applied

sciences
Article
Accurate MRI-Based Brain Tumor Diagnosis: Integrating
Segmentation and Deep Learning Approaches
Medet Ashimgaliyev 1 , Bakhyt Matkarimov 1, * , Alibek Barlybayev 1,2 , Rita Yi Man Li 3
and Ainur Zhumadillayeva 1,4, *

1 Faculty of Information Technologies, L.N. Gumilyov Eurasian National University,

Astana 010008, Kazakhstan; [email protected] (M.A.); [email protected] (A.B.)
2 Higher School of Information Technology and Engineering, Astana International University,
Astana 010008, Kazakhstan
3 Department of Economics and Finance, Hong Kong Shue Yan University, Hong Kong, China; [email protected]
4 Department of Computer Engineering, Astana IT University, Astana 010000, Kazakhstan
* Correspondence: [email protected] (B.M.); [email protected] (A.Z.); Tel.: +7-707-715-0588 (B.M.);
+7-7025295999 (A.Z.)

Abstract: Magnetic Resonance Imaging (MRI) is vital in diagnosing brain tumours, offering crucial in-
sights into tumour morphology and precise localisation. Despite its pivotal role, accurately classifying
brain tumours from MRI scans is inherently complex due to their heterogeneous characteristics. This
study presents a novel integration of advanced segmentation methods with deep learning ensemble
algorithms to enhance the classification accuracy of MRI-based brain tumour diagnosis. We conduct
a thorough review of both traditional segmentation approaches and contemporary advancements
in region-based and machine learning-driven segmentation techniques. This paper explores the
utility of deep learning ensemble algorithms, capitalising on the diversity of model architectures
to augment tumour classification accuracy and robustness. Through the synergistic amalgamation
of sophisticated segmentation techniques and ensemble learning strategies, this research addresses
the shortcomings of traditional methodologies, thereby facilitating more precise and efficient brain
tumour classification.
Citation: Ashimgaliyev, M.;
Matkarimov, B.; Barlybayev, A.; Li,
Keywords: ensemble algorithm; image segmentation; magnetic resonance imaging; brain neoplasm;
R.Y.M.; Zhumadillayeva, A. Accurate
neural networks; categorisation
MRI-Based Brain Tumor Diagnosis:
Integrating Segmentation and Deep
Learning Approaches. Appl. Sci. 2024,
14, 7281. https://doi.org/10.3390/
app14167281 1. Introduction
The human brain, the body’s most complex organ, regulates various physiological
Academic Editors: Thomas Lindner
and Qi-Huang Zheng
functions, including sensory integration. Brain tumours (BTs), among the most common
global malignancies, disrupt these functions, leading to severe consequences, including
Received: 12 June 2024 death [1,2]. Average cellular turnover involves programmed cell death and regeneration,
Revised: 29 July 2024 but BTs cause uncontrolled cell proliferation, impairing brain functions. BTs can be malig-
Accepted: 13 August 2024 nant or benign, with symptoms like fever, headaches, and cognitive decline, often leading
Published: 19 August 2024 to fatality [3,4].
The early and accurate detection of BTs is crucial for improved patient outcomes.
Medical imaging modalities like MRI, CT, and PET scans are used. MRI, favoured for
high-resolution images, uses contrast agents like gadolinium to differentiate pathological
Copyright: © 2024 by the authors.
tissues. Computer-aided diagnosis (CAD) systems analyse these tissues for precise BT
Licensee MDPI, Basel, Switzerland.
This article is an open access article
detection [5].
distributed under the terms and
Brain tumours are categorised according to their origin, type, and malignancy, ranging
conditions of the Creative Commons from benign Grade I to aggressive Grade IV, influencing treatment approaches [6–12]. Both
Attribution (CC BY) license (https:// Machine Learning (ML) and Deep Learning (DL) improve the accuracy of brain tumour
creativecommons.org/licenses/by/ classification [1]. While ML methods such as SVM and KNN are effective, they necessitate
4.0/).

Appl. Sci. 2024, 14, 7281. https://doi.org/10.3390/app14167281 https://www.mdpi.com/journal/applsci

 Appl. Sci. 2024, 14, x FOR PEER REVIEW 2 of 2

[6–12]. Both Machine Learning (ML) and Deep Learning (DL) improve the accuracy o
Appl. Sci. 2024, 14, 7281 brain tumour classification [1]. While ML methods such as SVM and KNN 2 ofare
20 effective
they necessitate manual feature extraction. In contrast, DL, mainly through Convolutiona
Neural Networks (CNNs), automates feature learning, enhancing diagnostic precisio
[13,14].
manual feature extraction. In contrast, DL, mainly through Convolutional Neural Networks
Recent
(CNNs), automates advancements
feature in AI, especially
learning, enhancing DL, have
diagnostic revolutionised
precision [13,14]. BT detection. CNN
and
Recent ensemble learning
advancements methods DL,
in AI, especially combine multiple models
have revolutionised to enhance
BT detection. CNNsclassificatio
and ensemble learning
accuracy andmethods combine
robustness multiple
[15]. This papermodels to enhance
explores classification
integrating accuracy
segmentation and ensembl
and robustness [15].techniques
learning This paperfor
explores integrating
MRI-based segmentation
BT classification, and ensemble
discussing learning clinica
advancements,
techniquesimplications,
for MRI-based BT classification, discussing advancements, clinical implications,
and future research directions to improve patient treatment and outcome
and future[16–43].
research directions to improve patient treatment and outcomes [16–43].
An overviewAn of such primary
overview of suchdeep learning-based
primary brain tumour brain
deep learning-based segmentation
tumour tech-
segmentatio
niques shows the effectiveness
techniques shows theof these methods
effectiveness (Figure
of these 1):
methods (Figure 1):

Figure 1. of
Figure 1. Overview Overview
key deepoflearning-based
key deep learning-based brain
brain tumour tumour segmentation
segmentation techniques techniques
[10,12]. [10,12].

These techniques highlight highlight

These techniques the advancements in segmentation
the advancements methods, methods,
in segmentation contributing
contributin
to more accurate
to moreand efficient
accurate andbrain tumour
efficient brainclassification. This paperThis
tumour classification. explores
paperintegrating
explores integratin
these sophisticated segmentationsegmentation
these sophisticated techniques with ensemble learning
techniques strategies, addressing
with ensemble learning strategies
traditionaladdressing
methodologies’ shortcomings,
traditional therebyshortcomings,
methodologies’ facilitating more precise
thereby and efficient
facilitating more precis
brain tumour classification [44].
and efficient brain tumour classification [44].

2. Materials and Methods

2. Materials and Methods
Integrating segmentation methods and
Integrating segmentation deep learning
methods ensemble
and deep algorithms
learning encom-
ensemble algorithm
passes diverse techniques and architectures, each contributing to enhancing MRI-based
encompasses diverse techniques and architectures, each contributing to enhancing MRI
brain tumour classification.
based brain tumour Deep learning models
classification. are at themodels
Deep learning core ofare
thisatintegration, which
the core of this integration
leverage neural networks’ hierarchical representation learning capabilities to extract dis-
which leverage neural networks’ hierarchical representation learning capabilities t
criminative features from MRI data. Some such blocks contain DCNNs, Deep Convolutional
extract discriminative features from MRI data. Some such blocks contain DCNNs, Dee
Neural Networks (DCNNs), convolutional neural networks (CNNs), Recurrent Neural
Convolutional Neural Networks (DCNNs), convolutional neural networks (CNNs
Networks (RNNs), Long Short-Term Memory networks (LSTMs), Deep Neural Networks
Recurrent Neural Networks (RNNs), Long Short-Term Memory networks (LSTMs), Dee
(DNNs), Deep Autoencoders (AEs), and Generative Adversarial Networks (GANs) [45–47].
Neural Networks (DNNs), Deep Autoencoders (AEs), and Generative Adversaria
DCNNs have emerged as the cornerstone of medical image analysis, offering superior
Networks (GANs) [45–47]. DCNNs have emerged as the cornerstone of medical imag
performance in image segmentation and classification tasks. By employing multiple lay-
analysis, offering superior performance in image segmentation and classification tasks. B
ers of convolutional filters, DCNNs can effectively capture spatial hierarchies and learn
employing multiple layers of convolutional filters, DCNNs can effectively capture spatia
complex patterns from MRI data, enabling accurate tumour segmentation and feature
extraction. CNNs, a subset of DCNNs, have demonstrated remarkable success in various
image analysis tasks, including brain tumour classification. By leveraging convolutional
layers and pooling operations, CNNs can extract local features from MRI images, provid-
ing valuable tumour characterisation and classification information. RNNs and LSTM
Appl. Sci. 2024, 14, 7281 3 of 20

