Submitted by:

Muhammad Ahmad Malik (BCS07203072)

Junaid Hassan (BCS07203064)
Sunaina Iltaf (BCS07203065)

Submitted To:

Sir Zaman

Degree:

Bs-Cs

Semester VII
 Topic: Brain Tumor Classification

Abstract:

Magnetic resonance imaging (MRI) manual tumor diagnosis is a laborious process that can result
in human error, erroneous tumor type identification, and false detection. In order to streamline
the complex medical procedures and alleviate the burden on doctors, a deep learning framework
is introduced for the categorization of brain tumors. Brain image analysis is done on publicly
accessible datasets like Brats and Kaggle. AlexNet, VGG16, and ResNet50 are three Deep
Convolution Neural Network (DCNN) architectures that have been trained beforehand, are used
to implement the suggested model. The pre-trained DCNN architecture's features include
retrieved using these transfer learning architectures, and the Support Vector Machine (SVM)
classifier is utilized to classify the extracted features. Data enhancement techniques are used with
magnetic resonance imaging (MRI) to prevent over fitting of the network. For the datasets from
Kaggle and Brat, the suggested methodology yields Overall accuracy of 98.28% and 97.87%
without data augmentation, and 99.0% and 98.86% with data augmentation, respectively. For the
identical datasets, the Receiver Operator Characteristic (ROC) Area Under Curve (AUC) is
0.9978 and 0.9850. When compared to the other two networks, the outcome demonstrates
ResNet50 is the most effective at categorizing brain tumors

Introduction

A brain tumor is the development of abnormal cells in the human brain. Although the exact cause
of brain tumors is unknown, several risk indicators can be used to determine the stage of the
tumor There are two types of brain tumors: primary and secondary fall. The main brain tumor
usually begins in the brain, or any part of the brain, and does not spread to other body organs.
Among tumors, malignant and benign are the most prevalent forms. The secondary brain tumor
spreads from many places of the body rather than starting in the brain itself. The majority of
malignant tumors are classified as secondary type tumors. Meningiomas and gliomas are two
other subtypes of benign tumors that are considered low-grade tumors. Glioblastoma and
astrocytoma are two types of high-grade tumors that are considered malignant.
Imaging techniques such as Single Photon Emission Computer Tomography (SPECT), Positron
Emission Tomography (PET), Computed Tomography (CT), and Magnetic Resonance Imaging
(MRI) are mostly used to examine brain images. Among these, magnetic resonance imaging
(MRI) and computed tomography (CT) are the most commonly utilized because of their
highresolution image quality and accessibility. MRI imaging is frequently used to detect brain
tumors instead of CT imaging since it may look at diseased or other physiological changes in live
tissues.. Adult brain tumors of the usual type are called gliomas, and they can be recognized
using magnetic resonance imaging (MR) scans of many types, including fluid-attenuated
inversion recovery (Flair), T2-weighted, T1-weighted, and T2-weighted contrast-enhanced (T1c).
A key component of the therapeutic process is early tumor detection. The radiologist classifies is
the MR image normal or abnormal using classification techniques. The type of tumor is
identified for the alternative treatment method if the outcome is unusual. Manual classification is
a costly and time-consuming task. Additionally, when it comes to differentiating tumors, they
may receive conflicting results from different observers or similar results from identical
observers. Automatic classification methods are therefore required. Deep transfer learning uses
an already-proven deep model to adjust it rather than creating a brand-new one. This lowers the
expenses associated with intricate layer parameters and time-consuming validation procedures.
The following is a list of the proposed framework's major contributions.

• When using a multi SVM classifier rather than a stand-alone Softmax classifier, a notable
improvement in performance is seen with the deep CNN features of transfer learned
models.
• Since the suggested framework uses SVM Classifier to train data augmentation for the
purpose of classifying brain MR images, it achieved the highest classification accuracy
when compared to previous relevant research efforts.
• Because of data augmentation, the suggested framework performs remarkably well
during training even with a small number of training examples.

• The capacity categorize brain MR pictures with a training accuracy of 99% without the
necessity for previous segmentation is among this research's most important
contributions..
• Moreover, our improved pretrained DCNN model with SVM classifier framework uses
fewer computing parameters than VGG-16, ResNet-50, and Alexnet due to its lower execution
time and superior accuracy and speed.

