UNVEILING THE FUTURE OF HEALTHCARE:

REVOLUTIONIZING BRAIN TUMOR DETECTION WITH

COMPUTER VISION AND AI

A PROJECT REPORT

Submitted by

Jaswanth Koripalli [Reg No:RA2011004010013]

P R S V Akhil [Reg No: RA2011004010030]
Pavan Kumar Pendyala [RegNo:RA2011004010060]
Under the guidance of
DR. K. Kalimuthu
(Associate professor, Department of Electronics & Communication Engineering)

in partial fulfillment for the award of the degree

of

BACHELOR OF TECHNOLOGY
in

ELECTRONICS & COMMUNICATION ENGINEERING

DEPARTMENT OF ELECTRONICS AND COMMUNICATION
ENGINEERING

COLLEGE OF ENGINEERING AND TECHNOLOGY

SRM. NAGAR, Kattankulathur-603203,

Chengalpattu District
MAY 2024
 SRM INSTITUTE OF SCIENCE AND
TECHNOLOGY
(Under Section 3 of UGC Act, 1956)

BONAFIDE CERTIFICATE

Certified that this project report titled “UNVEILING THE FUTURE OF

HEALTHCARE: REVOLUTIONIZING BRAIN TUMOR DETECTION
USING COMPUTER VISION AND AI” is the bonafide work of
“JASWANTH KORIPALLI [RA2011004010013]” who carried out the
project work under my supervision along with his batch mates P R S V
AKHIL [RA2011004010030] and PAVAN KUMAR PENDYALA
[RA2011004010060]. Certified further, that to the best of my knowledge the
work reported herein does not form any other project report on the basis of
which a degree or award was conferred on an earlier occasion for this or any
other candidate.

SIGNATURE SIGNATURE

Dr. K KALIMUTHU Dr. SHANTHI PRINCE

GUIDE HEAD OF THE DEPARTMENT
Associate Professor Dept. of Electronics &
Dept. of Electronics & Communication Engineering
Communication Engineering

Signature of the Internal Examiner Signature of the External Examiner

ii
 DECLARATION

We hereby declare that the Major Project entitled “UNVEILING THE FUTURE

OF HEALTHCARE: REVOLUTIONIZING BRAIN TUMOR DETECTION

WITH COMPUTER VISION AND AI” to be submitted for the Degree of

Bachelor of Technology is our original work as a team and the dissertation has not

formed the basis of any degree, diploma, associate-ship or fellowship of similar

other titles. It has not been submitted to any other University or institution for the

award of any degree or diploma.

Place:

Date:

Jaswanth Koripalli
[RA2011004010013]

P R S V Akhil
[RA2011004010030]

Pavan Pendyala
[RA2011004010060]

iii
 ACKNOWLEDGEMENTS

We would like to express my deepest gratitude to the entire management of SRM Institute of
Science and Technology for providing me with the necessary facilities for the completion of this
project.

We wish to express my deep sense of gratitude and sincere thanks to our Professor and Head of
the Department Dr. Shanthi Prince, for her encouragement, timely help, and advice offered to
me.

We are very grateful to my guide Dr. K. Kalimuthu Associate Professor, Department of

Electronics and Communication Engineering, who has guided me with inspiring dedication,
untiring efforts, and tremendous enthusiasm in making this project successful and presentable.

We would like to express my sincere thanks to the project coordinator Dr. A. Anilet Bala, panel
members Dr. A. Lavanya and Dr. A. Manikandan for their time and suggestions for the
implementation of this project.

We also extend my gratitude and heartful thanks to all the teaching and non-teaching staff of the
Electronics and Communications Engineering Department and to my parents and friends, who
extended their kind cooperation using valuable suggestions and timely help during this project
work.

iv
 ABSTRACT

In a groundbreaking development in medical science, scientists are leveraging the

capabilities of Computer Vision to revolutionize the detection of brain tumors. This

pioneering study explores the fusion of technology and healthcare, unveiling the

intricacies of the human brain through machine learning. This research uses a pre-

trained VGG-19 model, RESNET50 model, and a concatenated model,

incorporating model pruning techniques to enhance efficiency. Drawing on the vast

knowledge of ImageNet, Kera’s API, and TensorFlow play crucial roles in

seamlessly integrating cutting-edge algorithms with extensive datasets. As pixels

come together to convey a wealth of information, these tools unravel the mystery

of brain tumors, providing a glimpse into the future of precise diagnostics. Through

the collaboration of artificial intelligence and medical expertise, a new era in

healthcare emerges, where early detection is not just a possibility but a tangible

reality, reshaping patient care and instilling hope for many.

v
 TABLE OF CONTENTS

CHAPTER NO. TITLE PAGE NO

ABSTRACT v
LIST OF FIGURES viii
LIST OF TABLES ix
LIST OF ABBREVATIONS x

1 INTRODUCTION 1
1.1 Purpose 2
1.2 Aim and Objectives 3
1.3 Motivation 4
1.4 Innovative idea of the project 4
1.5 Scope and application of project 5

2 LITERATURE SURVEY 6
2.1 Works related to our project 6
2.2 Challenges and limitations 7

3 RESEARCH METHODOLOGY 9
3.1 Overview of the architecture 11
3.2 Data pre-processing 12
3.3 Realistic constraints 14

4 PROJECT IMPLEMENTATION 15
4.1 Software requirement specifications 15
4.2 Visual geometry group 19 (VGG19) 15
4.3 Residual network 50 (RESNET50) 16
4.4 Fusion net model 17

vi
5 FRAMEWORK 20
5.1 Feature Extraction of VGG19 20
5.2 Feature extraction of RESNET50 21
5.3 Feature extraction of FUSIONNET 22

6 RESULTS AND DISCUSSIONS 23

6.1 Evaluation 23
6.2 Hyperparameters 24
6.3 Results discussion 24

7 CONCLUSION AND FUTURE SCOPE 29

7.1 Conclusion 29
7.2 Future Scope 29

REFERENCES 41

vii
 LIST OF FIGURES

FIGURE NO. TITLE PAGE NO

3.1 System Architecture Diagram 11

3.2 Sample test images 11
3.3 Steps to Visualize how cropping works 13
3.4 Images After being cropped 13
3.5 Original Image 14
3.6 Images after Augmentation 14
4.1 VGG19 Architecture 15
4.2 RESNET50 Architecture 16
4.3 Fusion net Architecture 18
6.1 Confusion matrix of VGG19 25
6.2 Confusion matrix of RESNET50 25
6.3 Confusion matrix of Fusion net 25
6.4 Confusion matrix of Pruned model 26
6.5 Train and test accuracy plot of VGG19 26
6.6 Train and test accuracy plot of RESNET50 26
6.7 Train and test accuracy plot of Fusion net 27
6.8 Train and test accuracy plot of Pruned 27

viii
 LIST OF TABLES

TABLE NO. TITLE PAGE NO

6.1 Hyperparameters 24
6.2 Performance comparison of VGG19, RESNET50, 28
Fusion net and pruned models

ix
 LIST OF ABBREVIATIONS

AI Artificial Intelligence

MRI Magnetic Resonance Imaging

DLA Deep Learning Architecture

ML Machine Learning

VGG Visual Geometry Group

RESNET Residual Network

CNN Convolutional Neural Network

1D 1 Dimensional

2D 2 Dimensional

3D 3 Dimensional

SVM Support Vector Machine

DL Deep Learning

BT Brain Tumor

UCG Uncontrolled Cell Growth

WHO World Health Organization

CT Computed Tomography

x
 CHAPTER 1

INTRODUCTION

Brain tumors, the second leading cause of global mortality, can be benign or malignant. Benign
tumors pose minimal health risks due to slower growth and lack of infiltration, while malignant
tumors can cause significant dysfunction. Gliomas, meningiomas, and pituitary tumors are
common types, with gliomas being malignant, meningiomas benign, and pituitary tumors
generally benign [1]. Brain tumors, harmful intracranial masses, pose challenges in detection due
to the brain's complex structure. Medical imaging, particularly MRI, aids diagnosis, and image
processing techniques enhance tumor detection through features like contrast adjustment,
thresholding, and segmentation [2].

The central nervous system, comprising the brain and spinal cord, facilitates sensory information
transmission and actions throughout the body. The brain's anatomy includes the brain stem,
cerebrum, and cerebellum, each responsible for specific functions such as motor control, sensory
processing, and vital activities [3]. The article introduces an automatic method for brain tumor
detection and classification, incorporating preprocessing, feature extraction, and geometrical
family Support Vector Machine (SVM) for improved segmentation and classification, evaluated
on various datasets [4].

The paper employs transfer learning (TL) on MRI images for brain tumor diagnosis,
demonstrating high classification performance and accuracy across 5000 T2-weighted contrast-
enhanced images, addressing challenges posed by the complex brain structure and similar lesion
areas. The results highlight the effectiveness of TL models in identifying brain tumors in medical
images [5]. The paper reviews deep learning applications in brain tumor analysis, covering
segmentation, classification, and survival time prediction, discussing methods, frameworks, and
challenges. The survey highlights the increasing dominance of this topic in conferences and
journals since 2015 [6].

The research paper proposes an innovative system for brain tumor detection, employing
preprocessing, segmentation, feature extraction, and an optimized convolutional neural network
(CNN) with various optimization algorithms. Achieving a detection accuracy of 98.9%, the

1
system demonstrates superiority over contemporary optimization techniques, emphasizing its
potential for timely and accurate diagnosis [7]. The study conducts a bibliometric analysis on
machine learning-based automated brain tumor detection, emphasizing its emergence as a
significant research area for improved diagnosis and survival chances. The research assesses
trends, advancements, and future topics over the past 5 years to understand the state of the art in
brain tumor prediction [8].