networks offer unique advantages in processing sequential data, making them well-suited
for tasks involving temporal dynamics and spatial relationships within MRI sequences.
By modelling sequential dependencies, RNNs and LSTMs can capture temporal patterns
in tumour progression and enhance classification accuracy. DNNs encompass a broad
class of neural network architectures with multiple layers of interconnected nodes. These
networks can learn intricate representations from high-dimensional data, making them
suitable for complex tasks such as brain tumour classification. AEs leverage unsupervised
learning principles to extract latent representations of input data. By learning to reconstruct
input images from compressed representations, deep autoencoders can capture meaningful
features and reduce data dimensionality, facilitating more efficient classification [48]. GANs
offer a novel approach to data generation and representation learning by training two
neural networks simultaneously: a generator and a discriminator. GANs can generate
realistic synthetic images that resemble real MRI data, providing valuable augmentation
for training deep learning models and enhancing their generalisation capabilities [49,50].
Integrating suitable segmentation methods with deep learning-based ensemble al-
gorithms for MRI-based brain tumour classification aims to enhance the accuracy and
robustness of tumour analysis. Here is a brief overview of each component:
1. Segmentation: Suitable segmentation methods aim to accurately identify tumour
regions in MRI images, separating them from healthy tissues. This involves threshold-
ing, region growing, active contours, and machine learning-based approaches [51,52].
The choice of method depends on the tumour’s complexity and the data quality [53].
2. Ensemble algorithms: Deep learning-based algorithms combine multiple models to
create a more accurate and robust classification system. They can be used to train on
different types of data and integrate the predictions of various models to produce a
final result. This approach can help reduce errors and improve the overall accuracy of
the classification task [54].
By combining these two techniques, researchers can create a powerful tool for analysing
MRI images and detecting tumours with greater accuracy and reliability. Deep learning-
based ensemble algorithms utilise the diversity of multiple models to enhance classification
accuracy and robustness by combining their predictions. Techniques such as bagging,
boosting, and stacking combine the output of various models into a collective decision. In
the case of brain tumour classification, convolutional neural networks (CNNs) trained on
segmented regions of the tumour are utilised to extract discriminatory features that are
then fed into ensemble algorithms for final classification [52].
The process of combining segmentation methods with deep learning ensemble algo-
rithms includes the following steps:
Segmentation of tumour areas: Utilize suitable segmentation techniques to identify
tumour regions within MRI images. This involves preprocessing the data, applying appro-
priate segmentation algorithms, and creating segmented masks for the tumour areas [51].
Feature extraction: Extract features from segmented tumour regions using deep learn-
ing models such as CNNs. Train these models on the segmented regions to learn features
that can capture patterns and characteristics specific to brain tumours [52].
Ensemble learning: Use ensemble learning techniques to combine the predictions of
multiple deep learning models. This could involve training different CNN architectures
with different initialization parameters, and then combining their predictions by averaging,
voting, or other ensemble methods [55].
Classification: Use the combined predictions from the ensemble to classify tumours
into different categories. The final classification can be based on the consensus of the
individual models’ predictions, with each model’s contribution weighted according to its
performance or confidence level [53].
By integrating appropriate segmentation techniques with deep learning ensemble
algorithms, researchers seek to enhance the accuracy, dependability, and generalizability
of MRI-driven brain tumour classification models. This integrated strategy enables more
Appl. Sci. 2024, 14, 7281 4 of 20

accurate tumour analysis, facilitating improved treatment planning and patient care in
neuro-oncology [51,56].
The study employed transfer learning with five pre-trained CNN models: AlexNet,
VGG16, GoogleNet, ResNet18, and ResNet50, all originally trained on the ImageNet dataset.
It also provides a detailed discussion of each model’s architecture:
1. AlexNet:
Alex Krzyzewski introduced AlexNet in the 2012 Large-Scale Visual Recognition
Challenge (ILSVRC). It won the competition with a top-five error rate of 15.3%, surpassing
the state-of-the-art models of that time, which had error rates of 10.8% or higher [51].
Initially trained on two GPU machines, AlexNet now requires only one GPU. It is a
relatively shallow network with eight layers, comprising five convolutional layers (Conv)
followed by three fully connected (FC) layers [52].
The architecture includes layers with three different filter sizes (11 × 11, 5 × 5, and
3 × 3), data augmentation, dropout, and max pooling operations [53]. Notably, the tradi-
tional sigmoid activation function SF ( x ) was replaced with the rectified linear unit (ReLU)
activation function to address the vanishing gradient problem associated with sigmoid
functions [53]. This change enhanced the training stability of the model by preventing
the learning process from halting when the gradient approaches zero using the following
equations [52]:
1. Sigmoid function (Equation (1)):

1
SF ( x ) = (1)
1 + exp (− x )
2. Rectified linear unit (ReLU) function (Equation (2)):

Re Lu( x ) = Max (0, x ) (2)

The transition from the sigmoid to the ReLU activation function was due to the
vanishing gradient problem associated with the sigmoid function. The gradients become
very small as inputs move away from zero, causing the learning process to stall. In contrast,
ReLU produces non-zero gradients for positive inputs, which prevents gradient saturation
and speeds up learning [55].
2. VGG16:
VGGNet, developed by the Visual Geometry Group (VGG) Lab at Oxford University
in 2014, secured top positions in the ILSVRC 2014 challenge [53]. Designed by Karen
Simonyan and Andrew Zisserman, VGGNet achieved a top 5% test accuracy score of 92.7%.
This study uses the VGG16 version, which includes 13 convolutional layers and 3 fully
connected layers [52]. It utilises small (3 × 3) convolution filters with a stride size of 1 and
the same padding throughout all convolutional layers [51]. Additionally, it employs (2 × 2)
filters for pooling with a stride size of 2. The default input size for images in VGG16 is
224 × 224 [52].
The VGG16 model is a convolutional neural network (CNN) architecture proposed by
the Visual Geometry Group at the University of Oxford. Its name derives from the fact that
it is composed of 16 convolutional and fully connected layers, resulting in a deep network
architecture. Despite this considerable depth, VGG16 follows a simple and uniform design,
utilising small 3 × 3 convolutional filters with max pooling layers interspersed throughout
the network. While this approach may seem relatively straightforward, VGG16 had a
competitive performance on various image classification benchmarks. Consequently, the
VGG16 architecture has been widely adopted as a baseline model in numerous studies and
applications across the field of computer vision [12].
Appl. Sci. 2024, 14, 7281 5 of 20

3. GoogleNet (Inception v1):

GoogleNet, introduced by Google’s research group in 2014 with the paper “Going
Deeper with Convolutions”, secured first place in the ILSVRC 2014 competition with a
remarkable top 5 error rate of 6.67%. With 22 layers, GoogleNet’s design was inspired
by LeNet. Its architecture uses tiny (1 × 1) convolution filters to reduce the number of
intermediate parameters, lowering the parameter count from 60 million in AlexNet to
4 million [53].
In 2014, researchers at Google developed GoogleNet, also known as Inception v1, a
convolutional neural network (CNN) architecture. The critical characteristic of GoogleNet
is its deep and wide architecture, which features multiple parallel convolutional pathways.
The core innovation of this model was the introduction of inception modules, which per-
form convolutions with multiple filter sizes and concatenate the output feature maps. By
leveraging diverse receptive fields at different scales, GoogleNet achieved competitive
performance using fewer parameters than traditional CNN architectures. The inception
module concept has been further refined in subsequent versions of the model, such as
Inception v2 and Inception v3, to improve efficiency and performance. Some of the other
notable architectural choices in GoogleNet use 1 × 1 filters to limit the number of param-
eters, the incorporation of global average pooling at the end to reduce feature map size
and increase accuracy while decreasing trainable parameters, and use inception modules
that employ fixed convolutions of different sizes (1 × 1, 3 × 3, 5 × 5) along with 3 × 3 max
pooling. This multi-scale approach helps the model effectively manage objects at various
scales. Additionally, GoogleNet introduced intermediate classifier branches, such as the
Auxiliary Classifier, to provide regularisation and address vanishing gradient issues during
training. These innovations have contributed to the strong performance and efficiency of
the GoogleNet family of models across various computer vision tasks [12].
4. ResNet18 and ResNet50:
The Residual Network (ResNet), created by Kaiming and his team at Microsoft Re-
search, was introduced at ILSVRC 2015. It won the classification task with an impressive
3.57% error rate on the ImageNet test set [52]. ResNet is distinguished by its ability to
address the vanishing gradient problem, allowing for the effective training of intense
neural networks.
One of ResNet’s key advantages is its ability to manage deep architectures while
minimising computational complexity and training time. ResNet is available in various
versions, such as ResNet18, ResNet50, and ResNet101 [53].
In this study, ResNet18 and ResNet50 were employed. ResNet18 consists of 18 layers,
while ResNet50 comprises 50 layers. These variations allow flexibility in balancing model
complexity and computational resources, catering to different application needs [52].
ResNet, short for Residual Network, is a family of convolutional neural network
(CNN) architectures introduced by Microsoft Research in 2015. The critical innovation
of ResNet is its ability to address the problem of vanishing gradients in deep networks
by incorporating skip or residual connections. Two prominent variants of the ResNet
architecture are ResNet18 and ResNet50, which have 18 and 50 layers, respectively. These
models utilise residual blocks, where the input is added to the output of the block, allowing
for easier optimisation of very deep networks. This residual connection helps the model
learn the difference or “residual” between the input and the desired output rather than
learning the complete transformation from scratch. The ResNet architectures have demon-
strated superior performance in image classification tasks, particularly on datasets with
many classes or challenging visual patterns. By employing these residual connections,
ResNet models can overcome the vanishing gradient issue when training very deep neural
networks. This leads to improved performance and the ability to leverage the increased
depth of the model effectively [55].
Figure 2 shows a schematic diagram of an artificial intelligence-based system designed
for tumour detection and segmentation in medical images.
 with many classes or challenging visual patterns. By employing these residual
connections, ResNet models can overcome the vanishing gradient issue when training
very deep neural networks. This leads to improved performance and the ability to
leverage the increased depth of the model effectively [55].
Appl. Sci. 2024, 14, 7281 Figure 2 shows a schematic diagram of an artificial intelligence-based system
6 of 20
designed for tumour detection and segmentation in medical images.

