Sulekha Das / Afr.J.Bio.Sc. 6(5) (2024).

7113-7126 ISSN: 2663-2187

https://doi.org/10.33472/AFJBS.6.5.2024. 7113-7126

AfricanJournalofBiological
Sciences
A Novel Robust CNN Model for MRI Brain Tumor
Classification
Sulekha Das1, *Dr.Avijit Kumar Chaudhuri2, Dr.Partha Ghosh3, Swagato Sikdar4
1
Research Scholar, Information Technology, GCECT, Kolkata, West Bengal, India
2
Associate Professor, Computer Science and Engineering, Brainware University, Barasat, West Bengal, India
3
Assistant Professor, Computer Science and Engineering, GCECT, Kolkata, West Bengal, India
4
Student –CSE-AIML(2nd year), Techno Engineering College Banipur

shu7773sea@gmail.com,c.avijit@gmail.com,parth_ghos@rediffmail.com,sikdarswagato0@gmail.co
m

Abstract

Brain tumors rank as the 10th leading cause of death for men and
women, affecting both adults and children1. This underscores the urgent
need for effective strategies in prevention, diagnosis, and treatment to
Article History
address this significant health challenge. Magnetic resonance imaging
Volume 6, Issue 5, 2024 (MRI) is the preferred method for identifying brain tumors. Recent
Received: 15 May 2024
Accepted: 22 May 2024 advancements in image classification technology, particularly
doi:10.33472/AFJBS.6.5.2024. 7113-7126
Convolutional Neural Networks (CNNs), have greatly improved tumor
classification accuracy. The findings of this study have significant
implications for clinicians specializing in the early detection of brain
tumors. By leveraging advanced neural network models, healthcare
professionals can potentially improve the accuracy and efficiency of
tumor diagnosis, leading to better patient outcomes and possibly earlier
interventions. This underscores the importance of leveraging cutting-edge
technology, such as deep learning and neuroimaging, in medical
diagnostics. In this study, CNNs are used for brain tumor classification,
successfully categorizing brain images into two classes: benign and
malignant with an impressive accuracy of 97%.
1
www.cancer.net

Keywords: Convolutional Neural Network (CNN), Brain Tumor, Kappa score,

Area Under the Curve(AUC)
 Page 7114 of 7136
Sulekha Das / Afr.J.Bio.Sc. 6(5) (2024). 7113-7126

Introduction

Brain tumors rank as the 10th leading cause of death for men and women, affecting both
adults and children. This underscores the urgent need for effective strategies in prevention,
diagnosis, and treatment to address this significant health challenge. Magnetic resonance
imaging (MRI) is the preferred method for identifying brain tumors. Recent advancements in
image classification technology, particularly the brain is, by far, the master organ of the
human body, controlling the intricately intertwined processes of all other systems and making
critical choices in our lives (Saba, 2020). Being the coordinator of the central nervous system,
it controls both the voluntary activities and the involuntary processes that are a vital part of
our body's functioning (Özyurt et al., 2020). While the heart serves as a critical organ that
keeps us alive, the development of tumors within this vital structure poses considerable
threats to our health and wellness.

Brain and central nervous system cancers represent a growing, frightful global public health
challenge because of their high death rate, inflating costs to individuals and society, poor
survival rate, and extremely low quality of life. As specified in the GLOBOCAN 2020 data,
brain and central nervous system cancer is a significant share of the global causes of sickness
(Ilic&Ilic, 2023). They are the 19th most frequent cancer, accounting for 1.9% of all cancers,
and the 12th leading cause of cancer-related deaths, amounting to 2.5% of overall cancer
deaths. Comprehensive, in-depth knowledge of the many aspects of brain tumors, like their
symptoms, stage, and criticality, is very important for early diagnosis and solution. Magnetic
Resonance Imaging (MRI) has been a base for clinicians to uncover internal subtle
information about the brain structure and composition, including tumors (Sharif et al.,
2021). Advanced computational methods, including deep learning, have demonstrated their
capability to improve the accuracy and efficiency of tumor detection and classification.

On the contrary, it seems that a lot of articles have been written on this subject to summarize
the current trends in DL, but they usually center around a very narrow part of the issue, which
in the end leads to a situation in which we know too little about the field in its
totality. Therefore, people do not have much idea about the scope of DL developments and
how much they have grown up in this time.
 Page 7115 of 7136
Sulekha Das / Afr.J.Bio.Sc. 6(5) (2024). 7113-7126

In this regard, Convolutional Neural Networks (CNNs) have become an essential instrument
for discerning and classifying different types of brain tumors, which include gliomas,
meningiomas, pituitary tumors, and normal brain tissue. CNNs include multiple layers of
interconnected neurons. When all of them are connected, they form a network that can learn
complex patterns and features from data (Hu &Razmjooy, 2021). Datasets with labeled brain
images are the means of adjusting neurons' weights and biases across various layers for
CNNs during the training process. Through repeated comparisons of predicted outputs and
ground truth labels, the network refines its parameters to reduce errors and achieve the
highest accuracy. This pattern of iterative learning occurs through the activation of functions,
which introduces non-linearity and thus allows the network to capture internal relationships
within the data.