• This is how the rest of the framework is set up. Under Section 2, a discussion of relevant
works on the automation of brain tumor categorization techniques is conducted. The materials
and procedures needed for this work were briefly explained in Section 3. Section 4 of the
proposed technique explains the procedures involved in the Deep Convolution Neural Network
designs' (DCNN) pre-training Section 5 provides more information about the evaluation and
results validation of the suggested framework. Finally, Section 6 provides examples of
conclusions.

Literature Review:

Beyond machine learning, emerging technologies—especially artificial intelligence have a huge

impact on the medical field and greatly aid the diagnostic capabilities of the medical
professionals. In magnetic resonance imaging (MRI), machine learning techniques for picture
segmentation and classification help radiologists obtain a second opinion when interpreting
medical records. Consequently, in order to obtain accurate diagnosis and avoid surgery, it is
imperative to build an efficient diagnostic tool for tumor segmentation and classification using
MRI images. Research on the classification and division of brain MRI images has taken many
forms. We examined a few international journals that use deep learning to detect and classify
brain tumors: Pre-trained VGG16 was employed in a method proposed by Jain et al. [1] for the
diagnosis of Alzheimer's disease using MRI scan pictures. Yang et al. [2] used two distinct
architectures—AlexNet and GoogLeNet—to compare gliomas. The results of the simulation
indicate that GoogLeNet outperforms AlexNet in terms of performance between these two
networks. In order to detect meningiomas, gliomas, and pituitary tumors An Abiniwanda et al.
convolutional neural network was employed [3] without first segmenting the data. The results
showed an accuracy of 98.51% during training and 84.19% during validationThe primary benefit
of utilizing this classifier is its automatic nature, which eliminates the necessity for physical
tumor segmentation during collection. Zhang et al. [4] created a classifier for brain MR images
that incorporated SVM and particle swarm optimization (PSO). Instead of using kernel SVM, the
researchers employed the wavelet entropy technique. Wang et al. employed feedforward neural
networks, an artificial bee colony, and particle swarm optimization. [5]. A range of studies using
deep learning networks, such as Convolutional Neural Networks, in conjunction with brain tumor
classification and segmentation are reviewed [6]. Adaboost and SVM were coupled by the
researchers [7] to raise the model's accuracy to 99.45%. Saritha et al. [8] employed a
probabilistic neural network (PNN) using the wavelet entropy technique. Nayak et al. [9] used
probabilistic principal component analysis, the AdaBoost technique, and a random forest
classifier to extract features from 2-D DWT for feature reduction. A method for classifying brain
malignancies was created by Sheikh Basheera and Ram [10]. It involves first segmenting the
tumor from an MRI picture. Next a convolutional neural network that has been trained using
stochastic gradient descent is used to extract the segmented region Khwaldeh et al. [11] presented
a framework for classifying brain MRI images into normal and abnormal as well as a grading
system for recognizing unhealthy brain images as low and high grades by modifying the Alex-
Net CNN model. In order to classify multi-grade malignancies, Sajjad et al. [12] enhanced their
data augmentation method with a pre-trained VGG-19 CNN Model on Mri images. Using k-
nearest neighbor and multinomial logistic regression, Carlo et al. [13] proposed a classification
scheme for MRIs of pituitary adenomas tumors. The method produced an accuracy of 83% on
multinomial logistic regression and 92% on a knearest neighbor test, with an AUC curve of
98.4%. Das et al. [14] trained a CNN model to analyze several types of brain tumors using an
image processing methodology, achieving 94.39% accuracy and 93.33% precision. Rehman et al.
[15] used three different pre-trained CNN models (VGG16, AlexNet, and GoogleNet) to classify
brain malignancies into meningioma, glioma, and pituitary. Using this Transfer learning
mechanism, VGG16 attains the highest accuracy of 98.67%. Additionally, Swathi et al. [16]
devised a plan utilizing the CE-MRI dataset was utilized to test various types of algorithms, such
as Alxenet, VGG-16, and VGG-19. Other diseases like diabetes may be predicted in humans with
the maximum accuracy by using machine learning classifiers like SVM. [17]. A major obstacle to
the widespread use of deep learning in medical healthcare is the scarcity of labeled data. Recent
developments in deep learning models across other areas have demonstrated that the accuracy of
the result increases with the amount of data. In the literature search, multiple pre-trained CNNs
and deep learning are used to provide data augmentation and segmentation. The categorization
effectiveness of transfer learning is the main area of study. The most widely utilized pre-trained
models in the literature are VGG-16, ResNet50, and Alexnet, which are pretrained on numerous
datasets like ImageNet. The proposed system uses transfer learning to classify brain tumors from
MRI images using pre-trained models. Minor modifications are made to the transfer Learning
Models ResNet 50, AlexNet, and VGG16. This study suggests an effective and automated
categorization method for dividing MRI brain tumors into normal and abnormal categories. Brain
MR images are used to train a deep transmission of learning CNN model for feature extraction,
which are then classified using the well-known SVM classifier. Next, a thorough evaluation of
the suggested system is conducted. Among all the relevant articles, the suggested framework has
the highest classification accuracy when tested on the datasets of Kaggle [18] and Brats [19, 20].
Additionally, the suggested method produces satisfactory outcomes with fewer training examples
in less time.