Figure 2. Proposed framework of tumour detection and segmentation.

Figure 2. Proposed framework of tumour detection and segmentation.
This diagrammatic representation systematically encapsulates the fusion of convo-
This diagrammatic
lutional neural networks with representation
sophisticated systematically encapsulates
classification algorithms, the fusion
crafting a compre-of
convolutional
hensive medical neural
imagingnetworks
solution. withThissophisticated
integration classification
aims to elevate algorithms,
diagnosticcrafting
precisiona
comprehensive medical imaging solution. This integration
and contribute positively to enhanced healthcare outcomes. The pipeline initiates with aims to elevate diagnostic
precision and of
the acquisition contribute
source medicalpositivelyimages, to which
enhanced healthcare
are subjected to aoutcomes. The pipeline
series of pre-processing
initiates with the acquisition of source medical images, which
techniques. These techniques typically include normalisation to adjust the pixel are subjected to avalues
seriesfor
of
pre-processing techniques. These techniques typically include
uniformity across different images, as well as other image-enhancement methods aimed normalisation to adjust the
pixel values for
at improving theuniformity
visual clarity across
anddifferent images, as well
feature consistency as other
essential image-enhancement
for subsequent analysis.
methods aimed at improving the visual clarity and feature
The conditioned images are then input into a cascade of convolutional layers. Within consistency essential for
subsequent analysis. The conditioned images are then input
each layer, a set of trainable filters executes spatial feature extraction by convolving with into a cascade of
convolutional layers. Within each layer, a set of trainable filters
the image matrix, thereby capturing hierarchical patterns essential for complex pattern executes spatial feature
extraction
recognition. byPost
convolving
convolution, with athe imagelinear
rectified matrix, thereby
unit (ReLU) capturing
activation hierarchical
function ispatterns
applied
essential for complex
to each feature map topattern
introduce recognition.
non-linearPost convolution, which
transformations, a rectified linear unit
are crucial (ReLU)
for learning
activation
non-linearfunction is applied
complexities in imageto each feature
data. map toby
Followed introduce
selectivenon-linear transformations,
convolutional layers, max
which areoperations
pooling crucial forare learning
executednon-linear complexities
to downsample theinspatial
image data. Followed
dimensions of by
theselective
feature
convolutional
maps. This steplayers, maxinpooling
is pivotal reducing operations are executed
the computational demandto downsample
and mitigating the the
spatial
risk
dimensions of the feature maps. This step is pivotal in reducing
of overfitting by emphasising the most salient features and suppressing the less relevant the computational
demand
ones. Theand mitigating
architecture the risk
further of overfitting
includes additionalbyconvolutional
emphasisinglayers the most salientthe
that refine features
depth
and granularity
suppressingofthe less extraction.
feature relevant ones. The architecture
Transposed convolutional further includes
layers, additional
often referred to
as deconvolutional
convolutional layers layers,
that are employed
refine the depthto progressively
and granularityrescale ofthefeature
feature maps back
extraction.
to higher resolutions,
Transposed convolutionala process essential
layers, oftenforreferred
maintainingto as spatial integrity in tasks
deconvolutional suchare
layers, as
image segmentation.
employed The culmination
to progressively rescale theof the neural
feature mapsnetwork’s processing
back to higher is the generation
resolutions, a process
of a segmented image. This output distinctly isolates and delineates the regions of interest,
such as tumours, from the non-relevant background, enabling precise medical evaluations.
Concurrently, the feature maps are exploited in a classification framework, where an SVM
classifier operationalizes an entropy-based feature selection methodology. This approach
rigorously identifies and utilises the most discriminative features for robust classification
of tumour presence. To augment the model’s efficacy, fine-tuning is conducted using
MobileNetV2 (Figure 3), a streamlined deep neural network architected for enhanced com-
putational efficiency on mobile and edge devices. This step aims to optimise the trade-off
 the non-relevant background, enabling precise medical evaluations. Concurrently, the
feature maps are exploited in a classification framework, where an SVM classifier
operationalizes an entropy-based feature selection methodology. This approach
rigorously identifies and utilises the most discriminative features for robust classification
Appl. Sci. 2024, 14, 7281
of tumour presence. To augment the model’s efficacy, fine-tuning is conducted 7using of 20
MobileNetV2 (Figure 3), a streamlined deep neural network architected for enhanced
computational efficiency on mobile and edge devices. This step aims to optimise the trade-
between computational
off between speedspeed
computational and model accuracy.
and model The integrated
accuracy. system outputs
The integrated system both the
outputs
classification results—indicating the presence or absence of tumours—and the segmented
both the classification results—indicating the presence or absence of tumours—and the
images visually
segmented highlighting
images the tumour regions.
visually highlighting the tumour regions.

Figure 3. Fine-tuned
Figure3. Fine-tuned MobileNetV2.
MobileNetV2.

Several
Several keykeyperformance
performancemetrics
metricsare commonly
are commonly used in the
used context
in the of a binary
context clas-
of a binary
sification problem, such as tumour detection. True Positive (TP) represents
classification problem, such as tumour detection. True Positive (TP) represents the the number
of correctly
number of identified
correctly positive
identifiedsamples,
positivemeaning
samples,those with tumours.
meaning False
those with PositiveFalse
tumours. (FP)
refers to the
Positive (FP)number
refers of
to incorrectly
the number identified positive
of incorrectly samples positive
identified or those without
samples tumours.
or those
True Negative (TN) denotes the number of correctly identified negative
without tumours. True Negative (TN) denotes the number of correctly identified samples negative
or those
without
samples tumours. Conversely,
or those without False
tumours. Negative (FN)
Conversely, Falseindicates
Negativethe(FN) number of the
indicates incorrectly
number
identified negative samples or those with tumours. Accuracy, as described in a relevant
of incorrectly identified negative samples or those with tumours. Accuracy, as described
equation, represents the proportion of correctly identified samples in the data. These
in a relevant equation, represents the proportion of correctly identified samples in the
metrics provide a comprehensive assessment of the model’s performance in distinguishing
data. These metrics provide a comprehensive assessment of the model’s performance in
between the presence and absence of tumours in the given dataset [1]. Equations (5)–(9)
distinguishing between the presence and absence of tumours in the given dataset [1].
describe the commonly used evaluation criteria for this study:
Equations (5)–(9) describe the commonly used evaluation criteria for this study:
( T(𝑇𝑝 − 𝑇𝑛)
p − Tn )
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦==( T p + Fp + Tn + Fn)
Accuracy (3)
(3)
(𝑇𝑝 + 𝐹𝑝 + 𝑇𝑛 + 𝐹𝑛)
Precision is
Precision is aa performance
performance metric
metric representing
representing the
the proportion
proportion ofof correctly
correctly identified
identified
positivesamples
positive samplesamong
amongallallthose
thoseidentified
identifiedas
aspositive
positivebybythe
themodel.
model.Recall
Recallisis the
the ratio
ratio of
of
correctlyidentified
correctly identifiedpositive
positivesamples
samplestotothe
thetotal
totalnumber
numberof ofpositive
positivesamples.
samples. The
The F1F1 score
score
is aa composite
is compositemetric
metriccombining
combiningprecision
precisionand
andrecall
recallinto
intoaasingle
singlevalue.
value.

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 T p𝑇𝑝
Precision = = (𝑇𝑝 + 𝐹𝑝) (4)
(4)
( T p + Fp)

Tp
Recall = (5)
( T p + Fn)
2 ( Precision × Recall )
F1 − score = (6)
( Precision + Recall )
The Dice coefficient is like the Intersection over Union (IoU) and ranges from 0 to 1,
with 1 indicating the highest similarity between the predicted segmentation and the ground
truth. This coefficient evaluates the effectiveness of automated probabilistic fractional
Appl. Sci. 2024, 14, 7281 8 of 20

segmentation of MR images and the accuracy of manual segmentations regarding spatial

overlap.
Tp
SC =   (7)
1
2 ( 2T p + Fp + Fn )

Intersection over Union (IoU) is a performance metric used to evaluate the accuracy of
object detection models. It compares the overlap between the predicted positive regions
(including both true positives and false positives) and the actual positive regions (including
both true positives and false negatives) in the dataset. The intersection of these sets
represents the true positive (Tp) detections, while the union encompasses the true positives,
false positives (Fp), and false negatives (Fn).

