a stride of 1a ×
3 and stride
1. A size of 1 × 1.
normalizing
A normalizing
layer [52] and a 2D layer [52] and a 2D (two-dimensional)
(two-dimensional) max pooling layermax withpooling 2 × 2with
a size oflayer werea employed
size of 2 ×
to maximize
2 were the information
employed to maximize among thethe intermediate
information layer’s
among output.
the Similarly,layer’s
intermediate we integrated
output.
additional
Similarly, convolutional
we integratedlayers into the
additional model, utilizing
convolutional 32, into
layers 64, 128,
theand 256 filter
model, sizes.32,
utilizing Each
64,
filter utilized in this study had a kernel size of 3 × 3 and a stride size of 1 ×
128, and 256 filter sizes. Each filter utilized in this study had a kernel size of 3 × 3 and a 1, and the same
and valid
stride sizepadding
of 1 × 1,was and suitable for the
the same andexperiment.
valid paddingAs illustrated
was suitablein Figure
for the4, skip connections
experiment. As
were employed
illustrated
Brain Sci. 2023, 13, x FOR PEER REVIEW within
in Figure eachconnections
4, skip block to facilitate the information
were employed within flow
each by concatenating
block 8 of 23the
to facilitate the
outputs
informationof specific
flow by convolutional
concatenating layers. Subsequently,
the outputs a dense
of specific layer of 512layers.
convolutional neurons was
Subse-
employed, accompanied
quently, a dense layer ofby 512global
neuronsaverage pooling andaccompanied
was employed, activation through the rectified
by global averagelinear
pool-
unit (ReLU)
ing and function.
activation through the rectified linear unit (ReLU) function.

Figure 4. Illustration of the proposed architecture and various forms of brain tumors
Figure 4. Illustration of the proposed architecture and various forms of brain tumors.

ToTomitigate
mitigate the issue of overfitting, the dense layer was subjected to regulation using
the issue of overfitting, the dense layer was subjected to regulation using
L1 (10−−55) and L2 (10−4)−regularization techniques [53]. During the training process, the neu-
L1 (10 ) and L2 (10 4 ) regularization techniques [53]. During the training process, the
rons within a dropout layer [54] were randomly deactivated at a rate of 0.5% to enhance
neurons within a dropout layer [54] were randomly deactivated at a rate of 0.5% to enhance
regularization implementation further. Finally, the output layer employed the softmax al-
regularization implementation further. Finally, the output layer employed the softmax
gorithm [51] to compute the probability score for each class and classify whether the input
algorithm [51] to compute the probability score for each class and classify whether the
image exhibited a glioma, meningioma, pituitary, or no tumor. In addition, the model
employed the Adam optimizer [55,56], categorical cross-entropy for loss functions, and
the ReduceLROnPlateau callback to optimize the learning rate [57]. The model was trained
with a batch size of 8 for 30 epochs.
Convolutional neural networks are widely used for image classification tasks. In the
Brain Sci. 2023, 13, 1320 8 of 22

input image exhibited a glioma, meningioma, pituitary, or no tumor. In addition, the model
employed the Adam optimizer [55,56], categorical cross-entropy for loss functions, and the
ReduceLROnPlateau callback to optimize the learning rate [57]. The model was trained
with a batch size of 8 for 30 epochs.
Convolutional neural networks are widely used for image classification tasks. In the
proposed model, 2D convolution involved applying a kernel to the input data to extract
features. The convolution operation captures spatial dependencies and hierarchies within
the data. The convolution operation in a 2D CNN can be mathematically defined as follows:

Yij = ∑m ∑n X(i+m)( j+n) . Kmn (3)

where Yij represents the output element at the position i, j; X(i+m)( j+n) denotes the input
elements at the position (i + m, j + n); and K(mn) signifies the kernel element at the position
(m, n). The equation involves summing the element-wise multiplication of the input element
and corresponding kernel element across the indices m and n. This operation is applied
across the entire input to compute the element of the output feature map. The convolution
operation efficiently captures local patterns and interactions between neighboring elements,
enabling the network to learn the hierarchical representation and extract meaningful
features from the input data. Furthermore, the convolutional operation involved applying
the kernel to input using a sliding window. The kernel size determines the local region
considered, and the stride size controls the movement of the kernel. Padding preserves
spatial dimensions. The output size can be calculated using the following equation.

I − K + 2P
 
O= +1 (4)
S

where O represents the output size, I denotes the input size, K represents the kernel size, S
denotes the stride size, and P represents the padding size [51].

3.2.1. Batch Normalization


Batch normalization (BN) is used in deep neural networks to normalize the inter-
mediate layers’ outputs. It suits internal covariate shifts, improving training, stability,
and performance. In our proposed model, we incorporated the BN layer, following the
skip connections and preceding the Max Pooling layer. The rationale behind this design
was attributed to the function of skip connections, which involves the concatenation of
feature maps originating from distinct layers. Including the BN layer immediately after
ensures that the aggregated feature maps undergo normalization, preserving a uniform
scale and distribution before pooling. In addition to normalization, the positioning of
BN also provides regularization, hence mitigating the potential issue of overfitting and
ensuring that the pooling layer functions on standardized activations. The equation can
represent the normalization process.

x−µ
y= .γ + β (5)
σ
where x is the input; µ and σ; are the mean and standard deviation computed over a mini-batch
size, respectively; and γ and β are learnable scaling and shifting parameters, respectively.

3.2.2. Pooling Layers


The pooling operation is used in a CNN for downsampling, and the input feature map
is divided into non-overlapping regions or pooling windows. The purpose is to calculate
the maximum value of each window, resulting in a downscaled output feature map. The
following equation represents the max pooling operation at the position (i, j) in the output
feature map.
Maxpooling( x )(i, j) = (∀m, n)max( x )(i + m, j + n) (6)
Brain Sci. 2023, 13, 1320 9 of 22

Max pooling ( x )(i, j) denotes the value at the position (i, j) in the output feature map
after max pooling. The term ∀m, n represents the double summation over the indices m
and n and covers all possible values within the pooling windows. max( x )(i + m, j + n)
represents the maximum value among the neighboring elements in the input feature map,
specifically at positions (i + m, j + n). The global average pooling (GAP) operation reduces
the spatial dimension of a feature map while capturing the average representation of the
entire feature map. The GAP can be formulated as follows:

1 k l
k × 1 i∑ ∑ xi,j
Gobal AvgPooling( x ) = (7)
=1 j =1

The equation illustrates the operational mechanism of GAP applied to a feature map
(x). The feature map is characterized by l dimensions for height, width, and channels (k).
The symbol ∑ denotes the mathematical operation of summation and the variables i and j
are employed to iterate through the spatial dimensions of the feature map. The k values in
the resulting vector correspond to the mean activation of the relevant channel across all
spatial positions in the feature map [53].

3.2.3. Activation and Loss Functions


ReLU is an activation function that introduces nonlinearity into a neural network [58].
It takes an input value and returns the maximum value and 0. Mathematically the ReLU
function can be defined as
ReLU ( x ) = max (0, x ) (8)
where x is the input value; if the input value is positive, ReLU outputs the same value. If
the input value is negative, ReLU outputs 0.
The utilization of the softmax function occurs in the output layer of the proposed
model planned for multi-classification tasks. The process converts a vector of real input
values into a probability distribution across different classes. The mathematical expression
for the softmax role is as follows:
exp( xi )
So f tmax ( xi ) = , f or i = 1, 2, 3, 4 (9)
∑4j=1 exp( x j )

The equation xi represents the i-th element of the input vector, and the softmax
function normalizes each probability by dividing it by the sum of the exponential value of
all probabilities in the vector. Furthermore, the loss function was utilized to measure the
discrepancy between the algorithm’s predictions and actual values. Various optimization
techniques can be applied to minimize this error. In addition, categorical cross-entropy was
chosen as the loss function. Categorical cross-entropy can be calculated as the error rate
using the equation.
N
Categorical Cross Entropy = −∑i ytrue [i ].log(y pred [i ]) (10)

where N is the number of classes, ytrue [i ] represents the true class probabilities, and y pred [i ]
denotes the predicted probabilities of each class.