Problem Statement

A brain tumor is a serious condition brought on by the unchecked proliferation of brain tumor-
causing cells. The yearly death rate from brain cancer is 21%. Magnetic reassurance images
contain disease-sensitive information that must be carefully recorded in order to identify the
region of concern, according to earlier studies. To classify images, saptio temporal relationships
are latent. Dealing with ROI requires the use of attention mechanisms.

Research Questions:

1. In what ways might sophisticated image processing methods be used to improve the
identification of modest, disease-sensitive information in brain tumor-related magnetic resonance
imaging (MRI)?

2. How can the region of interest (ROI) in brain tumor images be reliably captured? What
are the main obstacles to be overcome and what opportunities exist? How can these obstacles be
efficiently resolved to increase the accuracy of the diagnosis?

3. How can spatiotemporal relationships in pictures of brain tumors be efficiently described

and used to improve the reliability and accuracy of image classification?

4. How can critical areas in the brain images be better identified and analyzed, with a
particular emphasis on areas linked to tumor presence, by incorporating attention mechanisms
into the image classification process?
5. Considering the unchecked proliferation of malignant brain cells, what novel techniques
or approaches can be created to reduce the death rate related to brain cancer, and how can these
techniques is included into medical imaging technologies?

6. How can the intricacy of brain tumor identification be handled by machine learning and
deep learning algorithms to provide early diagnosis and prompt intervention to enhance patient
outcomes?

7. Taking into account the complexity of the disease and the requirement for a wide range of
knowledge, what part might multidisciplinary collaboration play in the creation of all-
encompassing solutions for brain tumor detection and classification?

8. How can the incorporation of temporal data into MRI sequences help us comprehend the
features and course of brain tumors better, and how can we use this data to enhance the precision
of our diagnoses?

9. How can the application of cutting-edge medical imaging technology for brain tumor
diagnostics be handled ethically during the research and development phase? What are the
possible obstacles and ethical concerns?

10. How can the results of earlier investigations into the classification of brain tumors be
combined and improved upon to produce a more comprehensive and successful strategy for early
identification and treatment?

Methodologies

There are three main types of brain tumors in the Kaggle dataset used for brain tumor
classification: meningiomas, pituitary tumors, and gliomas. With their own features and clinical
ramifications, these tumors are different pathological entities. Brain glial cells are the source of
gliomas, which can range in aggressiveness from low-grade to high-grade cancers. The pituitary
gland is the site of pituitary tumor development, which impairs hormone production and causes a
range of endocrine diseases. Meningiomas are benign tumors that originate from the meninges,
which are the membranes that surround the brain and spinal cord. However, their size and
location can affect their ability to cause symptoms.

1. Glioma Tumor
 One kind of brain tumor that starts in the brain's glial cells is called a glioma. These
tumors are categorized according to where they are located in the brain and the kind of
cells they impact, and they can either be benign or malignant. Gliomas can develop in
different parts of the brain and are characterized by their infiltrative nature.