The structure of CNNs is mainly the input layer, output layer, and at least one hidden layer is
present. These layers act as key components in the feature extraction and abstraction
processes, which enables the network to detect these sensitive distinctions. Moreover, trial
and error with various parameters of the machine learning model like the number of layers,
input filters, and learning rates positively impacts the diagnosis of different tissues
(Mehnatkesh et al., 2023, Kumar &Mankame, 2020). Despite their computational
complexity, CNNs deliver a unique function in dealing with and processing medical images
and their capabilities to excel in the field of neuroimaging research cannot be devalued.
Researchers constantly attempt to increase the efficacy of these deep learning models to
enhance diagnostic precision, simplify clinical procedures, and finally, ensure the
improvement of patient outcomes in the diagnosis and treatment of brain cancer.

Working of the proposed CNN model

The architecture of the experimental layout of this research and model construction is
presented in Fig.1. A CNN in PyTorch is formulated from constituent
layers which include convolutional, pooling, and fully connected layers, to build the model
structure. These levels are necessary to build up hierarchical representations of features from
input data consideration, especially for tasks like image classification. The process consists
of creating an initial Python class consisting of the CNN model. All these layers are set as an
initial layer where parameters are defined, such as the kernel size, the number of filters, and
activation functions. For instance, the authors start with four convolution layers followed by
pooling layers to down sample the feature map and simply connected layers for classification.
After the creation of these layers, they are passed into the forward function of the class one
 Page 7116 of 7136
Sulekha Das / Afr.J.Bio.Sc. 6(5) (2024). 7113-7126

by one, which produces the outputs of the specific network. It is a key part of the network
where data undergoes convolutional operations, followed by pooling, and then steps through
fully connected layers ultimately. With this connected architecture, information flows in a
predefined order during both prediction and training (backward pass) to facilitate the efficient
propagation of information through the network.

The model is trained with image datasets after constructing architecture. This is done by
feeding the input images through the network one after another, computing loss between
predicted and actual labels, and then using techniques like backpropagation and optimization
algorithms (like Stochastic Gradient Descent, Adam) to update network parameters
iteratively. There is an automatic learning functionality enabled that allows the convolutional
and pooling layers to learn effective feature representations from the input images. This is
done through the process of optimization of parameters. One of the biggest strengths of
CNNs is their ability to learn detailed features directly from data making them suitable for
achieving outstanding performance across several computer vision tasks.

The authors employed several strategies to enhance the performance of their Convolutional
Neural Network (CNN) model for vision tasks:

Adding More Layers: Through an increase in the depth of the CNN architecture, the
model is essentially capable of learning more complex elements and features from the input
data. This leads to the acquisition of better precision and increases the overall quality of the
model.

Training for More Epochs: A rise of epochs means that the model will see the training data
more times, which helps higher convergence to fine-tuned features. Thus, the model can
create more complex embeddings that manage to encapsulate a greater diversity of
features that can be found in the input data.

Fine-tuning Hyperparameters: Hyperparameters like learning rate and regularization

strength are very important in the training process. Through the selection of grid search or
random search methods, the authors aim to find the exact setting of all hyperparameters that
will guarantee the perfect model to run. The adjustment of the learning rate determines the
step size during parameter update and thus affects the speed and stability of training.
Regularization techniques act as a buffer against overfitting in complex models by imposing
penalties.
 Page 7117 of 7136
Sulekha Das / Afr.J.Bio.Sc. 6(5) (2024). 7113-7126

Experimentation and Iterative Improvement: The authors use the cycle of trial and error
to improve the model by iteratively adjusting the model architecture, training procedures, and
parameters depending on the feedback. Therefore, this process involves systematically
exploring different configuration options and strategies to find the best-performing model to
match the specific vision tasks.

The authors will implement these strategies and will never cease refining their model through
experimentation to achieve the accuracy and robustness of the vision tasks given. Therefore,
this iterative process shows that a model should meet high initial standards and also be
improved and optimized during the whole learning process. The architecture for the proposed
classifier for brain cancer prediction has been depicted in Figure 1 below.

Figure 1: Brain Cancer Prediction Framework

Relevant Literature

According to McCaley(2015), knowledge classification (normal vs. mutated brains) should

prevail, and brain tumors have to be predicted more carefully. Within the scope of research,
the researchers employ a wide range of machine learning techniques, which consist of a
Support Vector Machine (SVM), K-nearest neighbor (KNN), and Hybrid Classifier (SVM-
KNN). Through the use of such tools, they classified the 50-image dataset. Data show that the
SVM-KNN classifier method has a top accuracy level of 98% when compared to the rest of
the trained methods. This paper’s main aim is to show the improvement in MRI brain cancer
classification accuracy, especially with the SVM-KNN hybrid classifier application.
 Page 7118 of 7136
Sulekha Das / Afr.J.Bio.Sc. 6(5) (2024). 7113-7126

Afshar et al., 2018 indicate, that the early and precise diagnosis of brain tumors is
complicated based on brain tumors' different severity and incidence across different age
groups. Convolution neural networks (CNNs) are suitable for this type of task, but they
require trained data and also cannot handle input transformations. Capsule Network
(CapsNets), the breakthrough in machine learning architectures, is a natural choice here since
it is more robust to transformations and requires a smaller training set, which makes it perfect
for such a small dataset and brain MRI images.