3.2.4. Optimization Techniques


Several regularization strategies were used in the proposed model, including dropout,
L1, L2, and ReduceLROnPlateau callbacks to reduce the overfitting in neural networks.
Dropout arbitrarily changes a small portion of the input units (neurons) to zero during the
training phase [59]. By preventing the network from being overly dependent on particular
units and encouraging generalization, this dropout process aids in the network learning
redundant representations. The model becomes more resilient and enhances its capacity to
perform effectively on unknown data by injecting this unpredictability through the 50%
Several regularization strategies were used in the proposed model, including drop-
out, L1, L2, and ReduceLROnPlateau callbacks to reduce the overfitting in neural net-
works. Dropout arbitrarily changes a small portion of the input units (neurons) to zero
during the training phase [59]. By preventing the network from being overly dependent
Brain Sci. 2023, 13, 1320 on particular units and encouraging generalization, this dropout process aids in the 10 of 22
net-
work learning redundant representations. The model becomes more resilient and en-
hances its capacity to perform effectively on unknown data by injecting this unpredicta-
dropout rate, thereby
bility through the 50%improving its thereby
dropout rate, overall performance. The 50%
improving its overall dropout example
performance. The 50% is
shown
dropoutinexample
Figure 5.is shown in Figure 5.

Figure 5.
Figure 5. The
The right
right side
side of
of the
the diagram
diagram visually
visually depicts
depicts aa dropout
dropout layer
layer characterized
characterized by
by aa dropout
dropout
rate of 50%.
rate of 50%.

L1 and
L1 and L2 L2strategies
strategiesare areemployed
employed in in
thethe
neural network
neural to mitigate
network the issue
to mitigate of over-
the issue of
fitting and enhance the accuracy when activated with novel data
overfitting and enhance the accuracy when activated with novel data from the problem from the problem do-
main [60].
domain These
[60]. techniques
These techniques were employed
were employed in the proposed
in the proposed model
model duedueto their effective-
to their effec-
ness among
tiveness amongthe standard
the standardregularization methods.
regularization L1 regularization
methods. L1 regularizationis alsoisknown as Lasso
also known as
regression,
Lasso and L2
regression, regularization
and is known
L2 regularization as weight
is known decaydecay
as weight or ridge regression.
or ridge The cost
regression. The
drives
cost for L1
drives forand L2 can
L1 and L2 becandefined as follows:
be defined as follows:
L1Regularization(LassoRegression) :
L1Regularization( LassoRegression) :
 i=1
N
CostCost
Function = Loss
Function = Funtion+
Loss Funtion +w λi∑iN=1 |wi |
(11)
(11)
L2Regularization (Weight Decay
L2Regularization(Weight Decay or RidgeRegression))::
or RidgeRegression
N 2
Cost Function = Loss Funtion + N λ ∑ i =1 wi
Cost Function = Loss Funtion+  i=1 wi2
where λ is the hyperparameter that regulates the strength of regularization, N is denoted as
where 𝜆 is the hyperparameter that regulates the strength of regularization, N is denoted
the model factors, wi embodies i-th parameters, and ∑ denotes the sum of all parameters.
as the model factors, 𝑤𝑖 embodies i-th parameters, and ∑ denotes the sum of all parame-
The cost function combines the loss, representing the error between predicted and target
ters. The cost function combines the loss, representing the error between predicted and
values, with a regularization term to form the overall objective function.
target values, with a regularization term to form the overall objective function.
In the proposed model, we utilized the ReduceLROnPlateau from Keras [61]. This
In the proposed model, we utilized the ReduceLROnPlateau from Keras [61]. This
callback is crucial in reducing the learning rate (LR) during the model training phase,
callback is crucial in reducing the learning rate (LR) during the model training phase, spe-
specifically when validation losses showed no further improvement. Incorporating this
cifically when validation losses showed no further improvement. Incorporating this
callback enabled the optimization process to take smaller steps toward minimizing the
callback enabled the optimization process to take smaller steps toward minimizing the
loss function, resulting in a more efficient model. During the training phase, the ReduceL-
loss function, resulting in a more efficient model. During the training phase, the ReduceL-
ROnPlateau callback monitored the chosen metric, such as validation loss. The system
ROnPlateau
recorded callbackobserved
the optimal monitored thefor
value chosen metric,
this metric andsuch as validation
assessed whether loss. The system
the current value
demonstrated improvement over a predetermined number of epochs. If the the
recorded the optimal observed value for this metric and assessed whether current
monitored
value demonstrated
metric did not exhibitimprovement
improvement,over a predetermined
the callback triggerednumber
a reductionof epochs. If the moni-
in the learning rate.
tored metric did not exhibit improvement, the callback triggered a
We employed a factor that was set while configuring the ReduceLROnPlateau callbacks to reduction in the learn-
ing rate.the
achieve Welearning
employed ratea reduction.
factor that Inwas theset while configuring
proposed model, we the ReduceLROnPlateau
initially set the learning
rate to 0.001 and utilized a reduction factor (F) of 0.4; the new learning ratewe
callbacks to achieve the learning rate reduction. In the proposed model, (Newinitially
LR) canset
be calculated by applying the given equation.

New LR = LR × F (12)

3.3. Pre-Trained Model


Pre-trained neural networks are ML models that have undergone training on exten-
sive datasets like ImageNet, consisting of various images belonging to various classes.
Pre-trained models have proven highly advantageous in various tasks, including image
Brain Sci. 2023, 13, 1320 11 of 22

classification and object detection. Pre-trained models are employed because of their abil-
ity to graph data patterns, allowing them to be used as a starting point for new tasks
without having to start the training process from scratch. This investigation included five
pre-trained models, namely VGG16, ResNet50, MobileNetV2, InceptionV3, and VGG19.

3.3.1. VGG16
The VGG16 model was initially presented in 2014 by Simonyan and Zisserman [62],
scholars affiliated with the Visual Geometry Group at the University of Oxford. The
architectural design incorporates filters of dimensions 3 × 3, a stride of 1, and 16 layers,
consisting of three fully connected layers and thirteen convolutional layers. The maximum
pooling layers employ pooling windows with dimensions of 2 by 2 and a stride of 2. VGG16,
a widely recognized choice for efficient feature extraction in transfer learning, boasts a
substantial parameter count of 138 million.

3.3.2. ResNet50
Deep neural networks demonstrate improved performance as their depth increases,
as evidenced in the literature [63]. The challenges related to this improvement arise from
vanishing or exploding gradients, manifesting as the neural network expands. To overcome
this impediment, the authors of [64] have proposed ResNet50, an innovative approach
that utilizes residual modules to facilitate the learning of residual mapping instead of
conventional input–output mapping. This innovative approach involves incorporating
the input into the output of the modules through shortcut connections that circumvent
certain levels. Consequently, including residual blocks effectively mitigates the problem of
vanishing gradients, thereby preventing a decline in performance as the network depth
increases. The ResNet50 architecture incorporates convolutional layers of varying filter
sizes (1 × 1, 3 × 3, 1 × 1) within bottleneck blocks interspersed with max pooling and
average pooling layers to facilitate extracting features from the input.