Tp
oU = (8)
( T p + Fp + Fn)

The Similarity Index (SI) is a performance metric used to assess the accuracy of
tumour detection models. It measures the similarity between the ground truth annotations
representing the actual tumour regions and the model’s segmentation output, which
identifies the predicted tumour regions. A higher SI value indicates a closer match between
the predicted segmentation and the ground truth, signifying a more accurate detection of
the tumour areas.
2T p
SI = (9)
(2T p + Fp + Fn)
These evaluation metrics are essential for assessing the performance of brain tumour
detection and classification models. They measure the models’ ability to correctly identify
positive (tumour-containing) and negative (tumour-free) samples and differentiate between
different types of tumours. Metrics such as True Positive, False Positive, True Negative,
False Negative, Accuracy, Precision, Recall, F1 score, Intersection over Union, and Similarity
Index comprehensively evaluate the model’s performance in various tumour detection and
classification aspects.

3. Dataset
To evaluate the performance of our methods, we used the BRATS 2018 dataset (Menze
et al., 2015; Bakas et al., 2017 a, b, c, 2018). The training set comprised images from
285 patients, including 210 with high-grade gliomas (HGG) and 75 with low-grade gliomas
(LGG). The validation set included MRI scans from 66 patients with brain tumours of an
unknown grade, as predefined by the BRATS challenge organisers. The test set consisted
of images from 191 brain tumour patients. Of these, 77 had undergone Gross Total Resec-
tion (GTR). Their data was used to evaluate the model’s ability to predict survival.Each
patient was scanned using T1, T1Gd, T2, and FLAIR. Each image was skull-stripped and
resampled to a uniform resolution of 1 mm3 in all dimensions. The sequences for each
patient were aligned to ensure consistency. Experts manually created the ground truth
segmentation masks.
The model’s performance was assessed on the validation and test sets using the CBICA
Image Processing Portal, available at ipp.cbica.upenn.edu. Segmentation annotations
included tumour subtypes such as necrotic/non-enhancing tumour (NCR), peritumoral
oedema (ED), and gadolinium-enhancing tumour (ET).
Below is listed the step-by-step process of training deep learning models for brain
tumour classification:
Data Preparation:
1. Dataset collection: We started by collecting a comprehensive dataset of MRI scans,
such as the BRATS (Brain Tumor Segmentation Challenge) series from 2013 to 2019.
This dataset included thousands of MRI images with corresponding annotations
indicating tumour regions.
Appl. Sci. 2024, 14, 7281 9 of 20

2. Image preprocessing: The collected images underwent several preprocessing

steps. This included bias field correction to address intensity inhomogeneities,
intensity normalisation to standardise pixel values, and data augmentation tech-
niques such as rotations, flips, and zooms to increase the dataset’s variability and
improve model robustness.
Segmentation:
3. Tumour segmentation: Suitable segmentation methods were applied to accurately
identify tumour regions within the MRI images. Techniques like thresholding, region
growth, and active contours created segmented masks that separated tumours from
healthy tissues.
Feature Extraction:
4. Feature extraction: Deep learning models extracted features from the segmented tumour
regions, particularly convolutional neural networks (CNNs). CNNs were trained in
these regions to learn distinctive features and patterns specific to brain tumours.
Model Selection and Initialization:
5. Model selection: We chose five pre-trained CNN models—AlexNet, VGG16, GoogleNet,
ResNet18, and ResNet50—each renowned for their effectiveness in image analysis
tasks. These models were initially trained on the ImageNet dataset.
6. Transfer learning: We applied transfer learning to tailor these models specifically for
brain tumour classification. This process involved fine-tuning the pre-trained models
on our MRI dataset, enabling them to utilise pre-learned features while adapting to
the nuances of medical images.
Training Process:
7. Training image pairs: The training involved using pairs of MRI images and their
annotated labels. Approximately 2000 MRI images with corresponding annotations
were used.
8. Training parameters: we experimented with various training parameters, such as
learning rate and batch size, to find the optimal settings for each model.
9. Regularization techniques: regularization methods, including dropout and early
stopping, were employed to prevent overfitting and ensure the models generalise
well to new data.
10. Training duration: the training process was conducted over 72 h on an NVIDIA Tesla
V100 GPU, utilising its computational power to handle extensive data and complex
computations.
Ensemble Learning:
11. Ensemble learning: To enhance classification accuracy and robustness, ensemble
learning techniques like bagging, boosting, and stacking were used. These meth-
ods combined the predictions of multiple models, producing a final, more accurate
classification result.
Evaluation and Fine-Tuning:
12. Model evaluation: The trained models were evaluated using various metrics, in-
cluding accuracy, sensitivity, specificity, and the Dice coefficient index. Based on the
evaluation results, fine-tuning was performed to further improve performance.
We meticulously followed these steps and ensured our models were well-prepared
and optimised for accurate and robust brain tumour classification [3,5,9].

4. Results
Figure 4 presents a sequence of MRI scans illustrating the detection and segmentation
process of a meningioma, a type of brain tumour.
 and optimised for accurate and robust brain tumour classification [3,5,9].

4. Results
Figure 4 presents a sequence of MRI scans illustrating the detection and segmentation
Appl. Sci. 2024, 14, 7281 10 of 20
process of a meningioma, a type of brain tumour.

Figure
Figure 4.
4. Meningioma
Meningioma identification
identification in
in brain
brain MRI.

This diagrammatic
This diagrammaticrepresentation
representation systematically
systematicallyencapsulates
encapsulatesthe fusion
the of convolu-
fusion of
tional neural networks
convolutional with sophisticated
neural networks classification
with sophisticated algorithms,algorithms,
classification crafting a comprehen-
crafting a
sive medical imaging
comprehensive medicalsolution.
imaging This integration
solution. Thisaims to elevate
integration aimsdiagnostic
to elevateprecision and
diagnostic
contribute positively to enhanced healthcare outcomes. The first image
precision and contribute positively to enhanced healthcare outcomes. The first image on on the left displays
a transverse
the section
left displays of a brainsection
a transverse MRI scan.of aItbrain
prominently
MRI scan.shows a large, hyperintense
It prominently mass
shows a large,
in the cerebral tissue, characteristic of a meningioma. The mass
hyperintense mass in the cerebral tissue, characteristic of a meningioma. The mass appears white on the
appears white on the T1-weighted MRI image due to its contrast enhancement, indicatinga
T1-weighted MRI image due to its contrast enhancement, indicating the presence of
dense,
the extra-axial
presence of atumour.
dense,The middle image
extra-axial tumour. is a binary mask highlighting
The middle image is a the segmented
binary mask
highlighting the segmented tumour. This mask simplifies the image to show onlyfrom
tumour. This mask simplifies the image to show only the tumour region, isolated the
the surrounding
tumour brain structures.
region, isolated from the Here, the tumour
surrounding is depicted
brain structures. as Here,
a purethe
white area on
tumour is
a black background,
depicted emphasising
as a pure white area on athe exact
black shape and boundaries
background, emphasisingofthe theexact
meningioma
shape and as
identified byofimage
boundaries processing algorithms.
the meningioma as identifiedThe by final
imageimage on the right
processing reintroduces
algorithms. the
The final
context of the original MRI scan, overlaying a yellow outline that demarcates
image on the right reintroduces the context of the original MRI scan, overlaying a yellow the perimeter
of the tumour. This outline visualises the tumour’s extent and spatial relationship to nearby
outline that demarcates the perimeter of the tumour. This outline visualises the tumour’s
brain structures, crucial for surgical planning or targeted therapies. This series effectively
extent and spatial relationship to nearby brain structures, crucial for surgical planning or
demonstrates the application of image processing techniques in medical imaging, specifi-
targeted therapies. This series effectively demonstrates the application of image
cally for tumour identification, isolation, and data preparation for further clinical analysis
processing techniques in medical imaging, specifically for tumour identification, isolation,
or intervention.
and data preparation for further clinical analysis or intervention.
Integrating suitable segmentation methods with a deep learning-based ensemble
Integrating suitable segmentation methods with a deep learning-based ensemble
algorithm offered several advantages, including improved accuracy, robustness, and
algorithm offered several advantages, including improved accuracy, robustness, and
interpretability—however, challenges such as dataset bias, class imbalance, and com-
interpretability—however, challenges such as dataset bias, class imbalance, and
putational resources needed to be addressed.
computational resources needed to be addressed.
Overall, the results demonstrated that the proposed approach could enhance the
Overall, the results demonstrated that the proposed approach could enhance the
performance of MRI-based brain tumour classification, contributing to improved patient
performance
outcomes andofpersonalised
MRI-based brain tumour
treatment classification, contributing to improved patient
strategies.
outcomes and personalised treatment strategies.
Table 1 compares the specific attributes and characteristics of commonly used pre-
trained CNN architectures, including AlexNet, VGG16, GoogleNet, ResNet18, and ResNet50.
Each architecture is evaluated based on layer count, input size, model description, unique
features, top-five error rates, and the total number of parameters in millions.
Appl. Sci. 2024, 14, 7281 11 of 20

Table 1. Key attributes and characteristics of the pre-trained CNN architectures used.