Experimental data show that CapsNet has a CNN (convolutional neural network) structure,
performing better than the others in the deep learning area, including imaging tasks,
especially in the field of medicine.

Zacharaki et al., 2019, focused on pattern classification methods capable of recognizing

cluster differences, for example, between primary gliomas and brain metastases, and glioma
grading. They invented unique methods of classification that are capable of integrating
conventional MRI and perfusion MRI for better differentiation ability. The methodology
consists of a series of steps such as region of interest (ROI) definition, feature extraction,
feature selection, and classification. The extended characteristics in principle include the
shape and intensity data and the rotation-invariant texture features. A feature subset is a result
of using a support vector machine algorithm with recursive feature elimination.

Sajjad et al., 2019 proposed a novel multi-grade classification system for brain tumor
diagnosis based on the CNN approach, which aims to alleviate the work burden of
radiologists and better utilize MRI data. The suggested system is verified through data
augmentation and the original data. It shows a better performance than existing ones. The
results have demonstrated the efficiency of this method in accurate and reliable brain tumor
classification. Thus, radiologists can work properly based on medical imaging analysis.

Deepak et al., 2019 present a new approach to brain tumor classification in which three
different types of tumors are identified: Glioma, meningioma, and pituitary are the major
three diseases to be detected by the deep convolutional transfer learning approach. The
methodology includes the use of the GoogLeNetmodel that has been pre-trained to produce
the image attributes from the MRI images of the brain. Subsequently, these features shall be
given to trained classifier models so that classification can be performed. This study
randomly applies a patient-level, five-fold cross-validation process to MRI data collected
through Figshare. The obtained results reveal a highly accurate mean classification accuracy
of 98%. Performance metrics, e.g., AUC, Precision, Recall, F-score, and Specificity are also
 Page 7119 of 7136
Sulekha Das / Afr.J.Bio.Sc. 6(5) (2024). 7113-7126

computed. On the other hand, this paper reveals the practicality of small training samples,
based on transfer learning, which is shown in this case. The outcomes prove that deep
transfer learning is an effective method in medical image analysis, especially when the data is
less available.

Deb et al.(2024) in their proposed work, make use of deep neural network form, which is a type
of convolutional neural network (CNN) to classify images into TUMOUR DETECTED and
TUMOUR NOT DETECTED. The model has achieved outstanding results in terms of
accuracy, with a mean accuracy score of 96.08% and a fscore of 97.3. This suggests that such
models can detect tumors precisely in MRIs and hence demonstrate the path with deep
learning methods used in the field of medical image analysis and diagnosis.

Ayadi et al., 2021 noted that the main aim of brain tumors classification results from their
grave consequences as well as different classifications. Moreover, the new CNN-based
model, which can differentiate between brain tumors in MRI images, is illustrated in the first
instance. This model has been made up of numerous layers of the network to achieve decent
classification accuracy for brain tumors.

Table 1. Summary of recent work that has been performed in brain cancer detection and in
this research using machine learning and deep learning algorithms

Author Feature/methods Performance(%)

Sasikala et al., DWT, GA, ANN Accuracy = 98

2008
Sensitivity = NA

Specificity = NA

AUC = NA

Verma et al., 2008 Bayesian, and SVM Accuracy = NA

Sensitivity = 91.84

Specificity = 99.57

AUC = NA

Zacharaki et al., SVM, RFE Feature Ranking, LDA, Accuracy = 97.8

 Page 7120 of 7136
Sulekha Das / Afr.J.Bio.Sc. 6(5) (2024). 7113-7126

2009 KNN, NL-SVM Sensitivity = 100

Specificity = 95

AUC = 98.6

Ryu et al., 2014 GLCM, Entropy, Histogram Accuracy = 84.4

Sensitivity = 81.8

Specificity = 90

AUC = 94.1

Machhale et al., SVM-KNN Sensitivity = 100

2015
Specificity = 93.75

Accuracy = 98

Cheng et al., 2015 Feature extraction: Intensity, Accuracy = 91.28

histogram, GLCM, BOW,
classification Methods: SVM,

SRC, KNN

Skogenet al., 2016 Statistical Analysis and Standard Accuracy = 84.4

Deviation
Sensitivity = 93

SPC = 81

AUC = 91

Paul et al., 2017 CNN Accuracy = 84.19

Afshar et al., 2018 Capsule network method Accuracy = 86.56

Zia et al., 2018 Window based image cropping Accuracy = 85.69

Sensitivity = 86.26

Specificity = 90.90
 Page 7121 of 7136
Sulekha Das / Afr.J.Bio.Sc. 6(5) (2024). 7113-7126