3.3.3. MobileNetV2
The architectural design aims to provide mobile and embedded applications, achiev-
ing a remarkable balance between high accuracy, lightweight computation, and optimal
memory usage. The employed model utilized three primary strategies: the inverted resid-
ual, the linear bottleneck, and the width multiplier parameters. Using convolutional layers
in the inverted residual technique increases network capacity while concurrently reducing
the computational requirements and memory usage. The input is improved by increasing
the number of channels and applying convolution using a small kernel size to achieve
this objective. Subsequently, the resulting output is projected onto a reduced number of
channels. In contrast, linear bottlenecks employ a linear activation function instead of
a non-linear one, aiming to minimize the number of parameters needed. Furthermore,
utilizing width multiplier parameters can adjust the number of channels within a network,
thereby introducing enhanced adaptability [65].

3.3.4. InceptionV3
The InceptionV3 architecture is a CNN that belongs to the inception series. It is
recognized for its significant advancements compared to previous iterations. The proposed
approach employs an advanced design strategy wherein the network’s capacity is expanded
by incorporating multiple kernel sizes at a given level instead of increasing depth through
stacked layers. The proposed methodology employs inception modules, which integrate a
max pooling layer with varying kernel sizes of 1 × 1, 3 × 3, and 5 × 5 to effectively capture a
wide range of features at different scales. The resulting output is obtained by concatenating
the outputs of these layers, which is achieved by including a 1 × 1 convolution layer before
the 3 × 3 and 5 × 5 convolutional layers. This additional layer decreases the number of
input channels and optimizes the utilization of computational resources [66].
Brain Sci. 2023, 13, 1320 12 of 22

3.3.5. VGG19
The VGG19 architecture modified the VGG16 architecture, encompassing nineteen
layers. This included sixteen convolutional layers, three fully connected layers, a compact
filter with dimensions of 3 × 3, and a stride size 1. Additionally, the model incorporated
max pooling layers that employ a pooling of size 2 × 2 and a stride size of 2. With a
parameter count of 144 million, this model surpasses VGG16 in terms of power, although
at the cost of increased computational requirements [62].

4. Experimental Results
This study employed the proposed model to categorize a substantial MRI dataset
comprising 7023 images. The dataset encompassed glioma, meningioma, pituitary cases,
and cases with no tumor. Initially, a preprocessing stage was incorporated to enhance
the feature extraction. In this stage, image enhancement techniques with Gaussian blur
and CLAHE were applied to improve the quality of the images. The dataset was divided
into subsets, namely training, validation, and testing. The dataset was trained using the
Adam optimizer and subsequently assessed through a fivefold cross-validation method.
Algorithm 1 presents the procedure for the training and evaluation process.

Algorithm 1: Training and Evaluation Process with 5-fold Cross-Validation


1. Initialize Metrics List
. final_test_metrics = []
2. Combine Training and Validation sets
. S = N train + N val where S represents the dataset
3. 5-Fold Cross - Validation
. For i in {1, 2, 3, 4, 5}:
3.1. Data Splitting
. Traini = S – Si
. Val i = Si
3.2. Train Model
.Train the model on Traini and validate on Val i
.Setup Callbacks and Optimizer
3.3. Evaluate on Test set (T) where T represents the testing data
.temp_metrics = Model. Evaluate (T)
.Append temp_metrics to final_test_metrics
4. Calculate Average Test Metrics
.Metrics final = 51 ∑5i=1 f inal_test_metrics[i ]
5. Output
. Metrics final contains the average values on the set T

The learning rates were optimized using the ReduceLROnPlateau callbacks, and a
batch size of 8 was utilized. Figure 6 presents the average accuracy and losses of the
model proposed in this study. During the initial stage of training, the graphs display
fluctuations, which can be attributed to the utilization of the ReduceLROnPlateau callback.
The primary objective of this callback is to dynamically modify the learning rate of the
optimizer during the training process, specifically when the loss function reaches a plateau.
After completing 12 epochs, the optimizer demonstrates a gradual convergence toward
an optimal configuration of weights, resulting in diminished fluctuations observed in the
accuracy and loss curves.
plotlib, Sklearn, Keras, and TensorFlow, to enhance the efficiency of data processing and
model development. The computation was performed on an Intel Core i7-7800 CPU oper-
ating at a clock speed of 3.5 GHz. The model training and tuning were managed using an
NVIDIA GeForce GTX 1080 Ti GPU. The selection of Python 3.7 as the primary program-
ming language for this study was based on its comprehensive set of tools for data manip-
Brain Sci. 2023, 13, 1320 ulation, analysis, and visualization. The platform successfully preserved the data 13 em-
of 22
ployed in this study due to its substantial RAM capacity of 32 GB.

Figure 6. Mean
6. Mean accuracy
accuracy andand losses
losses of proposed
of the the proposed
modelmodel
duringduring
5-fold 5-fold cross-validation.
cross-validation. (Left):
(Left): accuracy
mean mean accuracy progression
progression across across training
training folds. folds. (Right):
(Right): corresponding
corresponding mean mean loss trend.
loss trend. This
demonstrates consistent
This demonstrates accuracy
consistent improvement
accuracy and decreasing
improvement loss, highlighting
and decreasing effective
loss, highlighting model
effective
training.
model training.

ModelFurthermore,
Evaluation Matrices
the platform utilized several libraries, such as Numpy, Pandas, Mat-
plotlib, Sklearn, Keras, and TensorFlow,
The suggested framework to enhance
was subjected the efficiency
to a thorough of datawhich
evaluation, processing and
involved
model development. The computation was performed on an Intel Core i7-7800
an analysis of its precision, recall, F1-score, and accuracy. Precision evaluates the model’s CPU op-
erating at a clock speed of 3.5 GHz. The model training and tuning were managed
ability to minimize the misclassification of negative examples as positive, and the term “is using
an NVIDIA
derived from” GeForce GTX
refers to the 1080 Ti GPU.
calculation of The selection
a specific of Python
metric, which is 3.7obtained
as the primary pro-
by dividing
gramming language for this study was based on its comprehensive set of tools
the number of true positives by the sum of true positives and false positives. However, it for data
manipulation,
is important toanalysis,
note thatand visualization.
recall The measures
is a metric that platform successfully preserved
the model’s capacity tothe data
classify
employed in this study due to its substantial RAM capacity of 32 GB.
the appropriate tumor type accurately. This is calculated by dividing the number of true
positives by the sum
Model Evaluation of true positives and false negatives. The F1-score is a metric used in
Matrices
evaluation that quantifies the balance between precision and recall. It is calculated as the
The suggested framework was subjected to a thorough evaluation, which involved
harmonic mean of precision and recall, obtained by multiplying precision and recall and
an analysis of its precision, recall, F1-score, and accuracy. Precision evaluates the model’s
dividing the result by their sum, multiplied by two. In the context of classification models,
ability to minimize the misclassification of negative examples as positive, and the term “is
accuracy measures the model’s overall performance by quantifying the proportion of cor-
derived from” refers to the calculation of a specific metric, which is obtained by dividing
rect classifications. It is calculated by dividing the number of accurate predictions by the
the number of true positives by the sum of true positives and false positives. However, it
total number of predictions made. Equations (13)–(16) indicate the mathematical repre-
is important to note that recall is a metric that measures the model’s capacity to classify
sentations of precision, recall, F1-score, and accuracy [67].
the appropriate tumor type accurately. This is calculated by dividing the number of true
positives by the sum of true positives and false negatives. The F1-score is a metric used
in evaluation that quantifies the balance between precision and recall. It is calculated as
the harmonic mean of precision and recall, obtained by multiplying precision and recall
and dividing the result by their sum, multiplied by two. In the context of classification
models, accuracy measures the model’s overall performance by quantifying the proportion
of correct classifications. It is calculated by dividing the number of accurate predictions
by the total number of predictions made. Equations (13)–(16) indicate the mathematical
representations of precision, recall, F1-score, and accuracy [67].