2. Pituitary Tumor:
The pituitary gland is a little gland near the base of the brain where pituitary tumors
grow. Depending on the size and hormone-secreting activity of the tumor, they can cause
a range of symptoms and interfere with hormone synthesis.. Generally speaking, pituitary
tumors can be categorized as non-functional (not producing hormones) or functioning
(producing hormones).

3. Meningioma Tumor :
Tumors called meningiomas start in the meninges, which are the membranes that
surround the brain and spinal cord and act as protective barriers. Most of the time, these
tumors are benign and expand gradually. Depending on their size and location,
 meningiomas can develop anywhere throughout the meninges and may produce
symptoms.

4. No Tumor:
There are probably brain scan photos in this class that don't reveal any malignancies.
These images serve as a control group or baseline for comparison with images showing
tumors. Having these negative samples in the dataset is crucial for the algorithm to
accurately differentiate between photographs with and without tumors.

Incorporating these four categories into the dataset will hopefully aid in the development of a
classification algorithm capable of distinguishing between photos of healthy brain tissue and
images of different types of brain cancer. In the long run, this all-encompassing approach
enhances diagnostic accuracy and patient care in clinical settings by allowing the building of
trustworthy models for automated brain tumour detection and classification.
Because it aids in making accurate diagnosis and treatment plans, brain tumour classification is
crucial to medical diagnostics. Thanks to advancements in deep learning, convolutional neural
networks (CNNs) have become powerful tools for automated image interpretation and may one
day increase the efficiency and accuracy of diagnostics. In this research, we suggest using
ResNet50, a deep convolutional neural network (CNN) architecture that is very good at image
recognition tasks, to classify brain tumours. We cover topics such as selecting an appropriate
dataset, picture preprocessing, model construction, model training, classifier selection, evaluation
metrics, and testing and validation procedures tailored to ResNet50 for brain tumour
classification.
A revolutionary convolutional neural network architecture called ResNet-50—an abbreviation for
Residual Network with fifty layers—was created by Microsoft Research. It is noteworthy for its
thoroughness and efficiency in handling image classification jobs, such as the difficult problem of
brain tumour classification. Convolutional layers, pooling layers, fully linked layers, and residual
blocks—important shortcut connections—make up ResNet-50's 50 layers, making it a
complicated design in its entirety. Essential for the network's capacity to learn and overcome the
vanishing gradient problem, these residual blocks enhance the network's capacity to detect
complicated characteristics in the input data.
Furthermore, ResNet-50 is often pre-trained on large image datasets such as ImageNet. In order
to provide the groundwork for more specialised tasks, such brain tumour classification, pre-
training trains the model to recognise broad characteristics in a variety of pictures. Skip
connections, sometimes called shortcut connections or identity mappings, are another
characteristic that sets ResNet-50 apart. By facilitating the smooth flow of gradients during
backpropagation, these connections lessen the occurrence of degradation problems, which are
common in deeper networks, and enhance the network's ability to maintain optimum performance
at longer depths. On several image classification benchmarks, ResNet-50 has consistently shown
state-of-the-art performance. Its ability to integrate depth and skip connections allows it to
capture intricate details seen in input photos. This makes it a great fit for tasks requiring advanced
feature extraction, such brain tumour classification. Because of its pre-trained weights on large-
scale datasets, ResNet-50 is also ideal for transfer learning situations. By making use of these
learned representations, researchers and practitioners may accomplish competitive performance
levels with little effort by drastically lowering the quantity of training data required for certain
tasks. ResNet50 is a powerful tool for image classification tasks, especially for brain tumour
classification, because to its depth, skip connection integration, and pre-trained weights. Its
widespread use by academics and practitioners in the fields of deep learning and medical imaging
attests to its significance and efficacy in pushing the frontiers of both fields.