The study presents a deep learning-based method for automated metabolic tumor volume (MTV)
segmentation using amino acid PET scans, achieving reliable and robust results in brain tumor
patients. The automated segmentation demonstrated clinical value for response assessment,
showing agreement with physician assessments and significant impact on disease-free and overall
survival outcomes [9]. The research addresses the challenges of brain tumor segmentation using
deep learning networks, particularly focusing on the RESNET model to enhance computational
efficiency and training processes. The article outlines the organization with sections on existing
tumor segmentation work, materials, methods, results, discussion, and conclusions for
comprehensive coverage [10].

1.1 Purpose

The purpose of a brain tumor detection project employing the RESNET50 and VGG19 models
is to create a computer vision algorithm that can detect brain tumors in medical images effectively
and efficiently. This can assist doctors and medical workers in more swiftly and accurately
diagnosing brain tumors, resulting in earlier treatment and improved patient outcomes.
RESNET50 and VGG19 are two deep learning models that have been found to perform
exceptionally well in image classification tasks. We can educate these models to recognize tumor-
related patterns and characteristics by training them on a huge dataset of brain MRI images.

To make sure the project is original, it is essential to use real data and properly credit any sources
used. In addition, the methods and code used in the project have to be unique and not taken
directly from other sources. The process of concatenating the outputs of the VGG19 and
RESNET50 models entails integrating their feature maps or forecasts in order to take advantage
of each model's advantages and increase performance. Due to the distinct architectures of VGG19
and RESNET50, which may capture distinct facets of the data and produce complementing
information, this method may be advantageous.

2
It is critical to use authentic data and correctly cite any sources used in the project to ensure
that it is not plagiarized.

Furthermore, the code and methods utilized in the project must be original and not plagiarized
from other sources. Concatenation of both VGG19 and RESNET50 models involves combining
the outputs of these two models in some way, typically by merging their feature maps or
predictions to leverage the strengths of both models for improved performance. This technique
can be beneficial as VGG19 and RESNET50 have different architectures and might capture
different aspects of the data, leading to complementary information.

Model pruning is a technique used to reduce the size of deep learning models by removing
unnecessary parameters, which helps in reducing computational resources required for training
and inference without significantly sacrificing performance. This technique involves identifying
and removing redundant or less important connections, neurons, or layers from the model. In the
context of concatenating VGG19 and RESNET50, model pruning can be applied to the
concatenated model to further optimize its size and computational efficiency. By pruning
redundant connections or layers from the concatenated model, we can create a more streamlined
and efficient model while retaining the benefits of combining VGG19 and RESNET50. This can
result in faster inference times and lower resource requirements, making the concatenated model
more practical for real-world deployment in medical image analysis applications.

Overall, the purpose of a RESNET50 and VGG19 brain tumor identification project is to provide
a beneficial tool for medical practitioners that can assist improve patient outcomes and save lives.

1.2 Aim and Objectives

All that exists in a tumor is an overabundance of cells that develop out of control. The growth of
brain tumor cells eventually causes them to consume all the nutrition intended for the healthy
cells and tissues, leading to brain failure. A brain tumor is a highly serious condition that kills a
lot of people. To enable early diagnosis, a technique for detecting and classifying brain tumor’s
is available. In clinical diagnosis, classifying cancer is the most difficult endeavor.

This project focuses on one such system that classifies tumor’s using a VGG19 for MRI pictures
of various individuals and employs computerized techniques to find tumors. Brain tumors can be
found in MRI pictures of cancer patients using a variety of image processing techniques,
including feature extraction, image segmentation, and image enhancement.

3
Concatenation refers to the process of combining the outputs of both VGG19 and RESNET50
models. By combining the features learned by these models, the concatenated model can
potentially capture a more comprehensive representation of the input images, leading to improved
performance in tumor detection and classification tasks.

Additionally, the technique of model pruning is applied to the concatenated model. Model
pruning involves removing unnecessary connections, parameters, or layers from a neural network
without significantly impacting its performance. This process helps in reducing the computational
complexity and memory footprint of the model, making it more efficient for deployment in
resource-constrained environments such as medical diagnostic systems.

By leveraging the strengths of VGG19 and RESNET50 through concatenation and optimizing
the model using pruning techniques, the project aims to enhance the accuracy and efficiency of
brain tumor detection and classification in MRI images, ultimately aiding in early diagnosis and
treatment of this critical disease.

1.3 Motivation

The primary goal of the study is to identify the optimum Deep Learning Architecture for brain
tumor diagnosis. Therefore, it can be helpful in situations where we need to confirm if a tumor
is positive or negative. It can identify a tumor from an image and indicate whether a tumor is
positive or not.

This study focuses on a system that employs computer-based techniques to identify tumor
blocks and categories the type of tumor on MRI scans of various patients by applying the
Convolution Neural Network Algorithm.

1.4 Innovative Idea of the project

The main innovation of the project is the development of software aimed at aiding doctors in
accurately identifying tumors within the brain. This software utilizes deep learning architectures,
specifically VGG19 and RESNET50, to enhance the accuracy of tumor detection.

In addition to comparing the accuracies of these two architectures individually, the project
explores the effectiveness of combining them through concatenation. This concatenation
technique involves merging the outputs of both VGG19 and RESNET50 models to leverage the
strengths of each architecture and potentially improve overall performance in tumor detection.

4
Furthermore, the project incorporates a technique known as model pruning into the concatenated
model. Model pruning involves systematically removing unnecessary parameters from the model
to reduce its size and computational complexity while preserving or even enhancing its
performance. By implementing model pruning, the concatenated model becomes more
streamlined and efficient without sacrificing accuracy, ultimately enhancing its practical usability
in clinical settings.

1.5 Scope and application of project

The scope and application of deep learning architectures in brain tumor detection is vast and
promising. CNN, residual neural networks (RESNET), and other DL architectures have
demonstrated significant promise in the precise identification and diagnosis of brain tumors from
medical pictures.

Concatenating feature maps from other architectures, such VGG19 and RESNET50, is one
method for producing a more thorough representation of the input data. This concatenated model
may increase the precision and robustness of tumor identification by utilizing the advantages of
each design.

Moreover, model pruning is another technique that can be applied to the concatenated model.
Model pruning involves reducing the size of the network by removing unnecessary connections
or parameters, thus improving efficiency without sacrificing performance. This technique can
help streamline the concatenated model, making it more computationally efficient while
maintaining its diagnostic accuracy.

In summary, the concatenation of VGG19 and RESNET50, along with model pruning, presents
a promising approach in deep learning for brain tumor detection. By combining these techniques,
researchers can potentially develop more efficient and accurate models, leading to improved
diagnosis and patient outcomes.

Early brain tumor detection can lead to better treatment outcomes and higher survival rates. Deep
learning architectures can assist in this regard. Deep learning architectures have the potential to
significantly increase the speed and accuracy of brain tumor diagnosis, which would benefit
patients and increase survival rates.

5
 CHAPTER 2

LITERATURE SURVEY

2.1 Works related to our project

According to this study, a customized VGG19 network that uses both hand-crafted and deep MRI
scan features can be used to identify brain tumors. The method generated accuracy of 96.08%
and an Accuracy of 99% on a dataset of 306 MRI images, demonstrating that the VGG19 network
may perform better for this job when deep and handmade features are mixed [11].

This technique is used in the authors' proposed convolutional neural network (CNN)-based deep
learning strategy for brain tumor detection. The network receives input from MRI images, and
after a number of convolutional and pooling layers, moves on to fully linked layers for
classification. The proposed method was evaluated on a dataset of 150 MRI scans, and it
generated accuracy and sensitivity results of 98.7% and 97.1%, respectively. The results indicate
that the suggested CNN-based strategy has promise as a brain cancer identification tool [12].

The authors advocate using convolutional neural networks (CNNs), a deep learning approach, to
precisely identify brain tumors. The network receives input from MRI images, and after a number
of convolutional layers and pooling layers, moves on to fully linked layers for classification [13].

In the research by Bhandary A., a unique deep learning framework for the identification of lung
anomalies using a combination of chest X-ray and lung CT scan pictures is described. Although
some studies have employed deep learning approaches for lung problem detection, the suggested
framework is noteworthy for its distinct methodology and proven ability to identify multiple
kinds of abnormalities in the lungs [14].

The study makes a foundation with the purpose of providing doctors with reliable brain picture
analysis and therapy planning based on early diagnosis. Utilizing a convolutional neural network
(CNN), the framework for classification, correctly identified brain tumors using a set of brain
Magnetic resonance imaging scans with an accuracy of 97.14 percent. The paper was published
in Neural Computing and Applications in 2020 [15].

6
This study evaluated their suggested method using a dataset of 20 MRI images, indicating that
the proposed method can be an effective tool for segmenting brain regions in medical images.
The paper was published in Current Medical Imaging in 2016 [16].

The CNN model in this work, which was trained and tested using a dataset of brain MRI images,
classified brain tumors. The study demonstrates that deep learning can be an effective tool for
analyzing big data in medical imaging and improving the accuracy of brain tumor detection [17].

The study's proposed method was tested using a dataset of EEG signals from epileptic patients,
and it had an average detection accuracy of 91.6%. The research demonstrates that the suggested
technique can be a useful instrument for identifying interictal spike activity. in EEG signals,
which can aid in the diagnosis and treatment of epilepsy. The paper was published in Australasian
Physical & Engineering Sciences in Medicine in 2017 [18].

This study uses a dataset of MRI images to train and test their deep transfer learning model, which
achieved an accuracy of 97.9% for brain abnormality classification. The study demonstrates that
transfer learning can be an effective tool for automated medical image analysis, especially when
the amount of labeled data is limited. The paper was published in Cognitive Systems Research
in 2019[19].