TP
Precision = (13)
TP + FP

TP
Recall = (14)
TP + FN

Recall × Precision
F1 − Score = 2 × (15)
Recall + Precision
Brain Sci. 2023, 13, 1320 14 of 22

TP + TN
Accuracy = (16)
TP + TN + FP + FN
The evaluation results, including the average precision, recall, F1-score, and accuracy
for both the proposed and pre-trained models, are presented in Table 1. The suggested
framework demonstrated a notable accuracy rate of 97.84%. Moreover, it achieved precision
and recall values of 97.85% and an F1-score of 97.90%. On the contrary, the InceptionV3
model exhibited the lowest performance, achieving an accuracy of 88.15%, a precision rate
of 87.70%, a recall rate of 87.89%, and an F1-score rate of 87.60%. The observed variation in
the performance of InceptionV3 can be ascribed to its utilization of multiple and parallel
modules, which may not be well suited for the specific characteristics of this dataset, as
supported by our research findings. The pre-trained models VGG16, ResNet50, and VGG19
exhibited superior performance compared to MobileNetV2. Furthermore, the pre-trained
models employed the standard input dimensions, including VGG16, VGG19, ResNet50,
and MobileNetV2 with dimensions of 224 × 224 and InceptionV3 with dimensions of
299 × 229. In order to preserve the pre-existing weights, the layers of the base model were
designated as non-trainable.

Table 1. Evaluation results of proposed and pre-trained models.

Total Precision Recall F1-Score Accuracy Testing Time


Models Name
Params: Average (%) Average (%) Average (%) Average (%) Average (s)
VGG16 14,979,396 95.00 94.85 94.90 95.00 2.29
ResNet50 24,638,852 94.59 94.64 94.55 94.75 1.91
InceptionV3 55,362,340 87.70 87.89 87.60 88.15 2.61
MobileNetV2 2,915,908 91.65 91.40 91.60 91.73 0.99
VGG19 20,289,092 94.80 94.65 94.70 94.83 2.64
Proposed
1,708,356 97.85 97.85 97.90 97.84 0.83
Model

The utilization of the confusion matrix is a fundamental assessment instrument for


classification models [68]. The proposed network demonstrated robust capabilities in
accurately classifying various types of brain tumors, effectively identifying each type
during the examination. Figure 7 presents a visual representation of the results obtained
from the testing data, enabling a comparison between the proposed and pre-trained models.
The comparison reveals that the proposed model outperformed the pre-trained models
in performance. The proposed model demonstrated high accuracy in predicting glioma,
achieving 97%, and meningioma, achieving a 96% accuracy rate. Additionally, it achieved
a 99% accuracy rate in predicting pituitary and no-tumor cases. These results surpass the
performance of pre-trained models. However, it is crucial to emphasize that the efficacy of
treatment for glioma and meningioma in this study did not achieve comparable levels of
success. This finding underscores the necessity for additional research and investigation in
subsequent studies.
Furthermore, the Receiver Operating Characteristics (ROC) curve is a visual rep-
resentation of the performance of a classification model across different classification
thresholds [69]. The True Positive Rate (TPR) and False Positive Rate (FPR) are graphically
represented. The ROC curve illustrates the balance between correctly identifying positive
and incorrectly classifying negative instances as positive at all classification thresholds on
the testing set. The ROC curve provides insights into the model’s ability to differentiate
between different thresholds effectively.
Brain Sci. 2023, 13, 1320
Brain Sci. 2023, 13, x FOR PEER REVIEW
15 of 22
16 of 23

Figure 7. Confusion matrices of several models using the testing data. (a) The proposed model has
Figure 7. Confusion matrices of several models using the testing data. (a) The proposed model has a
a high level of accuracy, achieving a score of 97.84%. (b) VGG16 model achieved a classification
high level of accuracy, achieving a score of 97.84%. (b) VGG16 model achieved a classification accuracy
of 95.00%. (c) ResNet50 model achieved an accuracy of 94.75%. (d) The accuracy of InceptionV3 is
88.15%. (e) MobileNetV2 model achieved a classification accuracy of 91.73%. (f) VGG19 model achieved
a classification accuracy of 94.83%.
Furthermore, the Receiver Operating Characteristics (ROC) curve is a visual repre-
sentation of the performance of a classification model across different classification thresh-
olds [69]. The True Positive Rate (TPR) and False Positive Rate (FPR) are graphically rep-
resented. The ROC curve illustrates the balance between correctly identifying positive and
Brain Sci. 2023, 13, 1320 16 of 22
incorrectly classifying negative instances as positive at all classification thresholds on the
testing set. The ROC curve provides insights into the model’s ability to differentiate be-
tween different thresholds effectively.
The
Thepresent
presentinvestigation
investigationdemonstrates
demonstrates the proposed
the proposed framework’s
framework’ssuperior
superiordiagnostic
diagnos-
efficacy
tic efficacy compared to pre-trained designs. The findings of this study provideevidence
compared to pre-trained designs. The findings of this study provide evidence
supporting
supportingthe thesuggested
suggestedmodel’s
model’shigher
higherdiagnostic
diagnosticaccuracy
accuracycompared
comparedto tostate-of-the-
state-of-the-
art
art methodologies. When comparing the performance of the VGG16 architecture,ititwas
methodologies. When comparing the performance of the VGG16 architecture, was
observed
observedthat
thatititachieved
achievedscores
scoresofof0.95
0.95for
forglioma,
glioma,0.93
0.93for
formeningioma,
meningioma,0.970.97for
forpituitary,
pituitary,
and 0.98 for the no-tumor category. The ResNet50 architecture achieved classification scores
and 0.98 for the no-tumor category. The ResNet50 architecture achieved classification
of 0.92, 0.93, 0.97, and 0.98 for the glioma, meningioma, pituitary, and no-tumor classes,
scores of 0.92, 0.93, 0.97, and 0.98 for the glioma, meningioma, pituitary, and no-tumor
respectively. The InceptionV3 model yielded predictive scores of 0.84 for glioma, 0.81 for
classes, respectively. The InceptionV3 model yielded predictive scores of 0.84 for glioma,
meningioma, 0.96 for pituitary, and 0.97 for the no-tumor category. The MobileNetV2
0.81 for meningioma, 0.96 for pituitary, and 0.97 for the no-tumor category. The Mo-
design achieved scores of 0.90, 0.86, 0.97, and 0.98 for the glioma, meningioma, pituitary,
bileNetV2 design achieved scores of 0.90, 0.86, 0.97, and 0.98 for the glioma, meningioma,
and no-tumor categories, respectively. Additionally, the VGG19 architecture demonstrated
pituitary, and no-tumor categories, respectively. Additionally, the VGG19 architecture
classification scores of 0.92 for glioma, 0.93 for meningioma, 0.98 for the pituitary, and 0.98
demonstrated classification scores of 0.92 for glioma, 0.93 for meningioma, 0.98 for the
for the no-tumor category.
pituitary, and 0.98 for the no-tumor category.
The model under consideration demonstrates notable performance regarding ROC
The model under consideration demonstrates notable performance regarding ROC
scores. The achieved classification accuracies are as follows: 0.98 for glioma, 0.97 for
scores. The achieved classification accuracies are as follows: 0.98 for glioma, 0.97 for men-
meningioma, 0.99 for pituitary, and a flawless accuracy of 1.00 for the no-tumor category.
ingioma,
The robust0.99 for pituitary,
performance and
of the a flawless
model accuracybyofa1.00
is supported for theROC
collective no-tumor category.
score of 98.50%,The
as
robust performance of the model is supported
depicted in Figure 8, compared to pre-trained models. by a collective ROC score of 98.50%, as
depicted in Figure 8, compared to pre-trained models.