Figure 1: ResNet50 Model Architecture

Every machine learning endeavour relies on the dataset's variety and quality. Classification of brain tumours
requires a dataset that includes many photos of these tumours. Research institutions, medical imaging archives,
and publicly available datasets such as those on Kaggle are good places to find these types of datasets. Various
brain tumours, such as gliomas, meningiomas, pituitary tumours, and others, should be included in these
databases. To be more precise, a collection of around 3,000 images distributed across these four categories
would provide a comprehensive representation of the variety of brain tumours. If you want a full view of the
tumour and to train a good model, it's best to have these photographs taken using many imaging modalities,
such CT and MRI scans. The famous data science and machine learning site Kaggle provided the dataset, which
had over 3,000 images of various brain tumours. You can trust this source to provide you with the high-quality,
hand-picked data you need to train your machine learning models. With a dataset this large and diverse,
machine learning models built on ResNet50 might train to differentiate between different types of brain
tumours, which would improve diagnosis accuracy and patient care.

Dataset No. of Images Classes

Brain Tumor Classification 3000 4

In order to prepare the dataset for training, preprocessing is required. This requires scaling the pictures
consistently, which is often done according to the requirements of the chosen convolutional neural network
(CNN) design and computing constraints. Improving convergence and stabilising training are two additional
benefits of pixel value normalisation to a consistent range. Since images of brain tumours may reveal variations
in illumination and contrast, techniques like intensity normalisation and histogram equalisation can be used to
enhance the reliability and uniformity of these images.
For feature extraction, the CNN model used is ResNet50, a variant of the ResNet architecture. ResNet50's depth
and skip connections are famous for its ability to train deeper networks more easily and to reduce the impact of
the vanishing gradient problem. Using its pre-trained weights from large-scale picture datasets like picture Net,
ResNet50 can easily extract complicated characteristics from brain tumour photographs. This minimises the
demand for considerable training from scratch.
As a part of ResNet50 training, the pre-trained model is fine-tuned using the brain tumour dataset. During
training, you may tweak key parameters like learning rate, batch size, and epoch count to get the most out of
your model. By using transfer learning approaches, we can keep the learned representations and prevent
overfitting by changing the weights of the convolutional layers while keeping the deeper layers fixed. Data
augmentation methods like as flipping, scaling, and rotation may be used to increase the dataset size and
improve the generalizability of the model. A classifier is used to sort the collected characteristics into different
tumour types after ResNet50 is used for feature extraction. Additional classifiers such as fully connected neural
networks (FCNNs) or support vector machines (SVMs) may be considered, even though the initial concept was
for using a Random Forest classifier. In contrast to FCNNs, which provide design flexibility and can be tailored
to the specific needs of the job at hand, SVMs excel at binary classification tasks and are resistant to overfitting.
Evaluation metrics are crucial for gauging the efficacy of the brain tumour categorization model. Common
metrics that reveal how well the model performs and can correctly classify different types of tumours include F1
score, recall, accuracy, and precision. Furthermore, domain-specific measures such as specificity and sensitivity
are important in medical diagnostics since they assess the model's performance in detecting positive and
negative instances, respectively.
To find out how effectively the trained model generalises, it is essential to test it on a separate dataset that wasn't
used for training. This testing set should include a broad range of brain tumour pictures, including those with
different sizes, locations, imaging modalities, and acquisition characteristics. The model's robustness and
practicality may be better understood by conducting a quantitative evaluation of its performance on the testing
dataset using validation metrics. The use of ResNet50 for brain tumour classification offers a promising way to
enhance and automate the diagnostic process. By following the steps outlined in this paper, researchers and
practitioners may effectively use deep learning algorithms to evaluate pictures of brain tumours in a timely and
accurate manner. Future work may include further honing the model's structure and training procedures, as well
as incorporating advanced approaches to boost performance, such as attention processes and multi-modal
fusion. Overall, the proposed approach represents a significant advancement for computer-assisted diagnosis in
neuroimaging and improved patient outcomes in clinical practice.
Train the Model
1. In our pursuit of optimal model training efficiency, we devise a function with several input parameters aimed
at optimizing the training process. Modifying the learning dynamics and improving the model's usefulness
are both made possible by these parameters.

2. Model (model):

The pre-trained model is sent as an argument to the function. Usually, a benchmark dataset like
ImageNet-1K, CIFAR-100, or CIFAR-10 was used to pre-train this model. We can take
advantage of learned features and hasten convergence when training by utilizing pre-trained
models.