Attributes AlexNet VGG16 GoogleNet ResNet18 ResNet50

Number of Layers 8 16 22 18 50
Input Size 227 × 227 × 3 224 × 224 × 3 224 × 224 × 3 224 × 224 × 3 224 × 224 × 3
Model Five convolutional 13 convolutional 21 convolutional 17 convolutional 49 convolutional
Description layers, 3 FC layers layers, 3 FC layers layers, 1 FC layer layers, 1 FC layer layers, 1 FC layer
Local Response 1 × 1 Convolution,
Object Localization
Normalization, Global Average
Unique Features and Image Skip Connections Skip Connections
Overlapping Max Pooling, Inception
Classification
Pooling Module
Top-Five
15.3% 7.3% 6.67% 3.57% 3.57%
error rate
Number of
Parameters 60 138 4 11.4 23.9
(millions)

The data from Table 1 underscore the varied design principles and efficiencies inherent
in each CNN architecture, offering insights into their suitability for specific image process-
ing and pattern recognition tasks. The top column indicates the total number of layers in
each model, reflecting their depth: AlexNet has 8 layers, VGG16 has 16, GoogleNet has
22, ResNet18 has 18, and ResNet50 has 50 layers. Except for AlexNet, which processes
images of 227 × 227 pixels, all other models handle images of 224 × 224 pixels with three
colour channels (RGB). The Model Description column outlines the composition of each
architecture in terms of convolutional (Conv) and fully connected (FC) layers. For instance,
AlexNet includes 5 convolutional layers and three fully connected layers, while ResNet50
features 49 convolutional layers and 1 fully connected layer. Each model has unique fea-
tures that enhance its functionality. AlexNet incorporates local response normalization
and overlapping max pooling. VGG16 is noted for its deeper layer structure. GoogleNet
integrates object localization and image classification, employs 1 × 1 convolutions, utilises
global average pooling, and includes an inception module. ResNet architectures (ResNet18
and ResNet50) are characterised by their use of skip connections, which help mitigate the
vanishing gradient problem during training. The Top-Five Error Rate column indicates the
models’ accuracy in classifying the top five predictions, with lower percentages signifying
higher accuracy. ResNet18 and ResNet50 have the lowest error rates at 3.57%, followed
by GoogleNet at 6.67%, VGG16 at 7.3%, and AlexNet at 15.3%. The number of trainable
parameters varies significantly among the models, reflecting their complexity and com-
putational demands. VGG16 has the highest number of parameters at 138 million, while
GoogleNet is the most parameter-efficient with only 4 million.
These CNN architectures are extensively used in various computer vision tasks, such
as image classification, object detection, and semantic segmentation. They are potent tools
for feature extraction and representation learning, facilitating the development of advanced
deep-learning models for various applications [57].
By incorporating these diverse deep learning architectures into ensemble frameworks,
researchers can harness the complementary strengths of individual models and improve
classification performance. Ensemble learning techniques such as bagging, boosting, and
stacking further enhance the robustness and reliability of classification systems by aggregat-
ing predictions from multiple models. Through the synergistic integration of segmentation
methods and ensemble learning, researchers aim to unlock the full potential of MRI-based
brain tumour classification, paving the way for more accurate diagnosis and personalised
treatment strategies in neuro-oncology [44].
We fine-tuned the model parameters in our machine learning process to optimise perfor-
mance and enhance results. For our study, we trained various deep learning models—AlexNet,
VGG16, ResNet18, GoogLeNet, and ResNet50—initially using the ImageNet dataset. Sub-
 integration of segmentation methods and ensemble learning, researchers aim to unlock
the full potential of MRI-based brain tumour classification, paving the way for more
accurate diagnosis and personalised treatment strategies in neuro-oncology [44].
We fine-tuned the model parameters in our machine learning process to optimise
Appl. Sci. 2024, 14, 7281 performance and enhance results. For our study, we trained various deep learning 12 of 20
models—AlexNet, VGG16, ResNet18, GoogLeNet, and ResNet50—initially using the
ImageNet dataset. Subsequently, we carefully conducted additional training on
sequently, we
specialised carefully
datasets conducted
to adapt theseadditional
models fortraining on specialised
the specific datasetsimage
tasks of medical to adapt these
analysis,
models for the
particularly forspecific tasks of medical
the segmentation image analysis,
and classification ofparticularly for the segmentation and
brain tumours.
classification
Figure 5ofshows
brain tumours.
examples of segmentation outcomes contrasted with the ground
Figure
truth. 5 shows examples of segmentation outcomes contrasted with the ground truth.

Figure 5. The segmentation results were visualized and compared to the

the ground
ground truth.
truth.

segmentation results and the ground truth for a

The provided image compares the segmentation
medical image identified as Brats18_TCIA04_343_1_flair. The comparison is organised in
three rows and two columns.
– Left column (ground truth): These images represent the manually labelled ground
truth, indicating the regions of different tissue types within the brain scan. The colours
used are as follows:
– Green: edoema;
– Yellow: non-enhancing solid core;
– Red: enhancing core.
– Right column (prediction result): These images represent the automated segmentation
results produced by a model. The same colour coding is used as in the ground truth
images:
– Green: edoema;
– Yellow: non-enhancing solid core;
– Red: enhancing core.
Appl. Sci. 2024, 14, 7281 13 of 20

Table 2 presents a comprehensive evaluation of the proposed method using various

datasets, including the Brain Tumor Segmentation Challenge (BRATS) series from 2013
to 2019, the Alzheimer’s Disease Neuroimaging Initiative (ADNI), The Cancer Imaging
Archive (TCIA), and BrainWeb. These datasets were employed to train models on a broad
spectrum of realistic neuroimaging data. As detailed in Table 2, the proposed method
exhibits an exemplary performance in tumour detection using the Multi-Support Vector
Machine (M-SVM) classifier, particularly with the BRATS 2018 dataset. Table 3 demonstrates
that the proposed brain tumour detection method performs exceptionally well with M-SVM
classification.

Table 2. Summary of deep learning-based brain tumour segmentation techniques.

Model
Dataset Preprocessing Performance
Architecture
Bias field correction,
BRATS
intensity and patch DSC 88%, SEN 89%,
2013 and Custom CNN
normalisation, PR 87%
2015
augmentation
BRATS Intensity normalisation, SEN 95%, SPE 95.5%, PR 96.5%,
HCNN + CRF-RRNN 1
2013 augmentation RE 97.8%, ACC 98.6%
Residual Network +
BRATS
Z-score normalisation dilated convolution DSC 86%
2015
RDM-Net 2
Stack Multi-connection
BRATS
Z-score normalisation Simple Reducing_Net DSC 83.42%, PR 78.96%, SEN 90.24%
2015
(SMCSRNet)
BRATS Ensemble of a 3D-CNN
- DSC 90.6%
2019 and U-Net
BRATS Bias correction, Two-PathGroup-CNN
DSC 89.2%, PR 88.22%, SEN 88.32%
2015 intensity normalisation (2PG-CNN)
BRATS Hybrid two-track U-Net
- DSC 86.5%, SEN 88.3%, SPE 99.9%
2018 (HTTU-Net)
P-Net with bounding
BRATS
- box and image-specific DSC 86.29%
2015
fine tuning (BIFSeg)
Denoising,
Multi-scale CNN
ADNI skull stripping, ACC 90.1%
(MSCNN)
sub-sampling
Intensity normalisation,
BRATS
resizing, bias field Cascaded 3D U-Nets DSC 89.4%
2017
correction
BRATS
3D centre-crop BRATS 2015: DSC 88.4%, SEN 83.8%
2015 and Downsampling
dense block BRATS 2017: DSC 88.7%, SEN 84.3%
2017
BRATS BRATS 2018: DSC 90%, SEN 90.3,
Z-score normalisation,
2018 and 3D FCN 3 SPE 99.48%; BRATS 2019: DSC 89%,
cropping
2019 SEN 88.3%, SPE 99.51%
Appl. Sci. 2024, 14, 7281 14 of 20

Table 2. Cont.

Model
Dataset Preprocessing Performance
Architecture
Intensity normalisation,
BRATS removing 1% of DCNN BRATS 2018: DSC 86.2%, SEN 84.8%,
2018 highest and lowest (Dense-MultiOCM 4) SPE 99.5%
intensity
Image cropping, DSC 84%, SEN 92%,
TCIA padding, resizing, U-Net SPE 92%,
intensity normalisation ACC 92%
BRATS
BRATS 2013 DSC 86%,
2013,
- AFPNet 5 + 3D CRF BRATS 2015 DSC 82%,
2015,
BRATS 2018 86.58%
2018
Inception-based U-Net
BRATS
+ up skip connection +
2015, Z-score normalisation DSC 89%, PR 78.5%, SEN 89.5%
cascaded training
2017
strategy
Cropping,
BRATS
z-score normalisation, Triple-intersecting BRATS 2015: DSC 85%,
2015,
min–max normalisation U-Nets (TIU-Net) BrainWeb DSC 99.5%
BrainWeb
(BrainWeb)
BRATS LSTM multi-modal DSC 73.09%, SEN 63.76%,
-
2015 U-Net PR 89.79%
1 Heterogeneous CNN combined with conditional random fields and recurrent regression-based neural algorithm.

Table 3. Evaluation of proposed method.