Figure8.8.Illustration
Figure Illustrationofofaacomprehensive
comprehensivevisual
visualrepresentation
representationthat
thatcompares
comparesthe
theproposed
proposedmodel’s
model’s
overallROC
overall ROCscore
scorewith
withother
otherpre-trained
pre-trainedmodels.
models.

5.5.Discussion
Discussion
This
Thisinvestigation
investigationintroduces
introducesa novel methodology
a novel methodology for categorizing the Msoud
for categorizing dataset,
the Msoud da-
which consists of a varied assortment of 7023 brain images. The efficacy of the proposed
taset, which consists of a varied assortment of 7023 brain images. The efficacy of the pro-
system is demonstrated
posed system by its capacity
is demonstrated to attain
by its capacity highlyhighly
to attain precise prediction
precise outcomes,
prediction sur-
outcomes,
passing prior research endeavors with comparable aims. Moreover, this study proposes a
method that does not rely on segmenting brain tumor images for classification purposes.
The primary advantage of our approach resides in its capacity to substantially diminish
the requirement for manual procedures, such as feature extraction and tumor localization.
These processes are not only time-intensive but also susceptible to inaccuracies. By em-
ploying various enhancement techniques, including sharpening with Gaussian blur and
Contrast-Limited Adaptive Histogram Equalization (CLAHE), notable enhancements are
achieved in the quality of the brain images. The enhancement process plays a crucial role
in the refinement of edges and improving the overall image clarity, reducing the manual
effort needed for feature extraction.
Brain Sci. 2023, 13, 1320 17 of 22

Furthermore, our proposed model incorporates distinctive concatenation concepts


within the convolutional layers, demonstrating superior performance compared to alterna-
tive methods, as shown in Table 2. By incorporating these enhancement techniques, the
proposed model has demonstrated exceptional performance, surpassing the existing state-
of-the-art model in classifying brain tumors. The successful accomplishment is evidence
of the proposed model’s resilience and capacity to apply to a wide range of brain image
classification tasks, highlighting its potential for achieving precise and dependable results.
Integrating decreased manual intervention, enhanced image quality, and the suggested
model architecture renders our approach highly promising for practical implementations
in classifying brain tumors.

Table 2. Comprehensive comparison of the obtained and previous studies’ results.

Authors Year Methods Dataset Classes Precision Recall F1-Score Accuracy


Hybrid Figshare
Gumaei et al. [25] 2019 3 X X X 94.23
PCA-NGIST-RELM 3064 Images
Figshare
Swati et al. [40] 2019 VGG16 Fine tune 3 89.17 X 91.50 94.65
3064 Images
Figshare
Swati et al. [40] 2019 VGG19 Fine Tune 3 89.52 X 91.73 94.82
3064 Images
Figshare
Ghassemi et al. [32] 2019 CNN-based GAN 3 95.29 X 95.10 95.60
3064 Images
Figshare
Huang et al. [31] 2020 CNNBCN 3 X X X 95.49
3064 Images
HDWT-HOG- Figshare
Fouad et al. [27] 2020 3 X X X 96.40
Bagging 3064 Images
Figshare
Kaplan et al. [19] 2020 NLBP-αLBP-KNN 3 X X X 95.56
3064 Images
Figshare
Ayadi et al. [28] 2020 DSURF-HOG -SVM 3 X 88.84 89.37 90.27
3064 Images
InceptionV3 Figshare
Noreen et al. [34] 2021 3 93.00 92.00 92.00 94.34
Ensemble 3064 Images
Kaggle
Almalki et al. [38] 2022 SURF-KAZE-SVM 4 X X X 95.33
2870 Images
Benchmark BRATS 2015
Ekong et al. [41] 2022 Bayesian-CNN 4 94 95 94 94.32
4000 Images
Kaggle
Asiri et al. [42] 2023 GAN-Softmax 4 92 93 93 96.00
2870 Images
Figshare, SARTAJ and
Shilaskar et al. [43] 2023 HOG-XG Boost 4 92.07 91.82 91.85 92.02
Br35H 7023 images
Figshare, SARTAJ and
Image Enhancement
Our work - Br35H 4 97.85 97.85 97.90 97.84
+ Proposed Model
7023 images

The methodology of Gumaei et al. [25] introduced a combination of PCA, NGIST, and
RELM. While this hybrid approach attempted to capture a comprehensive feature set, PCA
might not always capture non-linear patterns inherent in brain images, potentially missing
crucial tumor-specific details and resulting in less accuracy. The methodologies of Swati
et al. [40] and Noreen et al. [34] relied on refining generic architectures, specifically state-of-
the-art models. Such fine-tuning of deep architectures can be resource-intensive. The intri-
cate process necessitates substantial computational resources and proves time-consuming,
given the need to adjust many parameters in these extensive networks. Contrarily, our
model is purposefully designed for brain tumor classification. It captures tumor-specific
attributes efficiently without the excessive computational demands typically associated
with deep architectures. As corroborated by Table 1, our method requires fewer parameters
than the state of the art and delivers faster testing times.
Brain Sci. 2023, 13, 1320 18 of 22

Ghassemi et al. [32] ventured into the territory of Generative Adversarial Networks,
leveraging CNN-based GANs. While GANs are adept at generating synthetic images, their
direct application to classification might introduce synthetic nuances that deviate from
real-world MRI variations, potentially affecting classification accuracy. Huang et al. [31]
introduced the CNNBCN, a model rooted in randomly generated graph algorithms, achiev-
ing an accuracy of 95.49% and demonstrating advancements in neural network design.
In contrast, our methodology performs superior classification on extensive tumor and
no-tumor images.
Techniques like HDWT-HOG-Bagging and NLBP-αLBP-KNN, as presented by
Fouad et al. [27] and Kaplan et al. [19], rely heavily on traditional feature extraction.
While computationally intensive, such methods might still miss subtle details and patterns
in the MRI scans, resulting in less accuracy. Ayadi et al. [28] employed DSURF-HOG com-
bined with SVM for classification, a method that might overlook hierarchical and spatial
patterns in MRI images, which deep learning models can capture more effectively.
Ekong et al. [41] introduced a Bayesian-CNN approach, and while Bayesian meth-
ods offer probabilistic insights, they might not always capture the intricate features of
brain tumors. While the GAN-Softmax approach by Asiri et al.’s [42] model offers certain
advancements, it is computationally more demanding. Moreover, the efficacy of method-
ologies such as HOG-XG Boost by Shilaskar et al. [43] and the SURF-KAZE technique
by Almalki et al. [38] might be constrained, particularly in their ability to capture spatial
and hierarchical MRI patterns—areas where contemporary deep learning models exhibit
proficiency as proved in this study.

Limitations
The usefulness of the proposed methodology for extracting features has been proven
by using a specific dataset obtained from MRI scans. In order to enhance the clarity of
the images, various techniques for image enhancement were employed. Although these
strategies can enhance visibility, it is crucial to acknowledge that, in specific circumstances,
it may impact classification accuracy. Therefore, comprehensive evaluations are necessary
to test the method’s suitability for different imaging modalities and clinical scenarios and
its flexibility for image enhancements.