This study uses a dataset of MRI images to train and test their deep learning model, which got an
accuracy of 99.6% for brain tumor detection and 97.7% for tumor segmentation. The work shows
that deep learning can be a useful technique for precise and automated brain tumor detection and
segmentation in medical pictures, which can help with brain tumor diagnosis and therapy [20].

2.2 Challenges and limitations

Limited Data: VGG19 can struggle with small datasets, which is a particular concern in medical
image analysis, where the number of images may be limited. In such cases, the model may overfit
to the training data, or it may fail to generalize well to new data.

Computational Resources: VGG19 is a very deep network that requires significant

computational resources to train and run. This can make it less practical for some settings,
especially those limited resources.

False Negatives: False negatives can occur when a model fails to detect a tumor that is present in
an image. false negatives, and it is important to ensure that the model designed to reduce risk.
7
Interpretability: While VGG19 is a powerful model, it can be difficult to interpret. This can be
a limitation in medical image analysis, where it is important to understand how a model is making
its decisions.

Limited Attention Mechanisms: VGG19 does not have any built-in attention mechanisms,
which are useful for highlighting regions of interest in an image.

8
 CHAPTER 3

RESEARCH METHODOLOGY

Numerous different DLAs, both standard and customized, are proposed throughout the research
literature as potential methods for locating anomalies in medical imaging. Constructing, training,
testing, and validating the architecture in order to meet the requirements of a particular test are
some of the many complex steps involved in the creation of a new DLA from the initial concept.
As a consequence of this, the majority of previous efforts to address a disease detection problem
have consisted of modifying the validated DLAs that can already be found in the literature. Before
selecting and putting an architecture into action, it is essential to have a solid understanding of
its structure, as well as the degree of effort involved in putting it into action, the initial tweaking,
and the validation procedures.

In this particular research project, the identification of brain tumors from the studied MRI
material is carried out with the use of the VGG-19 and RESNET50 Deep Learning Architectures.
Because the experiments conducted for this investigation have shown that RESNET50
outperforms VGG19 on several image classification benchmarks, especially on larger and more
complex datasets. However, VGG19 can still perform well and may be a better choice for smaller
datasets or when computational resources are limited. So, we have used them achieve greater
results.

In the realm of medical imaging, particularly in the detection of anomalies like brain tumors,
deep learning architectures (DLAs) have emerged as promising tools. These DLAs, both
standardized and customized, are extensively explored in research literature. However, creating
a new DLA involves intricate steps such as constructing, training, testing, and validating the
architecture to suit specific testing requirements. Consequently, many efforts in disease detection
focus on modifying existing DLAs rather than building entirely new ones. Understanding the
structure, effort involved, and validation procedures of an architecture before deployment is
crucial.

In this specific research project, the identification of brain tumors from MRI images is conducted
using two prominent deep learning architectures: VGG-19 and RESNET50. While experiments
indicate that RESNET50 generally outperforms VGG19 on various image classification
9
benchmarks, especially with larger and more complex datasets, VGG19 can still offer robust
performance, particularly with smaller datasets or limited computational resources. Hence, both
architectures are employed to potentially achieve superior results.

Concatenating VGG19 and RESNET50 involves merging the output features of both networks,
typically at a specific layer. This fusion enables the model to capture a broader range of features
from different depths of the networks, resulting in a more comprehensive representation for
classification tasks. By leveraging the complementary strengths of both architectures through
concatenation, the classifier can potentially benefit from enhanced feature extraction capabilities.

Model pruning is a technique aimed at reducing the size of deep learning models by eliminating
unnecessary parameters, thereby enhancing efficiency in terms of memory usage and
computational resources. Pruning involves identifying and removing unimportant connections or
weights in the model while preserving its performance. In the context of this research, applying
model pruning to the concatenated VGG19 and RESNET50 model can lead to a more streamlined
and efficient classifier for brain tumor identification. By trimming redundant parameters while
maintaining performance, the pruned model can optimize computational resources without
sacrificing accuracy.

In summary, the utilization of VGG19 and RESNET50 in tandem, coupled with model pruning
techniques, offers a promising approach for enhancing the efficiency and effectiveness of brain
tumor identification from MRI images. This strategy harnesses the strengths of both architectures
while mitigating resource constraints, ultimately contributing to improved diagnostic accuracy in
medical imaging applications.

Deep Learning Architectures that were regarded for use in this investigation and ultimately
utilized, which are then utilized for training, testing, and validating the classifier unit that
segments the input images in yes and no tumor classes.

The utilization of VGG19 and RESNET50 in tandem, coupled with model pruning techniques, offers a
promising approach for enhancing the efficiency and effectiveness of brain tumor identification from MRI
images. This strategy harnesses the strengths of both architectures while mitigating resource constraints,
ultimately contributing to improved diagnostic accuracy in medical imaging applications.

10
3.1 Overview of the architecture

Figure 3.1: System architecture diagram

The figure 3.1 depicts the flowchart shows how MRI images can be processed and analyzed to
create a model that can be used for medical diagnosis. The database that is considered for each
medical evaluation procedure in line with the issue that needs to be solved is significantly
responsible for the success of such an established system for making diagnoses. The majority of
the images used in the brain tumor identification competition were based on the widely used
benchmark pictures.

Figure 3.2: Sample test images

figure 3.2 shows the image dataset that was used for this investigation and can be viewed here.
The images connected to the skull portion are used for this evaluation. To to validate the proposed
Architecture, Pros cans Ltd. clinical quality MRIs are also taken into account. The large quantity
of test photographs for the tumor classifications of YES and NO was a result of this method.

11
3.2 Data pre-processing

Feature extraction is the primary stage in ML and DL techniques that helps extract important
information from a picture based on its structure and texture characteristics. The implemented
classifier units are trained, examined, and validated based on these attributes.
Calculating the key performance metrics is often how classifier performance is evaluated. The
performance of the classifier in this work will be evaluated using specific performance metrics.
In the initial evaluation, true-positive (TP), false-positive (FP), false-negative (FN) and true-
negative (TN) values are computed.

Data preprocessing is a crucial step in the deep learning pipeline, aiming to enhance the quality
of raw data to improve the performance and efficiency of deep learning models. This process
involves several essential techniques to clean, normalize, and engineer features before feeding
them into the model. One of the primary tasks in data preprocessing is data cleaning, which
addresses issues like missing values, outliers, and irrelevant features that could negatively impact
model performance. Techniques such as mean imputation, interpolation, and outlier detection
algorithms help to address these issues effectively.

1. Accuracy: The percentage of photos that were successfully classified.

2. Precision: The percentage of images properly identified as having AD out of all images
anticipated to have that condition.
3. Recall (sensitivity): The percentage of genuine photos of AD that are true positives.
4. F1 score: A precision and recall weighted average that strikes a balance between the two.
5. AUC-ROC, or the area under the receiver operating characteristic curve a metric for
evaluating the model's capacity to differentiate between photos of AD and those with normal
cognitive function.
Confusion matrices, a table that lists the total number of true positives, true negatives, false
positives, and false negatives, can also be used to assess the performance of the model. Confusion
matrices can help continue model development by shedding light on the precise types of errors
the model is committing.

12
In this part of the paper, the findings of the experiments and any subsequent commentary on those
results are included. MATLAB is used to conduct the research on a workstation equipped with
an Intel i7 processor, 12 gigabytes of random-access memory (RAM), and two gigabytes of
dedicated video RAM. During this stage of the procedure, an initial value is designated for each
DLA, as can be seen in the table below.

The epoch size is 15, the iteration size is 150, the number of iterations that occur in each epoch
is 50, the update frequency is once every 5 iterations, the learning error rate is 1e-5, and the
termination criterion is either the best validation or the maximum number of iterations. The epoch
size is 15, the iteration size is 150, the number of iterations that occur in each epoch is 50, and
the iteration size.

First, images taken from a dataset that contained MRI slices were rendered into a visual
representation. The photographs are all various sizes, so I have to crop the picture before I can
use it. On the other hand, the model's input can be a picture with dimensions of (224*224*3).
You will need to resize the image in order to accomplish this. Simply resizing an image without
paying attention to the results can result in severe image distortion. Therefore, crop the image
first, and then proceed to resize it. Problems with distortion will be reduced to a minimum as a
result. Finding the image contours in order to crop them is accomplished with the help of the
OpenCV library. The process of cropping an image consists of four steps.

Figure 3.3: Steps to visualize how cropping works

Figure 3.4: Images after being cropped

13
 Figure 3.5: Original image

Figure 3.6: Images after augmentation

The figure 3.3 shows steps to visualize how cropping works. The figure 3.4 shows images after
being cropped. The figure 3.5 shows Original Image. The figure 3.6 shows Images After
Augmentation. After an image has been cropped, it should be resized so that it does not suffer
from significant distortion or artifacts caused by resizing, and it should then be enlarged so that
more photographs can be saved.

3.3 Realistic Constraints

The availability of computational resources poses an actual constraint, as the training and
optimization of deep neural networks require significant computational power and memory.
Furthermore, the presence and excellence of datasets are crucial factors, as limitations frequently
arise from a lack of data or incomplete information, which can hinder the system's capacity to
generalize effectively. Furthermore, the limited amount of time available imposes restrictions on
the extent and thoroughness of experimentation, requiring careful selection and prioritization of
research directions. Moreover, the practical difficulties arise in domains where transparency and
accountability are of utmost importance, specifically, regarding the interpretability and
explainability of model decisions.