Sajjad et al., 2019 CNN with data augmentation Accuracy = 94.58

Sensitivity = 88.41

Specificity = 96.12

Badža&Barjaktarovi CNN Accuracy = 95.40

c, 2020

Huang et al., 2020 Convolutional Neural Network Based Accuracy = 95.49

on Complex Networks (CNNBCN)

This Study CNN Accuracy = 97

Sensitivity = 96

Specificity = 98

Kappa Score = 94

AUC/ ROC = 97

Methodology

Dataset Description

The dataset used in this study for the brain tumor classification was acquired from Kaggle,
containing around 3762 MRI images. The dataset contains 2079 images without any tumors
and 1683 images with tumors. These images, which are segmented into training and testing,
are used for the creation of models. The training data corresponds to the part of input
information used for the learning process, whereas the testing data is used to test
performance. The dataset has been divided into training and testing groups in various ratios,
such as 60/40, 70/30 and 80/20 train/test partitions, and 10-fold cross-validation. In cross-
validation, the dataset is split into mutually exclusive subsets, and the training and testing are
done through the different subset combinations. To illustrate this, Figure # demonstrates
some data that will be used for experimental purposes. The visualization part of the process
can, therefore, be of help in understanding the general qualities and features of the
dataset that is being examined.
 Page 7122 of 7136
Sulekha Das / Afr.J.Bio.Sc. 6(5) (2024). 7113-7126

Convolutional Neural network:

A Convolutional Neural Network (CNN) is an exemplary deep learning architecture that was
developed for dealing with structured grid data frameworks such as images. The structure
consists of several layers, wherein each layer performs specific functions and is stacked in
hierarchical order to extract the salient features from the input data.

Input Layer: The input layer receives the raw image data, typically represented as a matrix
of pixel values.

Convolutional Layer: A Convolutional Neural Network (CNN) is a typical deep learning

architecture that was created specifically to solve the problems of data with grid structures,
e.g. images. The structure is made up of a sequence of layers. Each layer has its own function
to accomplish, and each layer has a hierarchical order trying to extract the salient features of
the input data.

Pooling Layer:The pooling layer manages the spatial dimensions of the feature maps
obtained from the convolutional layers by downsizing. Max pooling, a widely used method,
picks the maximum from each of the local areas (e.g., 2x2) of the feature map. One of the
roles of Max pooling is to reduce computational complexity, make the network more tolerant
of minor input differences, and minimize over-fitting. In addition, pooling layers capture the
most important features and dispose of unrelated information while minimizing sensitivity to
the input translations.

Fully Connected (FC) Layer: The fully connected layers integrate the detected features by
the convolutional and pooling layers for the purpose of classification or regression. Every
single neuron in a fully connected layer is connected to every neuron in the previous layer, so
it forms a tightly packed structure. Unlike CNN, FC layers focus on learning the
interrelationships between features and making the final predictions. We may use activation
functions such as Softmax to assign class probabilities from the fully connected layer output.

Output Layer: The output layer generates the final predicted or classified results. For the
classification tasks, it mainly uses the softmax activation function to convert raw scores into
the degree of probability of belonging to classes. The network's output shows the confidence
interval for the output classes.
 Page 7123 of 7136
Sulekha Das / Afr.J.Bio.Sc. 6(5) (2024). 7113-7126

Feature Map: Feature maps, as the output of convolutional layers, encode learned features.
They referred to as convolutional neural network (CNN) is structured grids of activation
values, where each value represents either the existence or the lack of a specific feature at a
given position in the input. Feature maps are partitioned into the overlapped sets to perform
pooling operations, which are considered as the most important information either with the
max pooling or the average pooling. In summary, CNNs make use of convolutional and
pooling operations to extract hierarchical features from input images and finally fully
connected layers to make decisions based on these features. This architecture obtained
remarkable result in numerous computer vision tasks, such as image classification, object
detection, and segmentation. (Ting et al., 2019; Dang et al., 2019). Figure 2 demonstrate the
CNN architecture.

Figure 2. CNN architecture for brain tumor diagnosis

CNN in Cancer Research

CNNs have been of interest to many researchers in their ability to process and analyze images
as well as recognize patterns. Here are some key reasons why CNNs are extensively used in
cancer research:
 Page 7124 of 7136
Sulekha Das / Afr.J.Bio.Sc. 6(5) (2024). 7113-7126

Feature Extraction from Medical Images:

CNNs can learn hierarchical image representations automatically and consequently can
extract intricate features from MRI and CT images as a result. These features are particularly
important in detecting even small abnormalities suggestive of cancer(Litjens et al., 2017).

Classification and Diagnosis:

CNNs are capable of classifying medical images as belonging to various categories that in
turn may be utilized for different cancer diagnoses. They are trained to learn discriminative
features and then are able to give a precise classification of the available data to the
healthcare professionals for informative decisions(Esteva et al., 2017).

Segmentation of Tumor Regions:

The CNNs can segment tumour regions in medical images highlighting the edges of
cancerous tissues. Precise segmentation is vital for treatment planning, monitoring disease
progression and checking the effectiveness of treatment(Litjens et al., 2016)..