6. Conclusions
The present study introduced a novel approach to classify various categories of brain
tumors, such as primary, meningioma, pituitary, and instances with no tumor. This is
achieved by combining image enhancement techniques, namely, Gaussian-blur-based
sharpening and Contrast-Limited Adaptive Histogram Equalization (CLAHE), with a pro-
posed convolutional neural network. The findings of our study demonstrate a remarkable
level of accuracy, specifically 97.84%, which was achieved through a diligent evaluation
of the effectiveness of the suggested framework. The outcome of this study showcases
the model’s robust capacity for generalization, rendering it a valuable and dependable
tool within the medical field. The capacity of this method to facilitate expeditious and
accurate decision making by medical professionals in the realm of brain tumor diagnosis is
evident. To enhance patient care in the future, we intend to revolutionize medical imaging
methods. This will be accomplished by creating real-time brain tumor detection systems
and establishing three-dimensional networks to analyze other medical images.

Author Contributions: Conceptualization, Z.R.; data curation, M.Z.K.; formal analysis, Y.-K.M.;
funding acquisition, I.U., F.A. and A.M.S.; investigation, Y.Y.G.; methodology, Z.R.; project adminis-
tration, I.U., F.A. and A.M.S.; resources, Y.Y.G., M.Z.K., A.A. and A.M.S.; software, Z.R.; supervision,
Y.-K.M.; validation, Z.R., I.U., M.A.K. and A.A.; visualization, M.A.K. and F.A.; writing—original
draft, Z.R.; writing—review and editing, Y.-K.M. and I.U. All authors have read and agreed to the
published version of the manuscript.
Brain Sci. 2023, 13, 1320 19 of 22

Funding: This work was funded by the Researchers Supporting Project Number (RSP2023R509),
King Saud University, Riyadh, Saudi Arabia.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: The data will be available on reasonable request from the corresponding
author.
Acknowledgments: This work was funded by the Researchers Supporting Project Number
(RSP2023R509), King Saud University, Riyadh, Saudi Arabia.
Conflicts of Interest: The authors declare no conflict of interest.

Abbreviations

DE Differential Evolution AI Artificial Intelligence


SVM Support Vector Machine KNN K-Nearest Neighbors
WSVM Weight Kernel Width SVM DL Deep learning
HIK-SVM Histogram Intersection Kernel SVM ML Machine learning
HDWT Haar Discrete Wavelet Transforms MRI Magnetic Resonance Imaging
HOG Histogram of Oriented Gradients LPB Local Binary Pattern
MODE Multi-Objective Differential Evolution SURF Speeded Up Robust Feature
GAN Generative Adversarial Network WOA Whale Optimization Algorithm
CNNBCN Convolutional Neural Network based on Complex Network PCA Principal Component Analysis
RELM Regularized Extreme Learning Machine CNN Convolutional Neural Network
CLAHE Contrast-Limited Adaptive Histogram Equalization CPU Central Processing Unit

References
1. Khazaei, Z.; Goodarzi, E.; Borhaninejad, V.; Iranmanesh, F.; Mirshekarpour, H.; Mirzaei, B.; Naemi, H.; Bechashk, S.M.; Darvishi,
I.; Ershad Sarabi, R.; et al. The association between incidence and mortality of brain cancer and human development index (HDI):
An ecological study. BMC Public Health 2020, 20, 1696. [CrossRef]
2. GLOBOCAN. The Global Cancer Observatory—All Cancers. Int. Agency Res. Cancer—WHO 2020, 419, 199–200. Available online:
https://gco.iarc.fr/today/home (accessed on 12 February 2023).
3. Johns Hopkins Medicine. Gliomas. Available online: https://www.hopkinsmedicine.org/health/conditions-and-diseases/
gliomas (accessed on 12 February 2023).
4. Mayo Clinic. Pituitary Tumors—Symptoms and Causes. Available online: https://www.mayoclinic.org/diseases-conditions/
pituitary-tumors/symptoms-causes/syc-20350548 (accessed on 12 February 2023).
5. Johns Hopkins Medicine. Meningioma. Available online: https://www.hopkinsmedicine.org/health/conditions-and-diseases/
meningioma (accessed on 12 February 2023).
6. Merck Manuals Consumer Version. Overview of Brain Tumors—Brain, Spinal Cord, and Nerve Disorders. Available online:
https://www.merckmanuals.com/home/brain,-spinal-cord,-and-nerve-disorders/tumors-of-the-nervous-system/overview-
of-brain-tumors (accessed on 17 May 2022).
7. American Brain Tumor Association. American Brain Tumor Association Mood Swings and Cognitive Changes. 2014. Available
online: https://web.archive.org/web/20160802203516/http://www.abta.org/brain-tumor-information/symptoms/mood-
swings.html (accessed on 11 December 2022).
8. Tiwari, A.; Srivastava, S.; Pant, M. Brain tumor segmentation and classification from magnetic resonance images: Review of
selected methods from 2014 to 2019. Pattern Recognit. Lett. 2020, 131, 244–260. [CrossRef]
9. Iwendi, C.; Khan, S.; Anajemba, J.H.; Mittal, M.; Alenezi, M.; Alazab, M. The use of ensemble models for multiple class and
binary class classification for improving intrusion detection systems. Sensors 2020, 20, 2559. [CrossRef] [PubMed]
10. Ahmad, S.; Ullah, T.; Ahmad, I.; Al-Sharabi, A.; Ullah, K.; Khan, R.A.; Rasheed, S.; Ullah, I.; Uddin, M.N.; Ali, M.S. A Novel
Hybrid Deep Learning Model for Metastatic Cancer Detection. Comput. Intell. Neurosci. 2022, 2022, 8141530. [CrossRef] [PubMed]
11. Zhuang, Y.; Chen, S.; Jiang, N.; Hu, H. An Effective WSSENet-Based Similarity Retrieval Method of Large Lung CT Image
Databases. KSII Trans. Internet Inf. Syst. 2022, 16, 2359–2376. [CrossRef]
12. Li, C.; Lin, L.; Zhang, L.; Xu, R.; Chen, X.; Ji, J.; Li, Y. Long noncoding RNA p21 enhances autophagy to alleviate endothelial
progenitor cells damage and promote endothelial repair in hypertension through SESN2/AMPK/TSC2 pathway. Pharmacol. Res.
2021, 173, 105920. [CrossRef]
13. Deng, X.; Liu, E.; Li, S.; Duan, Y.; Xu, M. Interpretable Multi-Modal Image Registration Network Based on Disentangled
Convolutional Sparse Coding. IEEE Trans. Image Process. 2023, 32, 1078–1091. [CrossRef]
Brain Sci. 2023, 13, 1320 20 of 22