14
 CHAPTER 4

PROJECT IMPLEMENTATION AND ALGORITHMS

4.1 Software requirement specifications

Operating System: Windows 7 and above or Linux Platform

Special Tools: NumPy, Pandas, Seaborn, TensorFlow, Sklearn

4.2 Visual Geometry Group 19 (VGG19)

VGG is a multi-layered, standard deep convolutional neural network design. A convolutional

neural network with 19 layers is called VGG-19. With VGG-19 having 19 layers, the term "deep"
refers to the quantity of layers. Very tiny convolutional filters are used in the construction of the
VGG network.

Figure 4.1: VGG19 model architecture

The figure 4.1 shows the architecture of VGG19 model. The latest iteration of the VGG
models, VGG19, is strikingly similar to VGG16.The architecture remains unchanged and is still
built around five convolutional blocks. However, by including a convolutional layer in the final
three blocks, the network's depth has been substantially enhanced.

The Visual Geometry Group at the University of Oxford has devised the deep convolutional
15
neural network architecture known as VGG-19. It belongs to the VGG model family, which is
renowned for being easy to use and efficient when it comes to image identification tasks.

The number "19" in VGG19 stands for the total number of layers in the network, which consists
of three fully linked layers and sixteen convolutional layers. The architecture follows a simple
and uniform design principle, where the convolutional layers have a small receptive field of 3x3
pixels and are stacked on top of each other, followed by max-pooling layers to down sample the
feature maps.

The deep design and simplicity of VGG19, which enable it to extract intricate features from
images, have contributed to its rise in popularity. It has been extensively employed as a baseline
model for a number of computer vision applications, such as segmentation, object identification,
and image classification.

Despite its simplicity, VGG19 has a large number of parameters, which can make training and
inference computationally expensive, especially when dealing with large datasets. However, its
straightforward architecture makes it easy to understand and modify, and it serves as a solid
foundation for more complex network designs.

4.3 Residual Network 50 (RESNET50)

A deep convolutional neural network architecture is called RESNET50. The 50-layer RESNET50
architecture is made up of many residual blocks. A huge dataset of labelled photos is utilized to
train the RESNET50 architecture using supervised learning, which determines the weights of the
network.

Figure 4.2: RESNET50 model architecture

16
The figure 4.2 represents RESNET50 model architecture. The deep convolutional neural network
(CNN) architecture known as RESNET50 rose to fame due to its efficacious training of extremely
deep neural networks. Since its introduction in 2015 by Microsoft Research researchers, it has
gained popularity as a solution for a range of computer vision applications, such as segmentation,
object identification, and image classification.

The term "Residual Network," or "RESNET," alludes to the fundamental innovation included in
this architecture: the residual learning blocks. The vanishing gradient problem is a common
difficulty with traditional deep neural networks, which makes it challenging to train incredibly
deep networks efficiently. In order to overcome this problem, RESNET adds shortcuts or skip
links that let the gradient pass through the network more directly during training.

The identity path and the shortcut path are the two primary paths that make up the residual blocks,
which are the fundamental building blocks of RESNET. The shortcut path offers a direct
connection from the block's input to its output, whereas the identity path is the typical order of
convolutional layers followed by activation functions. RESNET efficiently enables the network
to learn residual functions by appending the input to the output through the shortcut path, which
facilitates the training of deep networks.

RESNET50 is the name given to a version of the RESNET architecture that has 50 layers total,
including fully connected, pooling, and convolutional layers. It consists of many residual blocks
at different levels, where more abstract features are captured in the deeper layers. Large-scale
datasets like ImageNet have been used to pretrained RESNET50, enabling it to learn rich
representations of pictures. These representations may then be adjusted or customized for
particular tasks using smaller datasets, such medical image analysis such as the detection of brain
tumors from MRI scans.

4.4 Fusion net model

Concatenation: The model takes two inputs of size 224x224x3, which are likely image patches.
These patches are fed into separate convolutional models, VGG19 and RESNET50. The outputs
of these two models are then concatenated into a single feature vector. This means that the model
combines the features extracted by both VGG19 and RESNET50.

Both the VGG19 and RESNET50 sub models in the architecture use multiple convolutional
layers. Convolutional layers are essential for extracting features from images. They apply a filter
17
to the input data, producing a feature map that highlights specific patterns in the image.

Figure 4.3: Architecture of FUSIONNET model

This figure 4.3 represents the fusion net model architecture consists of concatenation of VGG19
and RESNET50 Sub models: The VGG19 sub model follows a typical convolutional neural
network architecture with multiple convolutional layers, followed by pooling layers. The
RESNET50 sub model uses residual connections, which skip some layers in the network and
allow the model to learn more complex features.

Multiple Convolutional Layers: Both the VGG19 and RESNET50 sub models in the architecture
use multiple convolutional layers. Convolutional layers are essential for extracting features from
images. They apply a filter to the input data, producing a feature map that highlights specific
patterns in the image.

Pooling Layers: In order to down sample the feature maps generated by the convolutional layers,
the architecture additionally incorporates pooling layers. By doing this, the data's dimensionality
is decreased and overfitting is less likely to occur.

A sigmoid activation layer is a type of activation function commonly used in neural networks,
especially in the context of binary classification problems. The sigmoid function, also known as
the logistic function, maps any real-valued number to the range (0, 1).

In the context of a neural network, the sigmoid activation function is applied to the output of a
18
neuron, transforming it into a value between 0 and 1. This transformed value can be interpreted
as a probability. If the output is close to 0, it indicates a low probability, and if it's close to 1, it
indicates a high probability.

The sigmoid activation function is particularly useful at the output layer of a binary classification
neural network where the goal is to predict probabilities of belonging to one of the two classes.

functions like ReLU (Rectified Linear Unit) due to the vanishing gradient problem, which can
impede training in deep networks.

The identity path and the shortcut path are the two primary paths that make up the residual blocks,
which are the fundamental building blocks of RESNET. The shortcut path offers a direct
connection from the block's input to its output, whereas the identity path is the typical order of
convolutional layers followed by activation functions. RESNET efficiently enables the network
to learn residual functions by appending the input to the output through the shortcut path, which
facilitates the training of deep networks.

19
 CHAPTER 5

FRAMEWORK

5.1 Feature extraction of VGG19

Deep convolutional neural network (CNN) architecture VGG19 has shown to be very successful
in a variety of computer vision applications, such as feature extraction, object detection, and
image categorization. The entire number of layers, including convolutional and fully linked
layers, is indicated by the "19" in VGG19.

By running an image through the network and recording the activations of specific layers as
features, features may be extracted using VGG19. At various levels of abstraction, these
activations represent the learnt representations of the input image. The max-pooling layers in
VGG19 come after a series of convolutional layers that gradually decrease the input's spatial
dimensions while increasing its depth or number of channels.

During feature extraction, the input image is passed through the network until a certain layer
is reached. The precise job at hand and the required level of abstraction in the features determine
which layer is best for feature extraction. Typically, deeper layers in the network capture higher-
level semantic information, while earlier layers capture low-level features such as edges and
textures.

Once the desired layer is reached, the activations of the neurons in that layer are extracted and
used as features. These activations form a high-dimensional representation of the input image,
where each dimension corresponds to the output of a particular neuron in the layer. This
representation can then be used as input to downstream machine learning models for tasks such
as classification, clustering, or feature visualization.

One advantage of using pre-trained models like VGG19 for feature extraction is that they have
been trained on large-scale datasets such as ImageNet, which contain a diverse range of images.
As a result, the features extracted by these models are often generic and transferable across
different tasks and domains.

20
5.2 Feature extraction of RESNET50

RESNET50 is a deep convolutional neural network architecture that has gained significant
popularity for tasks like image classification and feature extraction. In theory, understanding how
RESNET50 performs feature extraction requires delving into its architectural design and the
underlying principles of convolutional neural networks (CNNs).

Multiple layers of convolutional, pooling, and fully linked layers make up the foundation of
RESNET50. The use of residual learning blocks in RESNET50 distinguishes it from previous
CNN architectures and allows for the training of much deeper networks without experiencing
vanishing gradient issues. By introducing skip links, these leftover blocks enable more direct
information flow throughout the network, which makes it easier to learn more complicated
properties.

During the feature extraction process, an input image is passed through the network, layer by
layer. At each layer, convolutional filters are applied to the input, extracting various features such
as edges, textures, and patterns. These features become increasingly abstract and high-level as
the input progresses deeper into the network. In RESNET50, the initial layers typically capture
low-level features like edges and colors, while deeper layers learn more complex features like
object parts and textures.

A sigmoid activation function, which outputs the probabilities of the input image belonging to
each class in a classification job, usually comes after fully connected layers in the last layers of
RESNET50. However, during feature extraction, these final classification layers are often
removed, and the output is obtained from an intermediate layer of the network. This intermediate
feature representation can be used for various tasks such as feature visualization, transfer
learning, or feeding into another machine learning model for downstream tasks.

Overall, RESNET50's feature extraction capabilities stem from its deep architecture, which
enables it to learn hierarchical representations of input images. By leveraging the rich set of
features extracted by RESNET50, researchers and practitioners can tackle a wide range of
computer vision tasks with high accuracy and efficiency.

21
5.3 Feature extraction of FUSIONNET model

Utilizing the combined strengths of two potent convolutional neural networks (CNNs), feature
extraction utilizing a concatenation of VGG19 and RESNET50 includes extracting high-level
features from images. Well-known CNN designs VGG19 and RESNET50 each have unique
advantages in capturing various facets of visual data.

A variation of the VGG architecture called VGG19 is distinguished by its uniformity and
simplicity. It is made up of 19 layers, the first of which are max-pooling layers and small 3x3
convolutional filters. Its massive stack of convolutional layers allows it to capture finely detailed
textures and patterns found in photos. Feature maps representing the hierarchical visual properties
that the network learns from basic edges and textures to more intricate structures can be obtained
by running a picture through VGG19.