Prediction of Treatment Response:

CNNs can forecast the response of tumors to specific treatment with help of pre-treatment
imaging data. Through the analysis of tumor characteristics change, CNNs play role in
tailoring treatments to specific patients to achieve better outcomes(Parmar et al., 2017).

Prognostication and Survival Prediction:

CNNs can analyze medical images to extract prognostic biomarkers indicative of patient
outcomes. By identifying imaging features associated with disease aggressiveness, CNNs can
aid in predicting patient survival and stratifying risk(Wulczyn et al., 2020).

The extensive integration of CNNs in cancer research spanning diverse applications,

highlighting their capacity to transform cancer diagnosis, treatment, and prognosis
forecasting.

Results and Discussion

A comprehensive breakdown of various performance metrics commonly used in binary

classification problems is described below.
 Page 7125 of 7136
Sulekha Das / Afr.J.Bio.Sc. 6(5) (2024). 7113-7126

Accuracy: Accuracy can measure the overall correctness of the classifier. It records the ratio
of correctly classified instances (both positive and negative) that are classified correctly to the
total number of instances, as depicted in the equation (1)

(1)

Sensitivity(Recall): Sensitivity, which for most cases is known as predictive value or true
positive rate, is defined as the probability of getting the correct disease class and it is depicted
in the formula (2).

(2)

Specificity: Specificity is defined as the proportion of actual true negative cases that are
classified as negative by the method. It is depicted as shown in the formula (3).

. (3)

Precision: The measures of precision are the number of instances of positive test results
correctly diagnosed over all the positive tests. It is shown in the formula (4).

(4)

False Positive Rate (FPR): The FPR is a ratio of the proportion of true negative to that of the
corresponding false positive and represented by the mathematical function (5).

(5)

F1 Score: F1 score is the harmonic mean (HM) of precision and recall rates. It gives a
balance between precision and recall, and used when classes are imbalanced. It is represented
by the mathematical function (6).

(6)
 Page 7126 of 7136
Sulekha Das / Afr.J.Bio.Sc. 6(5) (2024). 7113-7126

(7)

(8)

The accuracy, recall, specificity, Kappa statistics and AUCs are the basis of comparison and
are depicted in Tables 2-5, respectively.

Table 2. Comparison of accuracies for train/ test partitions(Epoch 30)

Train/ Test Accuracy Tumours

Partitions Metrics Prediction
Accuracy()

60-40 Accuracy 92.35

Sensitivity 92

Specificity 93

AUC 92

Kappa Score 84

70-30 Accuracy 92

Sensitivity 91

Specificity 92

AUC 91

Kappa Score 83

80-20 Accuracy 91

Sensitivity 86

Specificity 95

AUC 90
 Page 7127 of 7136
Sulekha Das / Afr.J.Bio.Sc. 6(5) (2024). 7113-7126

Kappa Score 80

Table 3. Comparison of accuracies for 10 fold cross validations(Epoch 30)

Train/ Test Accuracy Tumours

Partitions Metrics Prediction
Accuracy()

1st fold Accuracy 91

Sensitivity 82

Specificity 97

AUC 97

Kappa Score 80

2nd fold Accuracy 92

Sensitivity 86

Specificity 97

AUC 90

Kappa Score 83

3rd fold Accuracy 94

Sensitivity 92

Specificity 95

AUC 95

Kappa Score 87

4th Accuracy 94.5

Sensitivity 92.34

Specificity 96.22
 Page 7128 of 7136
Sulekha Das / Afr.J.Bio.Sc. 6(5) (2024). 7113-7126

AUC 94

Kappa Score 88.3

5th Accuracy 93.45

Sensitivity 93

Specificity 95

AUC 92

Kappa Score 87

6th Accuracy 93

Sensitivity 87

Specificity 96

AUC 92

Kappa Score 85

7th Accuracy 92.38

Sensitivity 89

Specificity 95.4

AUC 94

Kappa Score 85

8th Accuracy 94

Sensitivity 91

Specificity 96

AUC 93

Kappa Score 87

9th Accuracy 91
 Page 7129 of 7136
Sulekha Das / Afr.J.Bio.Sc. 6(5) (2024). 7113-7126

Sensitivity 87

Specificity 94

AUC 94

Kappa Score 81

10th Accuracy 97

Sensitivity 96

Specificity 98

AUC 93

Kappa Score 94

This research provides insight into how MRI technology plays a crucial role in detecting
brain tumors, with the main focus being to interpret MRI images using deep learning
algorithms and detect cancer cells. The authors aimed to integrate these algorithms into
medical practice and in particular use data from open sources and multiple medical centres.

The authors used a CNN to successfully achieve their goals which they implemented using
the PyTorch machine learning library and coded in Python. We split the data into the 10
fold cross validation and various train-test splits (60-40, 70-30, and 80-20) and retrained the
model for the 30 epochs, with batch size 32 for every epoch.