14. Zhang, K.; Yang, Y.; Ge, H.; Wang, J.; Lei, X.; Chen, X.; Wan, F.; Feng, H.; Tan, L. Neurogenesis and Proliferation of Neural
Stem/Progenitor Cells Conferred by Artesunate via FOXO3a/p27Kip1 Axis in Mouse Stroke Model. Mol. Neurobiol. 2022, 59,
4718–4729. [CrossRef]
15. Wang, F.; Wang, H.; Zhou, X.; Fu, R. A Driving Fatigue Feature Detection Method Based on Multifractal Theory. IEEE Sens. J.
2022, 22, 19046–19059. [CrossRef]
16. Gao, Z.; Pan, X.; Shao, J.; Jiang, X.; Su, Z.; Jin, K.; Ye, J. Automatic interpretation and clinical evaluation for fundus fluorescein
angiography images of diabetic retinopathy patients by deep learning. Br. J. Ophthalmol. 2022, 1–7. [CrossRef] [PubMed]
17. Xu, H.; Van Der Jeught, K.; Zhou, Z.; Zhang, L.; Yu, T.; Sun, Y.; Li, Y.; Wan, C.; So, K.M.; Liu, D.; et al. Atractylenolide I enhances
responsiveness to immune checkpoint blockade therapy by activating tumor antigen presentation. J. Clin. Investig. 2021, 131,
e146832. [CrossRef] [PubMed]
18. Ao, J.; Shao, X.; Liu, Z.; Liu, Q.; Xia, J.; Shi, Y.; Qi, L.; Pan, J.; Ji, M. Stimulated Raman Scattering Microscopy Enables Gleason
Scoring of Prostate Core Needle Biopsy by a Convolutional Neural Network. Cancer Res. 2023, 83, 641–651. [CrossRef] [PubMed]
19. Kaplan, K.; Kaya, Y.; Kuncan, M.; Ertunç, H.M. Brain tumor classification using modified local binary patterns (LBP) feature
extraction methods. Med. Hypotheses 2020, 139, 109696. [CrossRef]
20. Rathi, V.G.P.; Palani, S. Brain Tumor Detection and Classification Using Deep Learning Classifier on MRI Images. Res. J. Appl. Sci.
Eng. Technol. 2015, 10, 177–187. [CrossRef]
21. Cheng, J.; Huang, W.; Cao, S.; Yang, R.; Yang, W.; Yun, Z.; Wang, Z.; Feng, Q. Enhanced Performance of Brain Tumor Classification
via Tumor Region Augmentation and Partition. PLoS ONE 2015, 10, e0140381. [CrossRef]
22. McBee, M.P.; Awan, O.A.; Colucci, A.T.; Ghobadi, C.W.; Kadom, N.; Kansagra, A.P.; Tridandapani, S.; Auffermann, W.F. Deep
Learning in Radiology. Acad. Radiol. 2018, 25, 1472–1480. [CrossRef]
23. Lu, S.; Yang, J.; Yang, B.; Yin, Z.; Liu, M.; Yin, L.; Zheng, W. Analysis and Design of Surgical Instrument Localization Algorithm.
Comput. Model. Eng. Sci. 2022, 137, 669–685. [CrossRef]
24. Afshar, P.; Plataniotis, K.N.; Mohammadi, A. Capsule Networks for Brain Tumor Classification Based on MRI Images and Coarse
Tumor Boundaries. In Proceedings of the ICASSP 2019—2019 IEEE International Conference on Acoustics, Speech and Signal
Processing (ICASSP), Brighton, UK, 12–17 May 2019; pp. 1368–1372. [CrossRef]
25. Gumaei, A.; Hassan, M.M.; Hassan, M.R.; Alelaiwi, A.; Fortino, G. A Hybrid Feature Extraction Method with Regularized
Extreme Learning Machine for Brain Tumor Classification. IEEE Access 2019, 7, 36266–36273. [CrossRef]
26. Rezaei, K.; Agahi, H.; Mahmoodzadeh, A. A Weighted Voting Classifiers Ensemble for the Brain Tumors Classification in MR
Images. IETE J. Res. 2020, 68, 3829–3842. [CrossRef]
27. Fouad, A.; Moftah, H.M.; Hefny, H.A. Brain diagnoses detection using whale optimization algorithm based on ensemble learning
classifier. Int. J. Intell. Eng. Syst. 2020, 13, 40–51. [CrossRef]
28. Ayadi, W.; Charfi, I.; Elhamzi, W.; Atri, M. Brain tumor classification based on hybrid approach. Vis. Comput. 2020, 38, 107–117.
[CrossRef]
29. Srujan, K.S.; Shivakumar, S.; Sitnur, K.; Garde, O.; Pk, P. Brain Tumor Segmentation and Classification using CNN model. Int. Res.
J. Eng. Technol. 2020, 7, 4077–4080.
30. Tejaswini, G.P.; Sreelakshmi, K. Brain Tumour Detection using Deep Neural Network. Wutan Huatan Jisuan Jishu 2020, XVI, 27–40.
31. Huang, Z.; Du, X.; Chen, L.; Li, Y.; Liu, M.; Chou, Y.; Jin, L. Convolutional Neural Network Based on Complex Networks for
Brain Tumor Image Classification with a Modified Activation Function. IEEE Access 2020, 8, 89281–89290. [CrossRef]
32. Ghassemi, N.; Shoeibi, A.; Rouhani, M. Deep neural network with generative adversarial networks pre-training for brain tumor
classification based on MR images. Biomed. Signal Process. Control 2020, 57, 101678. [CrossRef]
33. Deepak, S.; Ameer, P.M. Automated Categorization of Brain Tumor from MRI Using CNN features and SVM. J. Ambient Intell.
Humaniz. Comput. 2020, 12, 8357–8369. [CrossRef]
34. Noreen, N.; Palaniappan, S.; Qayyum, A.; Ahmad, I.; Alassafi, M.O. Brain Tumor Classification Based on Fine-Tuned Models and
the Ensemble Method. Comput. Mater. Contin. 2021, 67, 3967–3982. [CrossRef]
35. Shaik, N.S.; Cherukuri, T.K. Multi-level attention network: Application to brain tumor classification. Signal Image Video Process.
2022, 16, 817–824. [CrossRef]
36. Ahmad, B.; Sun, J.; You, Q.; Palade, V.; Mao, Z. Brain Tumor Classification Using a Combination of Variational Autoencoders and
Generative Adversarial Networks. Biomedicines 2022, 10, 223. [CrossRef]
37. Alanazi, M.F.; Ali, M.U.; Hussain, S.J.; Zafar, A.; Mohatram, M.; Irfan, M.; Alruwaili, R.; Alruwaili, M.; Ali, N.H.; Albarrak,
A.M. Brain Tumor/Mass Classification Framework Using Magnetic-Resonance-Imaging-Based Isolated and Developed Transfer
Deep-Learning Model. Sensors 2022, 22, 372. [CrossRef] [PubMed]
38. Almalki, Y.E.; Ali, M.U.; Ahmed, W.; Kallu, K.D.; Zafar, A.; Alduraibi, S.K.; Irfan, M.; Basha, M.A.A.; Alshamrani, H.A.; Alduraibi,
A.K. Robust Gaussian and Nonlinear Hybrid Invariant Clustered Features Aided Approach for Speeded Brain Tumor Diagnosis.
Life 2022, 12, 1084. [CrossRef] [PubMed]
39. Kavin Kumar, K.; Dinesh, P.M.; Rayavel, P.; Vijayaraja, L.; Dhanasekar, R.; Kesavan, R.; Raju, K.; Khan, A.A.; Wechtaisong, C.;
Haq, M.A.; et al. Brain Tumor Identification Using Data Augmentation and Transfer Learning Approach. Comput. Syst. Sci. Eng.
2023, 46, 1845–1861. [CrossRef]
40. Swati, Z.N.K.; Zhao, Q.; Kabir, M.; Ali, F.; Ali, Z.; Ahmed, S.; Lu, J. Brain tumor classification for MR images using transfer
learning and fine-tuning. Comput. Med. Imaging Graph. 2019, 75, 34–46. [CrossRef] [PubMed]
Brain Sci. 2023, 13, 1320 21 of 22