Conversely, RESNET50 tackles the deterioration issue with very deep neural networks by
introducing the idea of residual learning.
In order to accomplish this, RESNET50 makes use of skip connections, which let the network
discover residual mappings rather than fitting the mapping straight away. This architecture is
especially useful for tasks like object detection and localization since it is good at capturing high-
level semantic characteristics and can learn complex spatial relationships within images.

To perform feature extraction using the concatenation of VGG19 and RESNET50, we first pass
an image through both networks individually. This results in two sets of feature maps, each
capturing different aspects of the image content.

Next, we concatenate the feature maps obtained from both networks along the channel dimension.
This concatenation merges the complementary information captured by VGG19 and RESNET50,
resulting in a combined feature representation that encompasses both detailed texture information
and high-level semantic features. By doing so, we create a more comprehensive and informative
feature representation that can potentially enhance the performance of downstream tasks such as
image classification, object detection, or image retrieval.

Finally, the concatenated feature representation can be fed into subsequent layers of a neural
network for further processing or used directly as input to a machine learning model for various
computer vision tasks. Through this approach, we leverage the strengths of both VGG19 and
RESNET50 to extract rich and discriminative features from images, ultimately improving the
performance of visual recognition systems.
22
 CHAPTER 6

RESULTS AND DISCUSSIONS

6.1 Evaluation

In this part of the paper, the findings of the experiments and any subsequent commentary on those
results are included. MATLAB is used to conduct the research on a workstation equipped with
an Intel i7 processor, 12 gigabytes of random-access memory (RAM), and two gigabytes of
dedicated video RAM. During this stage of the procedure, an initial value is designated for each
DLA, as can be seen in the table below.

The epoch size is 15, the iteration size is 150, the number of iterations that occur in each epoch
is 50, the update frequency is once every 5 iterations, the learning error rate is 1e-5, and the
termination criterion is either the best validation or the maximum number of iterations. The epoch
size is 15, the iteration size is 150, the number of iterations that occur in each epoch is 50, and
the iteration size. After an image has been cropped, it should be resized so that it does not suffer
from significant distortion or artifacts caused by resizing, and it should then be enlarged so that
more photographs can be saved.

First, images taken from a dataset that contained MRI slices were rendered into a visual
representation. The photographs are all various sizes, so I have to crop the picture before I can
use it. On the other hand, the model's input can be a picture with dimensions of (224*224*3).
You will need to resize the image in order to accomplish this. Simply resizing an image without
paying attention to the results can result in severe image distortion. Therefore, crop the image
first, and then proceed to resize it. Problems with distortion will be reduced to a minimum as a
result. Finding the image contours in order to crop them is accomplished with the help of the
OpenCV library. The process of cropping an image consists of four steps.

After an image has been cropped, it should be resized so that it does not suffer from significant
distortion or artifacts caused by resizing, and it should then be enlarged so that more photographs
can be saved.

23
6.2 Hyperparameters

Hyperparameters in Machine learning are those parameters that are explicitly defined by the user
to control the learning process. These hyperparameters are used to improve the learning of the
model, and their values are set before starting the learning process of the model. Below table 6.1
refers to the hyperparameters of the model used.

Table 6.1. Hyperparameters.

Hyper Parameters Values

Total Number of epochs 15

Batch size 50

Loss Function Binary Cross entropy

Learning Rate 1e^−4

Optimizer Adam Optimizer

6.3 Results discussion

After an image has been cropped, it should be resized so that it does not suffer from significant
distortion or artifacts caused by resizing, and it should then be enlarged so, that more photographs
can be saved. The figure 6.1 describes the confusion matrix of VGG19 model. By using this
diagram, we can calculate accuracy, precision and few other metrics. The figure 6.2 describes the
confusion matrix of RESNET50 model, using this diagram we can calculate accuracy, precision
and other metrics. The figure 6.3 depicts the confusion matrix of combined model which is the
combination of VGG19 and RESNET50 models. The figure 6.4 represents the confusion matrix
of the pruned model after applying model pruning technique on the concatenated model. The
figure 6.5 shows Train and test accuracy plot of VGG19 model which is plotted against epochs
and accuracy on x and y axis respectively. The figure 6.6 shows Train and test accuracy plot of
RESNET50 model which is plotted for epochs and accuracy on x and y axis respectively. The
figure 6.7 shows Train and test accuracy plot of concatenated model against epochs and accuracy
on x and y axis respectively. The figure 6.8 depicts Train and test
24
accuracy plot of the pruned model against epoch and accuracy on x and y axis respectively.

Figure 6.1: Confusion matrix of VGG19

Figure 6.2: Confusion matrix of RESNET50

Figure 6.3: Confusion matrix of combine model

25
 Figure 6.4: Confusion matrix of pruned model

Figure 6.5: Train and test accuracy plot of VGG19 model

Figure 6.6: Train and test accuracy plot of RESNET50

26
 Figure 6.7: Train and test accuracy plot of concatenated model

Figure 6.8: Train and test accuracy plot of VGG19

Furthermore, we have used few performance metrics on the datasets. we got 93.83 as an accuracy
score, 94.30 as a precision score, 93.83 as a Recall score, 93.81 as a F1 score for VGG19.
Whereas we got 93 as a accuracy score, 93.37 as a precision score, 93 as a Recall score, 92.99 as
a F1 score for RESNET50. here as we got 95.5 as an accuracy score, 95.54 as a precision score,
95.50 as a Recall score, 95.55as a F1 score for the concatenated model and here as we got 97.16
as an accuracy score, 97.20 as a precision score, 97.16 as a Recall score, 97.16 as a F1 score for
the pruned model.

27
 Table 6.2. Performance parameters of models

This table 6.2 represents the performance parameters of the model. The Model parameter is a
configuration variable that is internal to the model and whose value can be estimated from the
given data. They are required by the model when making predictions. The Relative classification
error in accuracy between the pruned model and VGG19 model is 0.035. The classification error
for the pruned model and RESNET50 is 0.044 and for the concatenated model is 0.017.

The future scope of the proposed research includes:

Improving the handcrafted quality of the feature vector by taking into account extra texture and
form characteristics. Making adjustments to the fully-connected and drop-out layers in order to
achieve more accuracy in the categorization. Making enhancements to the method of feature
concatenation in order to achieve superior results. Putting into practice the suggested DLAs in
order to categorize the gliomasas either low or high grade.

28
 CHAPTER 7

CONCLUSION AND FUTURE SCOPE

7.1 Conclusion

In conclusion, the use of deep learning models such as VGG19 and RESNET50 has shown
promising results in the detection of brain tumors. In addition to leveraging individual deep
learning models like VGG19 and RESNET50 for brain tumor detection, there's also an emerging
approach that combines the strengths of multiple models through concatenation. By
concatenating the outputs of these models and passing them through a sigmoid layer, we can
harness the complementary features and representations learned by each network.

This fusion technique aims to enhance the overall performance of the detection system. VGG19,
with its simpler architecture, may excel in capturing certain low-level features, while
RESNET50, with its deeper structure and skip connections, may excel in capturing more
complex and abstract features. By combining their outputs, we create a more robust and
comprehensive representation of the input data, potentially improving the accuracy of tumor
detection.

However, it's important to carefully evaluate the effectiveness of this concatenation approach,
as it introduces additional complexity and computational overhead. Additionally, fine-tuning
may be necessary to optimize the fusion process and ensure that the concatenated model
effectively exploits the strengths of each component network.

Further research and experimentation are required to assess the performance of concatenated
models in comparison to individual models and other fusion strategies. Additionally, validation
studies across diverse datasets and populations are crucial to ensuring the generalizability and
robustness of the concatenated approach in real-world clinical settings.

7.2 Future Scope

The future scope for brain tumor detection using deep learning models like VGG19 and

RESNET50 is vast. Here are a few potential areas for further research:

29
 Integration with other medical imaging techniques: The use of deep learning models
can be integrated with other medical imaging techniques such as PET scans and CT
scans to improve the accuracy of brain tumor detection.
Multimodal imaging: The combination of different imaging modalities, such as MRI
and diffusion-weighted imaging, can provide a more comprehensive view of brain
tumor detection. Deep learning models can be used to analyze these multimodal
images to improve the accuracy of diagnosis.
Personalized treatment plans: The use of deep learning models can be extended beyond
diagnosis to assist in the development of personalized treatment plans for patients with
brain tumors. These models can be used to predict treatment outcomes and help doctors
make informed decisions.
Real-time tumor detection: Real-time tumor detection using deep learning models can
allow for quicker and more accurate diagnosis of brain tumors. This can be especially
beneficial for emergency situations where time is critical.
Integration with Electronic Health Records (EHR): The integration of deep learning
models with EHR systems can help doctors make informed decisions based on the
patient's medical history, previous diagnoses, and treatment outcomes.
Overall, the future scope for brain tumor detection using deep learning models is extensiveand
has the potential to improve patient outcomes by providing earlier and more accurate diagnosis,
personalized treatment plans, and real-time monitoring.

30
 REFERENCES
[1] Rajinikanth, V.; Joseph Raj, A.N.; Thanaraj, K.P.; Naik, G.R. "A Customized VGG19
Network with Concatenation of Deep and Handcrafted Features for Brain Tumor
Detection". Applied Sciences 2020, vol. 82, no.12, p. 104570, 2022, doi:
10.1016/j.imavis.2022.104570.