Evaluation Metrics Performance

Accuracy 97.47%
Sensitivity 97.22%
Specificity 97.94%
Dice coefficient index 96.71%

The values reported indicate the high effectiveness of the method, with an accuracy
of 97.47%, demonstrating the model’s ability to identify both positive and negative cases
correctly. The sensitivity is recorded as 97.22%, indicating the model’s proficiency in
correctly identifying positive cases. The specificity is 97.94%, reflecting the model’s accuracy
in identifying negative cases. Finally, the Dice coefficient index, which measures the
similarity between the predicted and actual labels, is 96.71%. These metrics collectively
suggest that the proposed method exhibits robust performance in detecting brain tumours.
Table 4 provides a comparative analysis of various methodologies for brain tumour
classification using the BRATS 2018 dataset. The table outlines the classification accuracies
of different techniques, facilitating an evaluation of their efficacy in tumour detection.
The methods include combinations of CNN and the Local Binary Patterns (LBP) with
Particle Swarm Optimization (PSO) reported by Irfan et al. [44], achieving a 92.5% accuracy.
The LSTM by Amin et al. [51] shows slightly higher accuracy at 93.8%. The brain-storm
optimisation technique by Narmatha et al. [52] matches the first at 92.5%. At the same time,
a method integrating Discrete Cosine Transform (DCT), CNN, and the Extreme Learning
Machine (ELM) by Khan et al. [53] achieves a value of 93.4%. Notably, the proposed
model, which utilises a CNN, MobileNetV2, and Multi-Support Vector Machine (M-SVM),
significantly surpasses these, recording a classification accuracy of 97.5%. This indicates
Appl. Sci. 2024, 14, 7281 15 of 20

the superior performance of the proposed method in accurately classifying brain tumours,
highlighting its potential utility in clinical application.

Table 4. Comparative performance analysis of existing methods.

Methods Classification Accuracy

CNN, LBP, and PSO [44] 92.5%
LSTM [51] 93.8%
Brain-storm optimisation [51] 92.5%
DCT, CNN, and ELM [52] 93.4%
Proposed model (CNN, MobiNetV2, M-SVM) 97.5%

The researchers used a multifaceted approach to improve the results of their med-
ical image analysis models. First, they fine-tuned the model architectures to match the
characteristics of medical image data, such as by adding or changing particular layers.
This architectural adaptation enhanced the models’ ability to capture medical imagery’s
unique features and patterns. Next, the researchers experimented with different training
parameters, like learning rate and batch size, to find the optimal settings for each model
and dataset. By systematically exploring the parameter space, they sought to unlock the
full potential of their models and ensure optimal performance on specific medical image
analysis tasks. Additionally, the researchers applied regularisation techniques, such as
adding regularisation layers and using early stopping, to prevent the over-training of
the models. These strategies helped to improve the models’ generalisation capabilities,
ensuring that they could reliably perform well on new, unseen medical images. Finally,
the researchers tested various loss functions to identify the one best-suited for the task, to
achieve the best possible results for their models. By carefully selecting the appropriate loss
function, they could further refine the models’ performance and enhance the accuracy and
reliability of the medical image analysis outcomes. Through this multifaceted approach,
encompassing architectural adaptations, training parameter optimisation, regularisation
strategies, and loss function selection, the researchers worked diligently to enhance the
performance of their medical image analysis models and obtain the most accurate and
reliable outcomes.
After preliminary training on a dataset, each model was further trained on a specific
dataset to improve results in a particular area, in this case medical image processing.
Additional training allowed the models to better adapt to the specifics of classifying
and segmenting medical images, including brain scans. Additionally, methods of fine-
tuning hyperparameters were applied, and the optimisation of learning algorithms was
undertaken to achieve the best results in a specific task. This approach has significantly
improved the accuracy of image classification and segmentation, which is essential for
diagnosing and treating patients with brain tumours.
As a result of these efforts, we have obtained significant improvements in the accuracy
of segmentation and the classification of brain tumours, which makes our model more
accurate and useful for medical applications.

5. Discussion
This research incorporates sophisticated deep-learning frameworks to enhance the
precision of brain tumour identification and segmentation. Our study focuses on menin-
giomas, utilising a series of MRI images to showcase our model’s capability in detecting
and outlining tumour boundaries accurately.
The visual progression presented in Figure 2, from detecting a hyperintense menin-
gioma mass to its precise segmentation and contextual visualisation, highlights the effec-
tiveness of combining CNNs with advanced classification algorithms. This methodological
fusion is intended to refine diagnostic accuracy and facilitate clinical interventions by
Appl. Sci. 2024, 14, 7281 16 of 20

providing precise, actionable imaging data. The binary mask and the subsequent overlay
on the MRI scan are pivotal in clarifying the tumour’s spatial relationship with adjacent
cerebral structures, a critical factor in planning surgical or therapeutic procedures.
The selection of CNN architectures, as detailed in Table 1, underpins our strategy to
optimise image analysis. Each architecture, from AlexNet to ResNet50, is selected based
on inherent features that support deep learning tasks specific to medical imaging, such as
object localisation and reducing vanishing gradients via skip connections. The diversity
in layer depth, error rates, and computational efficiency across these models facilitates a
tailored approach to feature extraction and image classification, ensuring robustness in
tumour segmentation outcomes.
As summarised in Table 3, our empirical evaluations demonstrate the proposed
model’s superior performance in tumour detection metrics like accuracy, sensitivity, speci-
ficity, and the Dice coefficient index. These metrics not only corroborate the high efficacy of
the Multi-Support Vector Machine classifier when applied to the BRATS 2018 dataset, but
also reflect the model’s generalizability across diverse neuroimaging datasets, including
those from ADNI, TCIA, and BrainWeb.
Despite the promising outcomes, our research recognises inherent challenges
such as dataset bias and class imbalance, which could skew model training and affect
generalizability. Addressing these challenges involves refining data preprocessing and
augmentation techniques to ensure a balanced representation of tumour types and
stages across training sets.
Moreover, the comparative performance metrics in Table 4 emphasise the advancement
of our proposed method over existing techniques. By achieving a classification accuracy of
97.5%, our model sets a new benchmark in the field, underscoring its potential for clinical
application. It surpasses traditional methods and recent innovations, which have shown
lower accuracies in similar settings.
The integration of advanced CNN architectures and ensemble learning strategies has
proven highly effective in enhancing the accuracy and reliability of brain tumour clas-
sification systems. The continuous improvement of these systems through architecture
adaptation, the optimisation of training parameters, and advanced regularisation tech-
niques points towards an exciting future where deep learning can significantly contribute
to personalised medicine and improved patient outcomes in neuro-oncology. Our future
work will further enhance these models, address existing challenges, and expand their
applicability to other medical imaging tasks.

Limitations
While the proposed method for brain tumour classification using deep learning and
ensemble algorithms shows promising results, several limitations must be acknowledged.
The model’s performance relies heavily on the training datasets used, which may not fully
represent the diversity of real-world clinical data, affecting the generalisation to unseen data.
Brain tumour datasets often exhibit class imbalances, leading to biassed model training
and suboptimal performance for less common tumour types. Additionally, the training and
deployment of deep learning models require significant computational resources, which
may not be accessible in all clinical settings. Furthermore, the interpretability of these
models remains a challenge, as clinicians require clear and understandable explanations
for their predictions to aid in decision-making, which the current models may lack. These
limitations highlight the need for continued research and development to address dataset
bias, class imbalance, computational constraints, and model interpretability, ensuring the
deployment of robust and clinically relevant brain tumour classification models.
Robustness and generalisation are critical factors for successfully translating advanced
deep-learning models for brain tumour classification from research to clinical practise.
Ensuring robustness across different imaging conditions and scanners is crucial, as vari-
ations in MRI protocols, scanner types, and patient populations can impact the model’s
performance. To ensure generalisation, the models must be validated extensively across
Appl. Sci. 2024, 14, 7281 17 of 20

diverse clinical settings. Additionally, despite applying regularisation techniques and early
stopping, the risk of overfitting persists, especially when models are trained on limited
datasets, which can lead to poor performance on new, unseen data. Furthermore, integrat-
ing these complex models into clinical workflows poses practical challenges, requiring the
development of user-friendly interfaces and seamless integration with hospital information
systems to facilitate widespread adoption. Addressing these limitations, including those
relating to robustness, generalisation, overfitting, and clinical integration, is crucial for
the reliable and effective deployment of advanced deep learning models in real-world
medical applications.