41. Ekong, F.; Yu, Y.; Patamia, R.A.; Feng, X.; Tang, Q.; Mazumder, P.; Cai, J. Bayesian Depth-Wise Convolutional Neural Network
Design for Brain Tumor MRI Classification. Diagnostics 2022, 12, 1657. [CrossRef]
42. Asiri, A.A.; Shaf, A.; Ali, T.; Aamir, M.; Usman, A.; Irfan, M.; Alshamrani, H.A.; Mehdar, K.M.; Alshehri, O.M.; Alqhtani, S.M.
Multi-Level Deep Generative Adversarial Networks for Brain Tumor Classification on Magnetic Resonance Images. Intell. Autom.
Soft Comput. 2023, 36, 127–143. [CrossRef]
43. Shilaskar, S.; Mahajan, T.; Bhatlawande, S.; Chaudhari, S.; Mahajan, R.; Junnare, K. Machine Learning Based Brain Tumor
Detection and Classification using HOG Feature Descriptor. In Proceedings of the International Conference on Sustainable
Computing and Smart Systems, ICSCSS, Coimbatore, India, 14–16 June 2023; pp. 67–75.
44. Yadav, S. Analysis of k-fold cross-validation over hold-out validation on colossal datasets for quality classification. In Proceedings
of the 2016 IEEE 6th International Conference on Advanced Computing (IACC), Bhimavaram, India, 27–28 February 2016.
[CrossRef]
45. Nickparvar, M.; Brain_Tumor_MRI Dataset. Kaggle. Dataset. 2021. Available online: https://www.kaggle.com/datasets/
masoudnickparvar/brain-tumor-mri-dataset (accessed on 10 May 2023).
46. Cheng, J.; Brain Tumor Dataset. Figshare. 2017. Available online: https://figshare.com/articles/dataset/brain_tumor_dataset/
1512427 (accessed on 10 May 2023).
47. Kaggle. Brain Tumor Classification (MRI). Available online: https://www.kaggle.com/datasets/sartajbhuvaji/brain-tumor-
classification-mri (accessed on 10 July 2023).
48. Hamada, A. Br35H: Brain Tumor Detection. 2020. Available online: https://www.kaggle.com/datasets/ahmedhamada0/brain-
tumor-detection (accessed on 10 May 2023).
49. Wang, W.; Chen, Z.; Yuan, X. Simple low-light image enhancement based on Weber–Fechner law in logarithmic space. Signal
Process. Image Commun. 2022, 106, 116742. [CrossRef]
50. Wang, Y.; Su, Y.; Li, W.; Xiao, J.; Li, X.; Liu, A.A. Dual-path Rare Content Enhancement Network for Image and Text Matching.
IEEE Trans. Circuits Syst. Video Technol. 2023; Early Access. [CrossRef]
51. Goodfellow, I.; Bengio, Y.; Courville, A. Deep Learning; MIT Press: Cambridge, MA, USA, 2016; pp. 1–10. Available online:
https://www.deeplearningbook.org (accessed on 10 May 2023).
52. Ioffe, S.; Szegedy, C. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Proceedings
of the 32nd International Conference on Machine Learning, Lille, France, 6–11 July 2015; Volume 1, pp. 448–456.
53. Alzubaidi, L.; Zhang, J.; Humaidi, A.J.; Al-Dujaili, A.; Duan, Y.; Al-Shamma, O.; Santamaría, J.; Fadhel, M.A.; Al-Amidie, M.;
Farhan, L. Review of Deep Learning: Concepts, CNN Architectures, Challenges, Applications, Future Directions; Springer International
Publishing: Cham, Switzerland, 2021; Volume 8, ISBN 4053702100444.
54. Bin Tufail, A.; Ullah, I.; Rehman, A.U.; Khan, R.A.; Khan, M.A.; Ma, Y.K.; Hussain Khokhar, N.; Sadiq, M.T.; Khan, R.; Shafiq,
M.; et al. On Disharmony in Batch Normalization and Dropout Methods for Early Categorization of Alzheimer’s Disease.
Sustainability 2022, 14, 14695. [CrossRef]
55. Kingma, D.P.; Ba, J. Adam: A Method for Stochastic Optimization. In Proceedings of the 3rd International Conference on Learning
Representations, San Diego, CA, USA, 7–9 May 2015; Conference Track Proceedings. 2014; pp. 1–15.
56. Robbins, H.; Monro, S. A Stochastic Approximation Method. Ann. Math. Stat. 1951, 22, 400–407. [CrossRef]
57. Rasheed, Z.; Ma, Y.-K.; Ullah, I.; Al Shloul, T.; Bin Tufail, A.; Ghadi, Y.Y.; Khan, M.Z.; Mohamed, H.G. Automated Classification of
Brain Tumors from Magnetic Resonance Imaging Using Deep Learning. Brain Sci. 2023, 13, 602. [CrossRef]
58. Nair, V.; Hinton, G.E. Rectified linear units improve Restricted Boltzmann machines. In Proceedings of the 27th International
Conference on Machine Learning, Haifa, Israel, 21–24 June 2010; Association for Computing Machinery: New York, NY, USA,
2010.
59. Srivastava, N.; Hinton, G.; Krizhevsky, A.; Sutskever, I.; Salakhutdinov, R. Dropout: A Simple Way to Prevent Neural Networks
from Overfitting. J. Mach. Learn. Res. 2014, 15, 1929–1958.
60. Moradi, R.; Berangi, R.; Minaei, B. A Survey of Regularization Strategies for Deep Models. Artif. Intell. Rev. 2020, 53, 3947–3986.
[CrossRef]
61. ReduceLROnPlateau. Available online: https://keras.io/api/callbacks/reduce_lr_on_plateau/ (accessed on 24 May 2023).
62. Simonyan, K.; Zisserman, A. Very deep convolutional networks for large-scale image recognition. In Proceedings of the 3rd
International Conference on Learning Representations, San Diego, CA, USA, 7–9 May 2015. Conference Track Proceedings.
63. Glorot, X.; Bengio, Y. Understanding the difficulty of training deep feedforward neural networks. J. Mach. Learn. Res. 2010, 9,
249–256.
64. He, K.; Zhang, X.; Ren, S.; Sun, J. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA, 27–30 June 2016; Volume 2016, pp. 770–778. [CrossRef]
65. Sandler, M.; Howard, A.; Zhu, M.; Zhmoginov, A.; Chen, L.C. MobileNetV2: Inverted Residuals and Linear Bottlenecks. In
Proceedings of the 2018 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Salt Lake City, UT, USA, 18–23
June 2018; pp. 4510–4520. [CrossRef]
66. Szegedy, C.; Vanhoucke, V.; Ioffe, S.; Shlens, J.; Wojna, Z. Rethinking the Inception Architecture for Computer Vision. In
Proceedings of the 2016 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, Las Vegas, NV, USA,
27–30 June 2016; IEEE Computer Society. pp. 2818–2826.
Brain Sci. 2023, 13, 1320 22 of 22

67. Kuraparthi, S.; Reddy, M.K.; Sujatha, C.N.; Valiveti, H.; Duggineni, C.; Kollati, M.; Kora, P.; Sravan, V. Brain tumor classification
of MRI images using deep convolutional neural network. Trait. Signal 2021, 38, 1171–1179. [CrossRef]
68. Ting, K.M. Confusion Matrix. In Encyclopedia of Machine Learning and Data Mining; Springer: Boston, MA, USA, 2017; p. 260.
[CrossRef]
69. Hajian-Tilaki, K. Receiver operating characteristic (ROC) curve analysis for medical diagnostic test evaluation. Casp. J. Intern.
Med. 2013, 4, 627–635.

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.

You might also like