[2] B. Kalyan and P. Chandrasekhar Reddy, "Brain Tumor Detection Using Convolutional
Neural Networks," 2022 10th International Conference on Reliability, Infocom
Technologies and Optimization (Trends and Future Directions) (ICRITO), vol. 18, p.
104570, 2022, doi: 10.1016/j.imavis.2022.104570, Noida, India, 2022

[3] Khan, Md & Rahman, Anichur & Debnath, Tanoy & Razaul, Md & Nasir, Mostofa & S.
Band, Shahab & Mosavi, Amir & Dehzangi, Iman (Abdollah)."Accurate Brain Tumor
Detection Using Deep Convolutional Neural Network". Computational and Structural
Biotechnology Journal, vol. 76 2022

[4] Bhandary A., Prabhu G.A., Rajinikanth V., Thanaraj K.P., Satapathy S.C., Robbins D.E.,
Shasky C. , Raja N.S.M. "Deep-learning framework to detect lung abnormality – A study
with chest X-Ray and lung CT scan images”, vol. 512 (2020)

[5] Fernandes, S.L., Tanik, U.J., Rajinikanth."A reliable framework for accurate brain image
examination and treatment planning based on early diagnosis support for clinicians".
Neural Compute & Applications 32, vol. 1138, 15897–15908 (2020)

[6] Krishnan T. Palani, Balasubramanian Parvathavarthini and Krishnan Chitra,

"Segmentation of Brain Regions by Integrating Meta Heuristic Multilevel Threshold with
Markov Random Field", Current Medical Imaging 2016; vol. 12(1)

[7] Amin J., Sharif M., Yasmin M., Fernandes S.L.Big "data analysis for braintumor
detection: Deep convolutional neural networks”, Future Generation Computer
Systems(2018) vol. 60

31
[8] Thanaraj, P., Parvathavarthini, B."Multichannel interictal spike activity detection using
time–frequency entropy measure". Australas Phys Eng. Sci Med vol. 40, 413–425
(2017)

[9] Talo M., Baloglu U.B., Yildirim O., Rajendra Acharya U. "Application of deep transfer
learning for automated brain abnormality classification using MR images Cognitive
Systems Research", vol. 54, pp. 176-188. (2019)

[10] Sajid, S., Hussain, S. & Sarwar, "A Brain Tumor Detection and Segmentation in MR
Images Using Deep Learning". Arab J Sci Eng vol. 44, 9249–9261 (2019)

[11] Wang, Y., Liu, Z., Liu, X., & Du, X., "Classification of Brain Tumor MRI Images Using
RESNET50. Journal of Medical Imaging and Health Informatics, vol. 11(9), 2284-
2290(2021)

[12] Aslantas, V., & Tanrikulu, O. C., "Automatic Brain Tumor Classification with
VGG19". Journal of Medical Imaging and Radiation Sciences, vol 51(2), 262-
268(2020)

[13] Zhang, L., Wang, J., Xiong, Y., & Zhang, J., An "Improved Brain Tumor Classification
Method Based on Deep Learning". Journal of Healthcare Engineering, 2020, vol.1011.
(2020)

[14] Zhu,"Brain Tumor Detection and Classification Using RESNET50 with Feature E
xtraction"(2020), vol. 30

[15] Sharif M.I., Li J.P., Khan M.A., Saleem M.A. "Active deep neural network features
selection for segmentation and recognition of brain tumors using MRI images,
PatternRecognition Letters", v o l . 129, pp. 181-189,(2020)c

[16] Nadeem, M.W.; Ghamdi, M.A.A.; Hussain, M.; Khan, M.A.; Khan, K.M.; Almotiri,
S.H.; Butt, S.A "Brain Tumor Analysis Empowered with Deep Learning": A Review,
Taxonomy, and Future Challenges. Brain Sciences 2020

32
[17] Amin J., Sharif M., Yasmin M., Fernandes S.L. Big "data analysis for brain tumor
detection: Deep convolutional neural networks", Future Generation Computer Systems
(2018) vol. 1330

[18] Talo M., Baloglu U.B., Yildirim O., Rajendra Acharya U. "Application of deep transfer
learning for automated brain abnormality classification using MR images Cognitive
Systems Research", vol. 54, pp. 176-188. (2019)
[19] Sajid, S., Hussain, S. & Sarwar, "A Brain Tumor Detection and Segmentation in MR
Images Using Deep Learning". Arab J Sci Eng vol. 44, 9249–9261 (2019)

[20] Aslantas, V., & Tanrikulu, O. C. , "Automatic Brain Tumor Classification with
VGG19". Journal of Medical Imaging and Radiation Sciences, vol. 51(2), 262-
268(2020)

33
Final Report plaggg
ORIGINALITY REPORT

5 %
SIMILARITY INDEX
2%
INTERNET SOURCES
1%
PUBLICATIONS
3%
STUDENT PAPERS

PRIMARY SOURCES

1
Submitted to University of East London
Student Paper 1%
2 Submitted to Liverpool John Moores
University
<1%
Student Paper

<1%
Lemana Spahić, Zerina Mašetić, Almir
Badnjević, Asim Kurjak, Lejla Gurbeta Pokvić.
"Chapter 65 Artificial Intelligence-Based
Ultrasound Imaging Classification for Infant
Neurological Impairment Disorders: A
Review", Springer Science and Business
Media LLC, 2024
Publication

<1%
4
ceur-ws.org
Internet Source

<1%
5
d.docksci.com
Internet Source

<1%
6
fastercapital.com
Internet Source

Submitted to Brown University

34
 <1%
Student Paper

8
dspace.daffodilvarsity.edu.bd:8080
Internet Source <1%

9
Submitted to University Tun Hussein Onn
Malaysia
<1%
Student Paper

Submitted to University of Hertfordshire

Student Paper <1%
idr.nitkkr.ac.in:8080
Internet Source <1%
www.researchsquare.com
Internet Source <1%
Submitted to Manchester Metropolitan
University
<1%
Student Paper

Submitted to Istanbul Aydin University

Student Paper <1%
Submitted to City University
Student Paper <1%
Hanxiang Hao, David Guera, Janos Horvath,
Amy R. Reibman, Edward J. Delp. "Robustness
<1%
Analysis of Face Obscuration", 2020 15th IEEE
International Conference on Automatic Face
and Gesture Recognition (FG 2020), 2020
Publication

35
 <1%
Submitted to University of Birmingham
Student Paper

<1%
pt.scribd.com
Internet Source

<1%
www.frontiersin.org
19 Internet Source

<1%
Submitted to University of Wales Institute,
Cardiff
Student Paper

Exclude quotes Off Exclude matches < 10 words

Exclude bibliography Off

36
 Unveiling the future of healthcare:
Revolutionizing Brain Tumor Detection with
Computer Vision and AI

Jaswanth Koripalli P R S V Akhil Pavan Kumar Dr K Kalimuthui *

Department of Department of Department of Department of
Electronics and Electronics and Electronics and Electronics and
Communication Communication Communication Communication
Engineering, Engineering, Engineering, Engineering,
Faculty of Engineering Faculty of Engineering College of Engineering Faculty of Engineering
and Technology, and Technology, and Technology, and Technology,
SRM Institute of Science SRM Institute of Science SRM Institute of Science SRM Institute of Science
and Technology, and Technology, and Technology, and Technology,
SRM Nagar, SRM Nagar, SRM Nagar, SRM Nagar,
Kattankulathur 603203, Kattankulathur 603203, Kattankulathur603203, Kattankulathur-603203,
Tamil Nadu, India. Tamil Nadu, India. Tamil Nadu, India. Tamil Nadu, India.
[email protected] [email protected] [email protected] [email protected]

Abstract— In a groundbreaking development in medical I. INTRODUCTION

science, scientists are leveraging the capabilities of
Computer Vision to revolutionize the detection of brain Brain tumors, the second leading cause of global
tumors. This pioneering study explores the fusion of mortality, can be benign or malignant. Benign tumors
technology and healthcare, unveiling the intricacies of pose minimal health risks due to slower growth and
the human brain through machine learning. This lack of infiltration, while malignant tumors can cause
research uses a pre-trained VGG-19 model, RESNET50 significant dysfunction. Gliomas, meningiomas, and
model, and a concatenated model, incorporating model pituitary tumors are common types, with gliomas
pruning techniques to enhance efficiency. Drawing on
being malignant, meningiomas benign, and pituitary
the vast knowledge of ImageNet, Keras API, and
tumors generally benign [1] . Brain tumors, harmful
TensorFlow play crucial roles in seamlessly integrating
cutting-edge algorithms with extensive datasets. As
intracranial masses, pose challenges in detection due
pixels come together to convey a wealth of information, to the brain's complex structure. Medical imaging,
these tools unravel the mystery of brain tumors, particularly MRI, aids diagnosis, and image
providing a glimpse into the future of precise diagnostics. processing techniques enhance tumor detection
Through the collaboration of artificial intelligence and through features like contrast adjustment,
medical expertise, a new era in healthcare emerges, thresholding, and segmentation [2] . The central
where early detection is not just a possibility but a nervous system, comprising the brain and spinal cord,
tangible reality, reshaping patient care and instilling
facilitates sensory information transmission and
hope for many.
actions throughout the body. The brain's anatomy
includes the brain stem, cerebrum, and cerebellum,
each responsible for specific functions such as motor
Keywords— brain tumor; brain MRI images; VGG-19, control, sensory processing, and vital activities. [3] .
Deep Learning, ResNet50, Model Pruning The article introduces an automatic method for brain
tumor detection and classification, incorporating
preprocessing, feature extraction, and geometrical
family Support Vector Machine (SVM) for improved timely and accurate diagnosis [7] . The study conducts
segmentation and classification, evaluated on various a bibliometric analysis on machine learning-based
datasets. [4] . The paper employs transfer learning automated brain tumor detection, emphasizing its
(TL) on MRI images for brain tumor diagnosis, emergence as a significant research area for improved
demonstrating high classification performance and diagnosis and survival chances. The research assesses
accuracy across 5000 T2-weighted contrast-enhanced trends, advancements, and future topics over the past
images, addressing challenges posed by the complex 5 years to understand the state of the art in brain tumor
brain structure and similar lesion areas. The results prediction [8] . The study presents a deep learning-
highlight the effectiveness of TL models in identifying based method for automated metabolic tumor volume
brain tumors in medical images. [5] . The paper (MTV) segmentation using amino acid PET scans,
reviews deep learning applications in brain tumor achieving reliable and robust results in brain tumor
analysis, covering segmentation, classification, and patients. The automated segmentation demonstrated
survival time prediction, discussing methods, clinical value for response assessment, showing
frameworks, and challenges. The survey highlights the agreement with physician assessments and significant
increasing dominance of this topic in conferences and impact on disease-free and overall survival outcomes
journals since 2015. [6] . The research paper proposes [9] . The research addresses the challenges of brain
an innovative system for brain tumor detection, tumor segmentation using deep learning networks,
employing preprocessing, segmentation, feature particularly focusing on the ResNet model to enhance
extraction, and an optimized convolutional neural computational efficiency and training processes. The
network (CNN) with various optimization algorithms. article outlines the organization with sections on
Achieving a detection accuracy of 98.9%, the system existing tumor segmentation work, materials,
demonstrates superiority over contemporary methods, results, discussion, and conclusions for
optimization techniques, emphasizing its potential for comprehensive coverage [10] .