3. Criterion (criterion):

The criterion uses a specified loss function to compute gradients. By measuring the discrepancy
between the ground truth and predicted labels, this function directs the training process and
allows the model to adapt its parameters.

4. Optimizer (optimizer):

During training, the optimizer function decides how to update the model's parameters. It depends
on two essential factors:

• Learning rate (lr): Determines the size of the steps executed during optimization, which
affects convergence stability and speed.

• Momentum: Improves optimization by utilizing previous gradients to reduce oscillations

and quicken convergence

4. Scheduler (scheduler):
During training, the learning rate is dynamically adjusted by the scheduler function. It adjusts the
learning rate depending on the following factors and is linked to an optimizer such as Stochastic
Gradient Descent (SGD):

• Optimizer (optimizer_ft): The optimizer that we usually employ is Stochastic Gradient

Descent (SGD).

• Step Size (step_size=7): Determines how frequently the rate of learning is modified.

• Gamma Value: Determines how much the learning rate can be adjusted.

5. Number of Epochs (num_epochs):

During training, this parameter specifies the number of complete runs over the whole dataset. In this case, we
are using an epoch range of 0–49. In each epoch, the model learns something new from a different set of data
via a series of repetitions. Over the duration of several training epochs, the model's performance is
incrementally enhanced by parameter adjustments. Based on our study, training the model throughout this range
of epochs achieves an accuracy of 98.28%, which shows that convergence and efficient dataset learning have
been achieved.

Expected Solutions:

Brain tumours are a major health concern, and one potential answer is to develop and implement
a state-of-the-art medical imaging system that can detect and diagnose these tumours more
accurately and with better treatment results. This method extracts subtle disease-sensitive
information from magnetic resonance imaging (MRI) using state-of-the-art image processing
algorithms. The main focus of this effort is to develop robust methods for reliable ROI detection
using machine learning models that can adapt to different patterns in MR images. Because brain
tumours evolve throughout time, it is necessary to use spatial-temporal modelling techniques,
such as 3D convolutional neural networks (3D CNNs) or recurrent neural networks (RNNs).
Attention processes will substantially facilitate the capacity to rank and examine relevant areas
inside brain pictures. Our ultimate goal is to optimise treatment strategies for brain cancer
patients based on their tumour characteristics in order to reduce the yearly mortality rate
associated with this disease. The ethical issues of patient privacy, data security, and appropriate
AI use in healthcare are also emphasized in the solution. To provide an all-encompassing and
efficient strategy, interdisciplinary collaboration between engineers, data scientists, and medical
specialists will be encouraged. Furthermore, the amalgamation of temporal data from magnetic
resonance imaging sequences and the amalgamation of outcomes from prior investigations will
augment the comprehensiveness and precision of brain tumor diagnosis.

Results and discussion

The classification performance of the VGG16, AlexNet, and ResNet50 models is evaluated. The
transfer learning methods on the Kaggle and Brats datasets use the VGG16, AlexNet, and
ResNet50 models. AlexNet achieved accuracy of 94.83% for Kaggle and 94.68% for the Brats
brain tumor dataset, VGG16 achieved 91.38% and 90.43% accuracy for the respective datasets.
Figure 2 displays the ROC curves for the Brats and Kaggle datasets using these networks

Figure 2: ROC curves of Kaggle datasets

.
AUC

Dataset VGG16 AlexNet ResNet50

KAGGLE 0.9325 0.9603 0.9978

Using the confusion matrix, Figures 3 and 4 compare pre-trained networks on the BRATS and
Kaggle datasets. The AUC values taken from the ROC curves are displayed in Table 5. This
result shows that by taking into account both confusion matrix and ROC features, ResNet50, one
of the three pre-trained models give good performance results for both the Kaggle and
BRATS datasets.