The results were promising and the model illustrated considerable accuracy rates, especially
97%accuracy(as shown in Table 3) for the final epoch with 10-fold cross-validation. This
accuracy was also similar irrespective of the train-test splits. Moreover, the model manifested
excellent sensitivity, specificity, and kappa scores, approximately 96%, 98%, and 94%
respectively, particularly for the discrimination between benign and malignant nodes. These
performance metrics still remain stable despite the variation of data splits.
 Page 7130 of 7136
Sulekha Das / Afr.J.Bio.Sc. 6(5) (2024). 7113-7126

Table 4. Comparison of AUCs(train/test partition) for30 epochs

50/50 60/40 70/30 80/20

Table 5. Comparison of AUCs(10 fold cross validation) for 30 epochs

10 th fold 9 th fold 8 th fold 7th fold

6th fold 5 th fold 4th fold 3rd fold

2nd fold 1st fold

 Page 7131 of 7136
Sulekha Das / Afr.J.Bio.Sc. 6(5) (2024). 7113-7126

The ROCs(Receiver Operating Characteristic) shown in Table 5 represent the performance

metrics of the designed CNN model in this experimentusing a 10-fold cross-validation
techniques and train/ test partitions. ROC charts are among the most widely used graphical
methods to evaluate binary classifiers. The receiver operating characteristic curve shows the
tradeoffs between true positive rate and false positive rate. The experiment proves that the
proposed classifier, developed by the authors in the said study, was better performing
compared to most of the models discussed in the literature review. The proposed classifier
showed outstanding performance when the AUC/ROC values were being evaluated using 10
fold cross validations and several train/ test partitions. The AUC value achieved by the
proposed model was 0.97, which is the highest possible value, illustrating great distinction
ability between the two classes. Hence this new CNN model seems to be very helpful in
distinguishing between the classes. High ROC values indicate the strong predictive ability of
model that can have practical implications in real world applications in the binary
classification problems.

One of the interesting findings of the research was the recognition of difficult cases where
benign nodules looked like malignant ones. In addition to these difficulties, the algorithm
proved to be quite robust in assignments, displaying its capabilities to recognize benign and
malignant nodules even in visually unclear cases.

Table 6. Accuracies for 80/ 20 train/ test partitions with reduced convolutional and max-

pooling layers

Train/ Test Accuracy Tumors

Partitions Metrics Prediction
Accuracy

80-20 Sensitivity 32

Specificity 07

Using Explainable Artificial Intelligence (XAI), we could explain the changes in model
performance that are observed when convolutional layers and max-pooling layers are reduced
by 1 as illustrated in Table 6. These techniques enable one to see the impact of changes in the
architecture of the model on different aspects of performance.

Here's a breakdown of the observed changes and potential explanations using XAI principles:
 Page 7132 of 7136
Sulekha Das / Afr.J.Bio.Sc. 6(5) (2024). 7113-7126

1. Recall for Malignancy Class Drop (0.86 to 0.32):

o Class-Specific Sensitivity: Recall evaluates the capacity of the model to
identify the positive samples of a certain class, in this case, the malignant
instances.

Explanation: The marked reduction in the recall shows that the proposed
model with fewer convolutional and max-pooling layers is not as successful in
identifying malignant instances. This is likely because:

Insufficient Feature Learning: However, this may lead to the fact that the
model does not learn rich or deep features that would help to distinguish
between malignant cases and others.

Loss of Fine-Grained Details: As one can gather, the model might neglect
important patterns necessary for classifying malignant cases due to the
simplification of the network architecture.

Class Imbalance Handling: If malignant cases are less frequent, decreasing

model capacity could bias the model for this class, which ultimately leads to a
higher number of missed malignant cases.

2. Specificity for Malignancy Prediction Drop (0.95 to 0.07):

o False Positives Increase: Specificity evaluates the model in correctly
identifying negative cases, which is a non-malignant instance.

Explanation: The dramatic reduction in specificity means that the model

misidentifies a large number of benign or normal cases as malignant. This
could be due to:

Lack of Feature Distinction: Such reduction may not help in making a

correct distinction between malignant and non-malignant cases due to poor
feature extraction.

Overgeneralization: If there were fewer layers, then the model would

become too general and may classify even benign cases as malignant.
 Page 7133 of 7136
Sulekha Das / Afr.J.Bio.Sc. 6(5) (2024). 7113-7126

Feature Overlap: It is possible that there are features that can assist in
discriminating malignancy that are not well learned or represented in the
model, thus misclassifying benign cases as malignant ones.

In general, this research reveals the ability of deep learning algorithms integrated into MRI
technology to correctly diagnose and classify brain cancer as a good instrument for medical
professionals in diagnosis and treatment planning.

Conclusion:

This paper demonstrates how transfer learning, especially the use of a deep convolutional
neural network (CNN), can help with the precise diagnosis of different types of brain cancer
through magnetic resonance imaging (MRI) scans. The model shown in the given paper uses
sigmoid and ReLU activation functions and successfully gets an outstanding score on every
evaluation metric that can be used to measure accuracy. Moreover, this shows how powerful
the model is in correctly classifying the images into their proper classes. The highest
accuracy that was achieved here manifests the model's success in amplifying early diagnosis
and customizing treatment plans for brain cancer patients for clinicians. Such screening is
expected to bring many benefits to the process, starting with the quick and precise
recognition of different types of tumors from MRI scans. This method has the prospect of
optimizing therapeutic outcomes and contributing to early detection that might ultimately
result in improved patient prognosis.