images, and after a number of convolutional and

pooling layers, moves on to fully linked layers for
II. LITERATURE SURVEY classification [13]. A novel deep learning framework
for the detection of lung abnormalities utilising a mix
According to this study, a customised VGG19 network
of chest X-ray and lung CT scan images is presented
that uses both hand-crafted and deep MRI scan
in the study by Bhandary A. While deep learning
features can be used to identify brain tumours. The
techniques have been used to detect lung problems in
method generated accuracy of 96.08% and an AUC of
a number of studies, the proposed framework stands
0.99 on a dataset of 306 MRI images, demonstrating
out for its unique approach and the demonstrated
that the VGG19 network may perform better for this
efficacy in detecting various types of lung
job when deep and handmade features are mixed [11].
abnormalities[14]. The study makes a foundation with
The authors' proposed convolutional neural network
the purpose of providing doctors with reliable brain
(CNN)-based deep learning method for detecting
picture analysis and therapy planning based on early
brain tumours uses this technology. The network
diagnosis. Utilizing a convolutional neural network
receives input from MRI images, and after a number
(CNN), the framework for classification, correctly
of convolutional and pooling layers, moves on to fully
identified brain tumours using a set of brain Magnetic
linked layers for classification.The proposed method
resonance imaging scans with an accuracy of 97.14
was evaluated on a dataset of 150 MRI scans, and it
percent.The paper was published in Neural
generated accuracy and sensitivity results of 98.7%
Computing and Applications in 2020 [15]. This study
and 97.1%, respectively. The results indicate that the
evaluated their suggested method using a dataset of 20
suggested CNN-based strategy has promise as a brain
MRI images, indicating that the proposed method can
cancer identification tool [12]. The authors advocate
be an effective tool for segmenting brain regions in
using convolutional neural networks (CNNs), a deep
medical images. The paper was published in Current
learning approach, to precisely identify brain
Medical Imaging in 2016 [16]. The CNN model in this
tumours. The network receives input from MRI
work, which was trained and tested using a dataset of
brain MRI images, classified brain tumours.The study detection problem have consisted of modifying the
demonstrates that deep learning can be an effective validated DLAs that can already be found in the

tool for analyzing big data in literature. Before selecting

medical imaging and improving the accuracy of brain and putting an architecture into action, it is essential to
tumor detection[17]. The study's proposed method have a solid understanding of its structure, as well as
was tested using a dataset of EEG signals from the degree of effort involved in putting it into action,
epileptic patients, and it had an average detection the initial tweaking, and the validation procedures.
accuracy of 91.6%. The research demonstrates that
the suggested technique can be a useful instrument for
identifying interictal spike activity. in EEG signals,
In this particular research project, the identification of
which can aid in the diagnosis and treatment of
brain tumours from the studied MRI material is carried
epilepsy. The paper was published in Australasian out with the use of the VGG-19 and ResNet50 Deep
Physical & Engineering Sciences in Medicine in 2017 Learning Architectures. Because the experiments
[18]. This study uses a dataset of MRI images to train conducted for this investigation have shown that
and test their deep transfer learning model, which ResNet50 outperforms VGG19 on several image
achieved an accuracy of 97.9% for brain abnormality classification benchmarks, especially on larger and
more complex datasets. However, VGG19 can still
classification. The study demonstrates that transfer
perform well and may be a better choice for smaller
learning can be an effective tool for automated datasets or when computational resources are limited.
medical image analysis, especially when the amount of So we have used them achieve greater results.
labeled data is limited. The paper was published in
Cognitive Systems Research in 2019[19]. This study Figure 1 depicts the VGG19 that was regarded for use
uses a dataset of MRI images to train and test their in this investigation and ultimately utilised.Principal
deep learning model, which got an accuracy of 99.6% component analysis (PCA) is a deep and meticulous
feature used to sort and serially integrate images. It is
for brain tumour detection and 97.7% for tumor
then used for training, testing, and validating the
segmentation.The work shows that deep learning can classification component that divides the input images
be a useful technique for precise and automated brain into yes and no cancer classes.
tumour detection and segmentation in medical
pictures, which can help with brain tumour diagnosis
and therapy [20].
The success of such an established system for
III. PROPOSED METHODOLOGY making diagnoses is largely based on the database
that is considered for each medical evaluation
Numerous different DLAs, both standard and procedure in accordance with the problem that
customised, are proposed throughout the research must be resolved. The widely recognised
literature as potential methods for locating anomalies benchmark photos served as the basis for the
in medical imaging. Constructing, training, testing, majority of the images that were utilised in the
and validating the architecture in order to meet the brain tumour detection competition. The image
requirements of a particular test are some of the many dataset that was utilised for this study is displayed
complex steps involved in the creation of a new DLA in Figure 2, which can be found here.
from the initial concept. As a consequence of this, the The glioma images connected to the skull portion
majority of previous efforts to address a disease are used for this evaluation. In order to validate
the proposed Deep Learning Architecture,
Proscans Ltd. clinical-grade Magnetic resonance object detection, segmentation, and even natural
imaging are also taken into account.This strategy language processing, has improved as a result.
contributed to the vast number of test photos for
the tumour classifications YES and NO. Residual Network 50 (ResNet50)

Microsoft Research presented ResNet50, a deep

convolutional neural network design, in 2015.
The architecture has 50 layers, including 1 fully
linked layer and 49 convolutional layers.

The introduction of residual connections, also

referred to as skip connections, which allow data
to escape specific network levels, is the main
innovation of ResNet50. As a result, the network
Figure 2. Sample test images can be trained more successfully and the
vanishing gradient problem that might arise in
Feature extraction is the primary stage in ML and very deep networks is lessened.
DL techniques that helps extract important
information from a picture based on its structure In our research, we implemented a novel
and texture characteristics.The implemented approach by combining the strengths of VGG19
classifier units are trained, examined, and and ResNet50 architectures into a concatenated
validated based on these traits. model, aiming to enhance the efficiency of brain
tumor detection. This concatenated model
Calculating the key performance metrics is often leverages the unique features of both
how classifier performance is evaluated. The architectures, capitalizing on VGG19's
performance of the classifier in this work will be effectiveness on smaller datasets and ResNet50's
evaluated using specific performance metrics. superior performance on larger and more
complex datasets. To optimize the concatenated
Visual Geometry Group 19 (VGG19) model further, we applied model pruning
techniques, a process that involves selectively
One of the first deep convolutional neural removing redundant or less impactful parameters,
networks, VGG19, was able to perform at the leading to a more streamlined and resource-
cutting edge on the difficult ImageNet dataset, efficient neural network.
which consists of more than a million photos
divided into 1000 different classes. The The introduction of the concatenated model and
popularity of deep learning for computer vision model pruning techniques represents a strategic
problems was aided by the performance of effort to boost the overall accuracy and
VGG19 on this dataset. effectiveness of brain tumor detection. By
synergizing the complementary attributes of
In contrast to larger filters with a longer stride, the VGG19 and ResNet50, we aim to achieve a
architecture of VGG19 is built on the idea of comprehensive and robust framework for
using smaller filters (3x3) with a stride of 1. This identifying and classifying brain tumors. This
improves performance by enabling the network to innovative approach not only harnesses the power
record more precise characteristics at each layer. of state-of-the-art deep learning architectures but
also emphasizes the importance of model
VGG19 has been utilised for transfer learning, optimization techniques to ensure optimal
where further computer vision tasks are launched performance in medical image analysis,
using the network's pre-trained weights, in ultimately advancing the field of diagnostic
addition to its success on image classification precision in neuroimaging.
tasks. Performance on a variety of tasks, such as
 Figure 3. Steps to Visualize how cropping works

Fig 3. describes the visualization of steps involved in

cropping of images
IV. PERFORMANCE EVALUATION

In this part of the paper, the findings of the

experiments and any subsequent commentary on those
results are included. Matlab is used to conduct the
research on a workstation equipped with an Intel i7
processor, 12 gigabytes of random access memory
(RAM), and two gigabytes of dedicated video RAM.
During this stage of the procedure, an initial value is
designated for each DLA, as can be seen in the table Figure 4. Images After being cropped
below.
Fig 4. describes the visualization of images after being
The epoch size is 15, the iteration size is 150, the cropped
number of iterations that occur in each epoch is 50, the
update frequency is once every 5 iterations, the After an image has been cropped, it should be resized
learning error rate is 1e-5, and the termination criterion so that it does not suffer from significant distortion or
is either the best validation or the maximum number artifacts caused by resizing, and it should then be
of iterations. The epoch size is 15, the iteration size is enlarged so that more photographs can be saved.
150, the number of iterations that occur in each epoch
is 50, and the iteration size

First, images taken from a dataset that contained MRI

slices were rendered into a visual representation. The
photographs are all various sizes, so I have to crop the
picture before I can use it. On the other hand, the
model's input can be a picture with dimensions of
(224*224*3). You will need to resize the image in
order to accomplish this. Simply resizing an image
without paying attention to the results can result in
severe image distortion. Therefore, crop the image
first, and then proceed to resize it. Problems with
distortion will be reduced to a minimum as a result.
Finding the image contours in order to crop them is
accomplished with the help of the OpenCV library.
The process of cropping an image consists of four
Figure 5. Images After Augmentation
steps.
Fig 5. describes the visualization of images from
augmentation of original image, In that augmentation
describes the creation of duplication of images from
original image.