6. Conclusions
This study successfully demonstrates the capability of an integrated deep-learning
approach to enhance the accuracy and efficiency of meningioma detection and segmen-
tation using MRI scans. The research findings underscore the effectiveness of employing
convolutional neural networks combined with advanced classification algorithms, collec-
tively contributing to more precise medical diagnoses and potentially better healthcare
outcomes. The utilisation of diverse CNN architectures, as elaborated in Table 1, such as
AlexNet, VGG16, GoogleNet, ResNet18, and ResNet50, highlights their distinct advantages
in handling complex image processing tasks. Each architecture’s specific features—from
local response normalization in AlexNet to the innovative skip connections in the ResNet
series—play a crucial role in improving the model’s ability to classify accurately and seg-
ment brain tumours. These architectures are carefully chosen based on their performance
metrics and suitability for the intricate requirements of medical image analysis, as reflected
in their error rates and computational efficiency.
The outcomes of this study, detailed in Table 3, reveal high accuracy, sensitivity,
specificity, and Dice coefficient indices for the proposed method, particularly notable
when applied to the BRATS 2018 dataset. These results validate the model’s robustness
and potential for clinical applications, providing a solid foundation for surgical planning
and targeted therapies. The comparative performance analysis in Table 4 emphasises the
superiority of our proposed model over existing methods. This superiority is a testament
to the careful integration of CNNs with MobileNetV2 and Multi-Support Vector Machine
classifiers, significantly enhancing tumour classification accuracy.
However, the study also acknowledges challenges such as dataset bias, class imbal-
ance, and the required high computational resources. These challenges necessitate refining
segmentation and training techniques to ensure the models remain effective across diverse
and realistic clinical scenarios. The research confirms that the strategic application of ma-
chine learning techniques, specifically tailored CNN architectures and ensemble learning
strategies, can significantly improve the performance of MRI-based brain tumour classi-
fication systems. This advancement paves the way for more accurate and personalised
diagnostic tools and promises to enhance therapeutic strategies for patients with brain
tumours. Future work will focus on expanding these techniques to other forms of medical
imaging and continuing to refine the models to handle the variability and complexity of
real-world clinical data.

Author Contributions: Conceptualization, M.A., A.Z. and A.B.; methodology, A.B., A.Z. and B.M.;
software, M.A.; validation, M.A., A.B. and A.Z.; formal analysis, R.Y.M.L.; investigation, M.A., A.B.,
B.M., R.Y.M.L. and A.Z.; resources, M.A. and A.B.; data curation, M.A. and R.Y.M.L.; writing—
original draft preparation, M.A., A.B. and A.Z.; writing—review and editing, M.A., A.Z. and R.Y.M.L.;
visualisation, M.A., A.B. and R.Y.M.L.; supervision, B.M. and A.Z.; project administration, B.M.;
funding acquisition, B.M. and A.Z. All authors have read and agreed to the published version of the
manuscript.
Funding: The research was supported by the Science Committee of the Ministry of Science and
Higher Education of the Republic of Kazakhstan (Grant no. of the research fund: AP14869848).
Institutional Review Board Statement: Not applicable.
Appl. Sci. 2024, 14, 7281 18 of 20

Informed Consent Statement: Not applicable.

Data Availability Statement: The original contributions presented in the study are included in the
article, further inquiries can be directed to the corresponding authors.
Conflicts of Interest: The authors declare no conflicts of interest. The funders had no role in the
study’s design; in the collection, analyses, or interpretation of data; in the writing of the manuscript;
or in the decision to publish the results.

References
1. Vidyarthi, A.; Agarwal, R.; Gupta, D.; Sharma, R.; Draheim, D.; Tiwari, P. Machine learning assisted methodology for multiclass
classification of malignant brain tumors. IEEE Access 2022, 10, 50624–50640. [CrossRef]
2. Steinmetz, J.D.; Seeher, K.M.; Schiess, N.; Nichols, E.; Cao, B.; Servili, C.; Cavallera, V.; Cousin, E.; Hagins, H.; Moberg, M.E.; et al.
Global, regional, and national burden of disorders affecting the nervous system, 1990–2021: A systematic analysis for the Global
Burden of Disease Study 2021. Lancet Neurol. 2024, 23, 344–381. [CrossRef]
3. Shah, H.A.; Saeed, F.; Yun, S.; Park, J.H.; Paul, A.; Kang, J.M. A robust approach for brain tumor detection in magnetic resonance
images using finetuned efficientnet. IEEE Access 2022, 10, 65426–65438. [CrossRef]
4. Musa, U.I.; Mensah, Q.E.; Falowo, R. Intracranial-tumor detection and classification system using convnet and transfer learning.
Int. Res. J. Eng. Technol. 2022, 10, 120–127.
5. Ahmad, S.; Choudhury, P.K. On the performance of deep transfer learning networks for brain tumor detection using MR images.
IEEE Access 2022, 10, 59099–59114. [CrossRef]
6. Islami, F.; Ward, E.M.; Sung, H.; Cronin, K.A.; Tangka, F.K.; Sherman, R.L.; Zhao, J.; Anderson, R.N.; Henley, S.J.; Yabroff, K.R.;
et al. Annual report to the nation on the status of cancer, part 1: National cancer statistics. JNCI J. Natl. Cancer Inst. 2021, 113,
1648–1669. [CrossRef]
7. Rizwan, M.; Shabbir, A.; Javed, A.R.; Shabbir, M.; Baker, T.; Obe, D.A.J. Brain tumor and glioma grade classification using
Gaussian convolutional neural network. IEEE Access 2022, 10, 29731–29740. [CrossRef]
8. Johnson, D.R.; O’Neill, B.P. Glioblastoma survival in the United States before and during the temozolomide era. J. Neuro-Oncol.
2012, 107, 359–364. [CrossRef] [PubMed]
9. Rehman, A.; Naz, S.; Razzak, M.I.; Akram, F.; Imran, M. A deep learning-based framework for automatic brain tumors
classification using transfer learning. Circuits Syst. Signal Process. 2020, 39, 757–775. [CrossRef]
10. Fernando, T.; Gammulle, H.; Denman, S.; Sridharan, S.; Fookes, C. Deep learning for medical anomaly detection—A survey. ACM
Comput. Surv. (CSUR) 2021, 54, 1–37. [CrossRef]
11. Somasundaram, K.; Genish, T.; Kalaiselvi, T. An atlas based approach to segment the hippocampus from MRI of human head
scans for the diagnosis of Alzheimers disease. Int. J. Comput. Intell. Inform. 2015, 5.
12. Xie, Y.; Zaccagna, F.; Rundo, L.; Testa, C.; Agati, R.; Lodi, R.; Tonon, C. Convolutional neural network techniques for brain tumor
classification (from 2015 to 2022): Review, challenges, and future perspectives. Diagnostics 2022, 12, 1850. [CrossRef]
13. Louis, D.N.; Perry, A.; Reifenberger, G.; Von Deimling, A.; Figarella-Branger, D.; Cavenee, W.K.; Ohgaki, H.; Wiestler, O.D.;
Kleihues, P.; Ellison, D.W. The 2016 World Health Organization classification of tumors of the central nervous system: A summary.
Acta Neuropathol. 2016, 131, 803–820. [CrossRef]
14. Montaha, S.; Azam, S.; Rafid, A.R.H.; Hasan, M.Z.; Karim, A.; Islam, A. Timedistributed-cnn-lstm: A hybrid approach combining
cnn and lstm to classify brain tumor on 3d mri scans performing ablation study. IEEE Access 2022, 10, 60039–60059. [CrossRef]
15. Lundervold, A.S.; Lundervold, A. An overview of deep learning in medical imaging focusing on MRI. Z. Für Med. Phys. 2019, 29,
102–127. [CrossRef]
16. Rundo, L.; Militello, C.; Vitabile, S.; Russo, G.; Pisciotta, P.; Marletta, F.; Ippolito, M.; D’Arrigo, C.; Midiri, M.; Gilardi, M.C.
Semi-automatic brain lesion segmentation in Gamma Knife treatments using an unsupervised fuzzy c-means clustering technique.
In Proceedings of the Advances in Neural Networks: Computational Intelligence for ICT. International Workshop on Neural
Networks, WIRN 2015, Vietri sul Mare, Italy, 20–22 May 2015; pp. 15–26.
17. Liang, N.Y.; Huang, G.B.; Saratchandran, P.; Sundararajan, N. A fast and accurate online sequential learning algorithm for
feedforward networks. IEEE Trans. Neural Netw. 2006, 17, 1411–1423. [CrossRef]
18. Nie, D.; Wang, L.; Gao, Y.; Lian, J.; Shen, D. STRAINet: Spatially varying sTochastic residual AdversarIal networks for MRI pelvic
organ segmentation. IEEE Trans. Neural Netw. Learn. Syst. 2018, 30, 1552–1564. [CrossRef] [PubMed]
19. Hollon, T.C.; Pandian, B.; Adapa, A.R.; Urias, E.; Save, A.V.; Khalsa, S.S.S.; Eichberg, D.G.; D’Amico, R.S.; Farooq, Z.U.; Lewis, S.
Near real-time intraoperative brain tumor diagnosis using stimulated Raman histology and deep neural networks. Nat. Med.
2020, 26, 52–58. [CrossRef]
20. Natarajan, A.; Kumarasamy, S. Efficient segmentation of brain tumor using FL-SNM with a metaheuristic approach to optimization.
J. Med. Syst. 2019, 43, 25. [CrossRef] [PubMed]
21. Muhammad, K.; Khan, S.; Del Ser, J.; De Albuquerque, V.H.C. Deep learning for multigrade brain tumor classification in smart
healthcare systems: A prospective survey. IEEE Trans. Neural Netw. Learn. Syst. 2020, 32, 507–522. [CrossRef] [PubMed]
22. Mohan, G.; Subashini, M.M. MRI based medical image analysis: Survey on brain tumor grade classification. Biomed. Signal
Process. Control 2018, 39, 139–161. [CrossRef]
Appl. Sci. 2024, 14, 7281 19 of 20