Bringing together several performance metrics like accuracy, sensitivity, specificity, kappa
and AUC (Area Under the ROC Curve) while analyzing brain cancer models is a significant
advance in the field of model evaluation. Normally, researchers usually pay attention to only
one or two metrics that typically comprise of accuracy. On the other hand, such a step might
leave out several other aspects of model performance, including a situation where different
metrics demonstrate various facets of the model's behaviour, especially in medical
applications.

Through the employment of all these metrics, our CNN model of brain cancer detection
offers a holistic assessment of its performance. Achieving the best accuracy across all the
metrics demonstrates our model's capability to excel in different modes of evaluation. This
integrational strategy is not only the evidence of model distinction but also the proof of
feasibility of such concept in practical health-care domain.
 Page 7134 of 7136
Sulekha Das / Afr.J.Bio.Sc. 6(5) (2024). 7113-7126

Finally, the use of advanced deep learning techniques in medical image analysis as presented
in the research indicates the likelihood of the development of early diagnosis and better living
of patients in the context of oncology, especially in the case of brain cancer. This advanced
approach shows us how to harness the potential of artificial intelligence to improve health
systems and thus patients' well-being.

Future Scope
The actual accuracy score of around 98% by the developed model with the help of all
performance metrics suggests best possibilities for future applications in the field of medical
image analysis. The efficacy of this model demonstrates its potential is not limited to other
disease datasets. Instead, it shows the way to the early identification and classification of
numerous other medical conditions.

In the upcoming period, the extent of this research can be augmented by working jointly with
healthcare agencies and hospitals to collect varied datasets. Researching different ways to do
the tasks, including applying reinforcement learning methods, can help to raise the level of
model performance. These may entail tweaking the current model architecture, implementing
new neural network machines, or application of advanced algorithms for the best outcomes.

Furthermore, this model can be developed for present-day diagnosis and treatment
monitoring, which helps in the creation of more personalized and timely interventions. The
integration of artificial intelligence into healthcare is a rapidly growing field, and the success
illustrated in this study gives the basis for further innovation and testing of various solutions
to enhance early disease diagnosis and improve patients' outcomes.

Data Availability Statement (DAS): The dataset used in this study for the brain tumor
classification was acquired from kaggle.com and this does not violate the protection of
human subjects, or other valid ethical, privacy, or security concerns

Data Availability Statement (DAS): The dataset used in this study for the brain tumor
classification was acquired from kaggle.com and this does not violate the protection of
human subjects, or other valid ethical, privacy, or security concerns

References
Afshar, P., Mohammadi, A., &Plataniotis, K. N. (2018, October). Brain tumor type

classification via capsule networks. In 2018 25th IEEE international conference on

image processing (ICIP) (pp. 3129-3133). IEEE.

 Page 7135 of 7136
Sulekha Das / Afr.J.Bio.Sc. 6(5) (2024). 7113-7126

Ayadi, W., Elhamzi, W., Charfi, I., &Atri, M. (2021). Deep CNN for brain tumor

classification. Neural processing letters, 53, 671-700.

Badža, M. M., &Barjaktarović, M. Č. (2020). Classification of brain tumors from MRI

images using a convolutional neural network. Applied Sciences, 10(6), 1999.

Cheng, J., Huang, W., Cao, S., Yang, R., Yang, W., Yun, Z., ... & Feng, Q. (2015). Enhanced

performance of brain tumor classification via tumor region augmentation and

partition. PloS one, 10(10), e0140381.

Dang, L. M., Hassan, S. I., Im, S., & Moon, H. (2019). Face image manipulation detection

based on a convolutional neural network. Expert Systems with Applications, 129,

156-168.

Deb, D., Khang, A., & Chaudhuri, A. K. (2024). Fuzzy Thresholding-Based Brain Image

Segmentation Using Multi-Threshold Level Set Model. In Driving Smart Medical

Diagnosis Through AI-Powered Technologies and Applications (pp. 118-129). IGI

Global.

Deepak, S., & Ameer, P. M. (2019). Brain tumor classification using deep CNN features via

transfer learning. Computers in biology and medicine, 111, 103345.

Esteva, A., Kuprel, B., Novoa, R. A., Ko, J., Swetter, S. M., Blau, H. M., &Thrun, S. (2017).

Dermatologist-level classification of skin cancer with deep neural networks. nature,

542(7639), 115-118.

Hu, A., &Razmjooy, N. (2021). Brain tumor diagnosis based on metaheuristics and deep

learning. International Journal of Imaging Systems and Technology, 31(2), 657-669.

Huang, Z., Du, X., Chen, L., Li, Y., Liu, M., Chou, Y., & Jin, L. (2020). Convolutional

neural network based on complex networks for brain tumor image classification with

a modified activation function. IEEE Access, 8, 89281-89290.

 Page 7136 of 7136
Sulekha Das / Afr.J.Bio.Sc. 6(5) (2024). 7113-7126

Ilic, I., &Ilic, M. (2023). International patterns and trends in the brain cancer incidence and

mortality: An observational study based on the global burden of disease. Heliyon,

9(7).