Figure 3: Comparison of results obtained from different pre-trained models on Kaggle dataset using
confusion matrix

Conclusion

In this framework, The three pre-trained models—AlexNet, ResNet-50, and VGG-16—are

effectively employed to classify brain tumors using the Transfer learning process. We assess the
three models mentioned above based on their performance indicators. In comparison to the other
two networks, ResNet50 outperforms them in the classification of brain tumors, according to the
simulation data. After training the framework with data augmentation and SVM classifier, the
suggested model, which used the Kaggle and BRATS datasets, achieved the best Classification
Accuracy in the shortest amount of computational time. As a result, brain tumor detection using
this study can be effectively applied in the medical field. These models have the potential to be
applied in the future to more comprehensive tumor categorization with fewer computational steps
and more subclasses.
Future Work;
Moving forward, there are various viable approaches to improving the deep learning framework for brain tumor
classification utilizing MRI data. For starters, using additional imaging modalities such as fMRI, DTI, or PET
scans might increase the amount of information available for categorization, thereby improving accuracy.
Second, fine-tuning hyperparameters such as learning rate and batch size might help to improve model
performance. Ensemble learning, which combines predictions from different deep learning models, provides a
way to build a more robust classifier. Furthermore, adding clinical data, such as demographics and genetic
information, may improve prediction power and provide individualized therapy recommendations. Creating an
interactive interface that allows physicians to interact with the model, view predictions, and offer feedback
would improve clinical decision-making. Exploring transfer learning approaches suited to specific patient
populations may increase generalizability. Implementing uncertainty estimate methods would provide
physicians with confidence intervals for predictions, which would help them make decisions. Optimizing the
algorithm to process MRI data in real time might allow for faster diagnosis and treatment planning. Validating
the model on external datasets and incorporating it into clinical trials would help to determine its efficacy and
influence on patient outcomes. Through these efforts, the deep learning framework can continue to evolve,
resulting in more accurate identification and treatment of brain tumors, which will benefit both patients and
physicians.

REFERENCES

[1] Jain, R., Jain, N., Aggarwal, A., Hemanth, D.J. (2019). Convolutional neural network based

Alzheimer’s disease classification from magnetic resonance brain images. Cognitive Systems
Research, 57: 147-159. https://doi.org/10.1016/j.cogsys.2018.12.015

[2] Yang, Y., Yan, L.F., Zhang, X., Han, Y., Nan, H.Y., Hu, Y.C., Ge, X.W. (2018). Glioma
grading on conventional MR images: A deep learning study with transfer learning. Frontiers
in Neuroscience, 12: 804. https://doi.org/10.3389/fnins.2018.00804

[3] Abiwinanda, N., Hanif, M., Hesaputra, S.T., Handayani, A., Mengko, T.R. (2019). Brain
tumor classification using convolutional neural network. In World Congress on Medical
Physics and Biomedical Engineering, 2018: 183-189. https://doi.org/10.1007/978-981-10-
9035-6_33

[4] Zhang, Y., Wang, S., Ji, G., Dong, Z. (2013). An MR brain images classifier system via
particle swarm optimization and kernel support vector machine. The Scientific World Journal,
2013: 130134. https://doi.org/10.1155/2013/130134
[5] Wang, S., Zhang, Y., Dong, Z., Du, S., Ji, G., Yan, J., Yang, J., Wang, Q., Feng, C., Phillips,

P. (2015). Feed‐ forward neural network optimized by hybridization of PSO and ABC for
abnormal brain detection. International Journal of Imaging Systems and Technology, 25(2):
153164. https://doi.org/10.1002/ima.22132

[6] Sravan, V., Swaraja, K., Meenakshi, K., Kora, P., Samson, M. (2020). Magnetic
resonance images based brain tumor segmentation-a critical survey. In 2020 4th International
Conference on Trends in Electronics and Informatics (ICOEI)(48184), pp. 1063-1068.
https://doi.org/10.1109/ICOEI48184.2020.9143045

[7] Ranjan Nayak, D., Dash, R., Majhi, B. (2017). Stationary wavelet transform and
AdaBoost with SVM based pathological brain detection in MRI scanning. CNS & Neurological
Disorders-
Drug Targets (Formerly Current Drug Targets-CNS & Neurological Disorders), 16(2): 137-149.