23. Hussain, L.; Saeed, S.; Awan, I.A.; Idris, A.; Nadeem, M.S.A.; Chaudhry, Q.U.A. Detecting brain tumor using machines learning
techniques based on different features extracting strategies. Curr. Med. Imaging 2019, 15, 595–606. [CrossRef] [PubMed]
24. Gurbină, M.; Lascu, M.; Lascu, D. Tumor detection and classification of MRI brain image using different wavelet transforms
and support vector machines. In Proceedings of the 2019 42nd International Conference on Telecommunications and Signal
Processing (TSP), Budapest, Hungary, 1–3 July 2019; pp. 505–508.
25. Sarkar, A.; Maniruzzaman, M.; Ahsan, M.S.; Ahmad, M.; Kadir, M.I.; Islam, S.T. Identification and classification of brain tumor
from MRI with feature extraction by support vector machine. In Proceedings of the 2020 International Conference for Emerging
Technology (INCET), Belgaum, India, 5–7 June 2020; pp. 1–4.
26. Sekhar, A.; Biswas, S.; Hazra, R.; Sunaniya, A.K.; Mukherjee, A.; Yang, L. Brain tumor classification using fine-tuned GoogLeNet
features and machine learning algorithms: IoMT enabled CAD system. IEEE J. Biomed. Health Inform. 2021, 26, 983–991. [CrossRef]
[PubMed]
27. Ramdlon, R.H.; Kusumaningtyas, E.M.; Karlita, T. Brain tumor classification using MRI images with K-nearest neighbor method.
In Proceedings of the 2019 International Electronics Symposium (IES), Surabaya, Indonesia, 27–28 September 2019; pp. 660–667.
28. Sathi, K.A.; Islam, M.S. Hybrid feature extraction based brain tumor classification using an artificial neural network. In
Proceedings of the 2020 IEEE 5th International Conference on Computing Communication and Automation (ICCCA), Greater
Noida, India, 30–31 October 2020; pp. 155–160.
29. Charan, K.S.; Chokkalingam, S.P.; Sundari, K.S. Efficiency of Decision tree algorithm for Brain Tumor MRI Images comparing
with SVM Algorithm. In Proceedings of the 2022 14th International Conference on Mathematics, Actuarial Science, Computer
Science and Statistics (MACS), Karachi, Pakistan, 12–13 November 2022; pp. 1–4.
30. Amran, G.A.; Alsharam, M.S.; Blajam, A.O.A.; Hasan, A.A.; Alfaifi, M.Y.; Amran, M.H.; Gumaei, A.; Eldin, S.M. Brain tumor
classification and detection using hybrid deep tumor network. Electronics 2022, 11, 3457. [CrossRef]
31. Sajjad, M.; Khan, S.; Muhammad, K.; Wu, W.; Ullah, A.; Baik, S.W. Multi-grade brain tumor classification using deep CNN with
extensive data augmentation. J. Comput. Sci. 2019, 30, 174–182. [CrossRef]
32. Ottom, M.A.; Rahman, H.A.; Dinov, I.D. Znet: Deep learning approach for 2D MRI brain tumor segmentation. IEEE J. Transl. Eng.
Health Med. 2022, 10, 1–8. [CrossRef]
33. Asif, S.; Yi, W.; Ain, Q.U.; Hou, J.; Yi, T.; Si, J. Improving effectiveness of different deep transfer learning-based models for
detecting brain tumors from MR images. IEEE Access 2022, 10, 34716–34730. [CrossRef]
34. Park, H.; Sjösund, L.L.; Yoo, Y.; Bang, J.; Kwak, N. Extremec3net: Extreme lightweight portrait segmentation networks using
advanced c3-modules. arXiv 2019, arXiv:1908.03093.
35. Chen, L.C.; Papandreou, G.; Schroff, F.; Adam, H. Rethinking atrous convolution for semantic image segmentation. arXiv 2017,
arXiv:1706.05587.
36. Chen, X.; Qi, D.; Shen, J. Boundary-aware network for fast and high-accuracy portrait segmentation. arXiv 2019, arXiv:1901.03814.
37. Toufiq, D.M.; Sagheer, A.M.; Veisi, H. A review on brain tumor classification in mri images. Turk. J. Comput. Math. Educ.
(TURCOMAT) 2021, 12, 1958–1969.
38. Hossain, M.F.; Islam, M.A.; Hussain, S.N.; Das, D.; Amin, R.; Alam, M.S. Brain Tumor Classification from MRI Images Using
Convolutional Neural Network. In Proceedings of the 2021 IEEE International Conference on Artificial Intelligence in Engineering
and Technology (IICAIET), Kota Kinabalu, Malaysia, 13–15 September 2021; pp. 1–6.
39. Saranya, N.; Renuka, D.K. Brain tumor classification using convolution neural network. In Journal of Physics: Conference Series;
IOP Publishing: Bristol, UK, 2021; Volume 1916, p. 012206.
40. Havaei, M.; Davy, A.; Warde-Farley, D.; Biard, A.; Courville, A.; Bengio, Y.; Pal, C.; Jodoin, P.-M.; Larochelle, H. Brain tumor
segmentation with deep neural networks. Med. Image Anal. 2017, 35, 18–31. [CrossRef]
41. Bayoumi, E.; Khalaf, A.A.; Gharieb, R.R. Brain tumor automatic detection from MRI images using transfer learning model with
deep convolutional neural network. J. Adv. Eng. Trends 2021, 41, 19–30. [CrossRef]
42. Minarno, A.E.; Mandiri, M.H.C.; Munarko, Y.; Hariyady, H. Convolutional neural network with hyperparameter tuning for brain
tumor classification. Kinet. Game Technol. Inf. Syst. Comput. Netw. Comput. Electron. Control 2021, 4, 127–132. [CrossRef]
43. Wolpert, D.H. Stacked generalization. Neural Netw. 1992, 5, 241–259. [CrossRef]
44. Nigam, I.; Huang, C.; Ramanan, D. Ensemble knowledge transfer for semantic segmentation. In Proceedings of the 2018 IEEE
Winter Conference on Applications of Computer Vision (WACV), Lake Tahoe, NV, USA, 12–15 March 2018; pp. 1499–1508.
45. Zhang, S.H.; Dong, X.; Li, H.; Li, R.; Yang, Y.L. PortraitNet: Real-time portrait segmentation network for mobile device. Comput.
Graph. 2019, 80, 104–113. [CrossRef]
46. Kim, Y.W.; Rose, J.I.; Krishna, A.V. Accuracy enhancement of portrait segmentation by ensembling deep learning models. In
Proceedings of the 2020 Fifth International Conference on Research in Computational Intelligence and Communication Networks
(ICRCICN), Bangalore, India, 26–27 November 2020; pp. 59–64.
47. Rohlfing, T.; Maurer, C.R., Jr. Shape-based averaging for combination of multiple segmentations. In Proceedings of the
International Conference on Medical Image Computing and Computer-Assisted Intervention, Palm Springs, CA, USA, 26–29
October 2005; Springer: Berlin/Heidelberg, Germany, 2005; pp. 838–845.
48. Hansen, L.K.; Salamon, P. Neural network ensembles. IEEE Trans. Pattern Anal. Mach. Intell. 1990, 12, 993–1001. [CrossRef]
49. Singh, V.; Mukherjee, L.; Peng, J.; Xu, J. Ensemble clustering using semidefinite programming with applications. Mach. Learn.
2010, 79, 177–200. [CrossRef]
Appl. Sci. 2024, 14, 7281 20 of 20

50. Zhao, B.; Feng, J.; Wu, X.; Yan, S. A survey on deep learning-based fine-grained object classification and semantic segmentation.
Int. J. Autom. Comput. 2017, 14, 119–135. [CrossRef]
51. Narmatha, C.; Eljack, S.M.; Tuka, A.A.R.M.; Manimurugan, S.; Mustafa, M. A hybrid fuzzy brain-storm optimization algorithm
for the classification of brain tumor MRI images. J. Ambient. Intell. Humaniz. Comput. 2020, 1–9. [CrossRef]
52. Khan, M.A.; Ashraf, I.; Alhaisoni, M.; Damaševičius, R.; Scherer, R.; Rehman, A.; Bukhari, S.A.C. Multimodal brain tumor
classification using deep learning and robust feature selection: A machine learning application for radiologists. Diagnostics 2020,
10, 565. [CrossRef] [PubMed]
53. 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] [PubMed]
54. Rokach, L. Ensemble-based classifiers. Artif. Intell. Rev. 2010, 33, 1–39. [CrossRef]
55. Goodfellow, I.; Bengio, Y.; Courville, A. Deep Learning; MIT Press: Cambridge, MA, USA, 2016.
56. Amin, J.; Sharif, M.; Raza, M.; Saba, T.; Sial, R.; Shad, S.A. Brain tumor detection: A long short-term memory (LSTM)-based
learning model. Neural Comput. Appl. 2020, 32, 15965–15973. [CrossRef]
57. Muthukrishnan, R.; Radha, M. Edge detection techniques for image segmentation. Int. J. Comput. Sci. Inf. Technol. 2011, 3, 259.
[CrossRef]

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.