Kumar, S., &Mankame, D. P. (2020). Optimization driven deep convolution neural network

for brain tumor classification. Biocybernetics and Biomedical Engineering, 40(3),

1190-1204.

Litjens, G., Kooi, T., Bejnordi, B. E., Setio, A. A. A., Ciompi, F., Ghafoorian, M., ... &

Sánchez, C. I. (2017). A survey on deep learning in medical image analysis. Medical

image analysis, 42, 60-88.

Litjens, G., Sánchez, C. I., Timofeeva, N., Hermsen, M., Nagtegaal, I., Kovacs, I., ... & Van

Der Laak, J. (2016). Deep learning as a tool for increased accuracy and efficiency of

histopathological diagnosis. Scientific reports, 6(1), 26286.

Machhale, K., Nandpuru, H. B., Kapur, V., &Kosta, L. (2015, May). MRI brain cancer

classification using hybrid classifier (SVM-KNN). In 2015 International Conference

on Industrial Instrumentation and Control (ICIC) (pp. 60-65). IEEE.

Mehnatkesh, H., Jalali, S. M. J., Khosravi, A., &Nahavandi, S. (2023). An intelligent driven

deep residual learning framework for brain tumor classification using MRI images.

Expert Systems with Applications, 213, 119087.

Özyurt, F., Sert, E., &Avcı, D. (2020). An expert system for brain tumor detection: Fuzzy C-

means with super resolution and convolutional neural network with extreme learning

machine. Medical hypotheses, 134, 109433.

Parmar, C., Grossmann, P., Bussink, J., &Lambin, P. (2017). Machine learning methods for

quantitative radiomic biomarkers. Machine learning applications for Radiomics, 5,

125.
 Page 7137 of 7136
Sulekha Das / Afr.J.Bio.Sc. 6(5) (2024). 7113-7126

Paul, J. S., Plassard, A. J., Landman, B. A., &Fabbri, D. (2017, March). Deep learning for

brain tumor classification. In Medical Imaging 2017: Biomedical Applications in

Molecular, Structural, and Functional Imaging (Vol. 10137, pp. 253-268). SPIE.

Ryu, Y. J., Choi, S. H., Park, S. J., Yun, T. J., Kim, J. H., & Sohn, C. H. (2014). Glioma:

application of whole-tumor texture analysis of diffusion-weighted imaging for the

evaluation of tumor heterogeneity. PloS one, 9(9), e108335.

Saba, T. (2020). Recent advancement in cancer detection using machine learning: Systematic

survey of decades, comparisons and challenges. Journal of Infection and Public

Health, 13(9), 1274-1289.

Sajjad, M., Khan, S., Muhammad, K., Wu, W., Ullah, A., &Baik, S. W. (2019). Multi-grade

brain tumor classification using deep CNN with extensive data augmentation. Journal

of computational science, 30, 174-182.

Sasikala, M., &Kumaravel, N. (2008). A wavelet-based optimal texture feature set for

classification of brain tumours. Journal of medical engineering & technology, 32(3),

198-205.

Sharif, M. I., Khan, M. A., Alhussein, M., Aurangzeb, K., & Raza, M. (2021). A decision

support system for multimodal brain tumor classification using deep learning.

Complex & Intelligent Systems, 1-14.

Skogen, K., Schulz, A., Dormagen, J. B., Ganeshan, B., Helseth, E., & Server, A. (2016).

Diagnostic performance of texture analysis on MRI in grading cerebral gliomas.

European journal of radiology, 85(4), 824-829.

Ting, F. F., Tan, Y. J., & Sim, K. S. (2019). Convolutional neural network improvement for

breast cancer classification. Expert Systems with Applications, 120, 103-115.

Verma, R., Zacharaki, E. I., Ou, Y., Cai, H., Chawla, S., Lee, S. K., ... &Davatzikos, C.

(2008). Multiparametric tissue characterization of brain neoplasms and their

 Page 7138 of 7136
Sulekha Das / Afr.J.Bio.Sc. 6(5) (2024). 7113-7126

recurrence using pattern classification of MR images. Academic radiology, 15(8),

966-977.

Wulczyn, E., Steiner, D. F., Xu, Z., Sadhwani, A., Wang, H., Flament-Auvigne, I., ...

&Stumpe, M. C. (2020). Deep learning-based survival prediction for multiple cancer

types using histopathology images. PloS one, 15(6), e0233678.

Zacharaki, E. I., Wang, S., Chawla, S., Soo Yoo, D., Wolf, R., Melhem, E. R., &Davatzikos,

C. (2009). Classification of brain tumor type and grade using MRI texture and shape

in a machine learning scheme. Magnetic Resonance in Medicine: An Official Journal

of the International Society for Magnetic Resonance in Medicine, 62(6), 1609-1618.

Zia, R., Akhtar, P., & Aziz, A. (2018). A new rectangular window based image cropping

method for generalization of brain neoplasm classification systems. International

Journal of Imaging Systems and Technology, 28(3), 153-162.