[8] Saritha, M., Joseph, K.P., Mathew, A.T. (2013). Classification of MRI brain images using
combined wavelet entropy based spider web plots and probabilistic neural network. Pattern
Recognition Letters, 34(16): 2151-2156. https://doi.org/10.1016/j.patrec.2013.08.017

[9] Nayak, D.R., Dash, R., Majhi, B. (2016). Brain MR image classification using
twodimensional discrete wavelet transform and AdaBoost with random forests. Neurocomputing,
177: 188-197. https://doi.org/10.1016/j.neucom.2015.11.034

[10] Basheera, S., Ram, M.S.S. (2019). Classification of brain tumors using deep features
extracted using CNN. In Journal of Physics: Conference Series, 1172(1): 012016.
https://doi.org/10.1088/1742-6596/1172/1/012016

[11] Khawaldeh, S., Pervaiz, U., Rafiq, A., Alkhawaldeh, R.S. (2018). Noninvasive grading of
glioma tumor using magnetic resonance imaging with convolutional neural networks. Applied
Sciences, 8(1): 27. https://doi.org/10.3390/app8010027

[12] 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. https://doi.org/10.1016/j.jocs.2018.12.003

[13] Carlo, R., Renato, C., Giuseppe, C., Lorenzo, U., Giovanni, I., Domenico, S., Mario, C.
(2019). Distinguishing functional from non-functional pituitary macroadenomas with a machine
learning analysis. In Mediterranean Conference on Medical and Biological Engineering and
Computing, Springer, 76: 1822-1829. https://doi.org/10.1007/978-3-030-31635-8_221

[14] Das, S., Aranya, O.R.R., Labiba, N.N. (2019). Brain tumor classification using
convolutional neural network. In 2019 1st International Conference on Advances in Science,
Engineering and Robotics Technology (ICASERT), pp. 1-5.
https://doi.org/10.1109/ICASERT.2019.8934603

[15] Rehman, A., Naz, S., Razzak, M.I., Akram, F., Imran, M. (2020). A deep learning-based
framework for automatic brain tumors classification using transfer learning. Circuits, Systems,
and Signal Processing, 39(2): 757-775. https://doi.org/10.1007/s00034-019-01246-3

[16] Swati, Z.N.K., Zhao, Q., Kabir, M., Ali, F., Ali, Z., Ahmed, S., Lu, J. (2019). Brain tumor
classification for MR images using transfer learning and fine-tuning. Computerized Medical
Imaging and Graphics, 75: 34-46. https://doi.org/10.1016/j.compmedimag.2019.05.001

[17] Reddy, D.J., Mounika, B., Sindhu, S., Reddy, T.P., Reddy, N.S., Sri, G.J., Swaraja, K.,
Meenakshi, K. and Kora, P. (2020). Predictive machine learning model for early detection and
analysis of diabetes. Materials Today: Proceedings. https://doi.org/10.1016/j.matpr.2020.09.522
[18] https://www.kaggle.com/jakeshbohaju/brain-tumor, accessed on 20 March 2021.

[19] Menze, B.H., Jakab, A., Bauer, S., Kalpathy-Cramer, J., Farahani, K., Kirby, J., Lanczi, L.

(2014). The multimodal brain tumor image segmentation benchmark (BRATS). IEEE

Transactions on Medical Imaging, 34(10): 1993-2024.

https://doi.org/10.1109/TMI.2014.2377694

[20] Kistler, M., Bonaretti, S., Pfahrer, M., Niklaus, R., Büchler, P. (2013). The virtual skeleton
database: an open access repository for biomedical research and collaboration. Journal of
Medical Internet Research, 15(11): e245. https://doi.org/10.2196/jmir.2930

[21] Krizhevsky, A., Sutskever, I., Hinton, G.E. (2012). ImageNet classification with deep
convolutional neural networks. Communications of the ACM, 60(6): 84-90.
https://doi.org/10.1145/3065386

[22] Simonyan, K., Zisserman, A. (2014). Very deep convolutional networks for large-scale
image recognition. arXiv preprint arXiv:1409.1556
[23] He, K., Zhang, X., Ren, S., Sun, J. (2016). Deep residual learning for image recognition. In Proceedings of
the IEEE Conference on Computer Vision and Pattern Recognition, pp. 770-778
https://doi.org/10.1109/CVPR.2016.90