The implementation of a concatenated model,

combining VGG19 and ResNet50 architectures, was
introduced to enhance efficiency. Model pruning
techniques were applied to optimize the concatenated
model. The code snippet demonstrates the
concatenated model's architecture, incorporating
additional layers, dropout, and batch normalization for
improved performance. The model summary outlines
the frozen layers and the compiled model metrics, Figure 7. Confusion matrix of RESNET50
including binary crossentropy loss, Adam optimizer,
and accuracy and AUC metrics. This concatenated
model, enriched with model pruning, represents an
innovative step toward optimizing brain tumor
detection through the synergistic power of VGG19 and
ResNet50 architectures.

V. RESULTS DISCUSSION

These datasets were gathered from Kaggle and

contain a total of 3,070 images, of which 2000 were
used for the training phase and 70 were used for the
testing operation. Accuracy of 94% is achieved while
using the selected VGG19 as the benchmark for
performance.
Figure 8. Confusion matrix of combine model

Figure 6. Confusion matrix of VGG19

Fig 6. describes the confusion matrix of the given

Figure 9. Confusion matrix of pruned model
dataset. By using this diagram we can calculate
accuracy,precision and few other metrics. Furthermore we have used few performance metrics
on the datasets. we got 93.83 as a accuracy score,
94.30 as a precision score, 93.83 as a Recall score, IV. Putting into practice the suggested VGG19
93.81 as a F1 score for VGG19. Whereas we got 93 as DLA in order to categorize the gliomas as
a accuracy score, 93.37 as a precision score, 93 as a either low or high grade.
V. The development of a neural network model
Recall score, 92.99 as a F1 score for ResNet50. hereas
for the propagation of ependymal tumors.
we got 95.5 as a accuracy score, 95.54 as a precision
score, 95.50 as a Recall score, 95.55as a F1 score for
the concatenated model and hereas we got 97.16 as a
accuracy score, 97.20 as a precision score, 97.16 as a
Recall score, 97.16 as a F1 score for the pruned model.

Model TP FP TN FN ACC Pre Recall F1-score

VGG19 297 3 266 34 93.83 94.30 93.83 93.81

RESNET50 293 7 265 35 93 93.37 93 92.99

Concatenated Model 291 9 282 18 95.50 95.54 95.50 95.55

Pruned Model 296 4 287 13 97.16 97.20 97.16 97.16

After the MRI slices had been separated into normal Table 1. describes the comparison of different metrics
and tumor classes using a DLA, the slices that included for four different algorithms.
tumors were subjected to additional scrutiny by a
medical professional. The primary focus of this study After the MRI slices had been separated into normal
was the implementation of a DLA. In order to and tumor classes using a DLA, the slices that included
construct a computerized model that can monitor the tumors were subjected to additional scrutiny by a
spread of ependymal tumors, it is possible to use the medical professional. The primary focus of this study
results of the system that has been proposed in was the implementation of a DLA. In order to
conjunction with the data that has been obtained construct a computerized model that can monitor the
clinically. spread of ependymal tumors, it is possible to use the
results of the system that has been proposed in
The following is future research's project scope: conjunction with the data that has been obtained
clinically.
I. Improving the handcrafted quality of the
feature vector by taking into account extra VI CONCLUSION
texture and form characteristics.
II. Making adjustments to the fully-connected
and drop-out layers in order to achieve more Based on the experimental results presented in the
accuracy in the categorization. table, several deep-learning architectures were
III. Making enhancements to the method of evaluated for their efficacy in enhancing brain cancer
feature concatenation in order to achieve diagnosis using 2D MRI data. The performance
superior results. metrics of each model, including VGG19,
RESNET50, a concatenated model, and a pruned
model, were thoroughly assessed in terms of true
positives (TP), false positives (FP), true negatives
(TN), false negatives (FN), accuracy (ACC), precision for segmentation and recognition of brain
(Pre), recall (Recall), and F1-score. tumors using MRI images,Pattern Recognition
Letters", 129 , pp. 181-189,(2020)
VGG19 demonstrated strong performance with
93.83% accuracy, 94.30% precision, 93.83% recall, 6. Nadeem, M.W.; Ghamdi, M.A.A.; Hussain,
and a corresponding F1-score of 93.81. Similarly, M.; Khan, M.A.; Khan, K.M.; Almotiri, S.H.;
RESNET50 exhibited competitive results with an Butt, S.A "Brain Tumor Analysis Empowered
accuracy of 93%, precision of 93.37%, recall of 93%, with Deep Learning": A Review, Taxonomy,
and an F1-score of 92.99. The concatenated model and Future Challenges. Brain Sciences 2020
further improved upon these metrics, achieving an
accuracy of 95.50%, precision of 95.54%, recall of 7. Praveen Kumar Ramtekkar, Anjana Pandey &
95.50%, and an F1-score of 95.55. Mahesh Kumar Pawar " Innovative brain
tumor detection using optimized deep learning
The pruned model exhibited the highest performance techniques": A Review Future Challenges.
among the evaluated architectures, achieving an Brain Sciences 2023
accuracy of 97.16%, precision of 97.20%, recall of
97.16%, and an F1-score of 97.16%. By leveraging 8. Rajan Hossain, Roliana Binti Ibrahim, Haslina
model pruning, the neural network's complexity was Binti Hashim " Automated Brain Tumor
significantly reduced without sacrificing predictive Detection Using Machine Learning: A
accuracy. This streamlined architecture not only Bibliometric Review": A Taxonomy, and
improved computational efficiency but also Future Challenges. Brain Sciences 2023
maintained high levels of precision and recall, crucial
9. Robin G utsche, Carsten Lowis, Karl
for reliable diagnosis in medical imaging applications.
Ziemons, Martin Kocher, Garry Ceccon"
Automated Brain Tumor Detection and
References
Segmentation for Treatment Response
Assessment Using Amino Acid PET": A
1. Wang, Y., Liu, Z., Liu, X., & Du, X., Review. Brain Sciences 2023
"Classification of Brain Tumor MRI Images
Using ResNet50. Journal of Medical Imaging 10. Mukul Aggarwal, Amod Kumar Tiwari, M
and Health Informatics, 11(9), 2284- Partha Sarathi & Anchit Bijalwan " An early
2290(2021) detection and segmentation of Brain Tumor
using Deep Neural Network": A Review,
2. Javaria Amin, Muhammad Sharif, Taxonomy, and Future Challenges. Brain
Anandakumar Haldorai, Mussarat Yasmin & Sciences 2023
Ramesh Sundar Nayak, " Brain tumor
detection and classification using machine 11. Rajinikanth, V.; Joseph Raj, A.N.; Thanaraj,
learning: a comprehensive survey". Journal of K.P.; Naik, G.R. "A Customized VGG19
Medical Imaging and Radiation Sciences, Network with Concatenation of Deep and
51(2), 262-268(2022) Handcrafted Features for Brain Tumor
Detection". Applied Sciences 2020
3. Vipin Y. Borole1 , Sunil S. Nimbhore2 , Dr.
Seema S. Kawthekar2,An " Image Processing 12. B. Kalyan and P. Chandrasekhar Reddy,
Techniques for Brain Tumor Detection". "Brain Tumour Detection Using
International Journal of Emerging Trends & Convolutional Neural Networks," 2022 10th
Technology in Computer Science (2020) International Conference on Reliability,
Infocom Technologies and Optimization
4. Javeria Amin a, Muhammad Sharif a, (Trends and Future Directions) (ICRITO),
Mussarat Yasmin a, Steven Lawrence Noida, India, 2022
Fernandes b," A distinctive approach in brain
tumor detection and classification using 13. Khan, Md & Rahman, Anichur & Debnath,
MRI"(2020). Tanoy & Razaul, Md & Nasir, Mostofa & S.
Band, Shahab & Mosavi, Amir & Dehzangi,
Iman (Abdollah)."Accurate Brain Tumor
Detection Using Deep Convolutional Neural
5. Sharif M.I., Li J.P., Khan M.A., Saleem M.A. Network". Computational and Structural
"Active deep neural network features selection Biotechnology Journal,2022
 CONFERENCE
Jaswanth Koripalli, P R S V Akhil, Pavan Kumar, Dr. K Kalimuthu, “Unveiling the Future of
HealthCare: Revolutionizing Brain Tumor Detection with Computer vision and AI”,
International Conference on Computational Intelligence and Communication Networks
(ICCICN-24) on 23 March 2024, Jaipur, India.