DEEP LEARNING MODELS TO DIAGNOSE BRAIN STROKE

BY
Asif Mahmud
ID: 183-15-2315
AND
Raisul Islam
ID: 183-15-2319

This Report Presented in Partial Fulfillment of the Requirements for the

Degree of Bachelor of Science in Computer Science and Engineering

Supervised By

Dr. S M Aminul Haque

Associate Professor
Department of CSE
Daffodil International University

Co-Supervised By

Md. Sabab Zulfiker

Lecturer (Senior Scale)
Department of CSE
Daffodil International University

DAFFODIL INTERNATIONAL UNIVERSITY

DHAKA, BANGLADESH
FEBRUARY 2023
 APPROVAL
This Project titled “Deep Learning Models To Diagnose Brain Stroke”, submitted by Asif
Mahmud, ID No: 183-15-2315, Raisul Islam, ID No: 183-15-2319 to the Department of Computer
Science and Engineering, Daffodil International University has been accepted as satisfactory for
the partial fulfillment of the requirements for the degree of Bachelor of Science in Computer
Science and Engineering and approved as to its style and contents. The presentation has been held
on date:

BOARD OF EXAMINERS

Name: Chairman
Designation
Department of Computer Science and Engineering
Faculty of Science & Information Technology
Daffodil International University

Name: Internal Examiner

Designation
Department of Computer Science and Engineering
Faculty of Science & Information Technology
Daffodil International University

Name: Internal Examiner

Designation
Department of Computer Science and Engineering
Faculty of Science & Information Technology
Daffodil International University

Name: External Examiner

Designation
Department of
University

©Daffodil International University

i
 DECLARATION

We hereby declare that, this project has been done by us under the supervision of Dr. S M Aminul
Haque, Associate Professor, Department of Computer Science and Engineering, Daffodil
International University. We also declare that neither this project nor any part of this project has
been submitted elsewhere for award of any degree or diploma.

Supervised by:

Dr. S M Aminul Haque

Associate Professor
Department of Computer Science and Engineering
Faculty of Science & Information Technology
Daffodil International University

Co-Supervised by:

Md. Sabab Zulfiker

Lecturer (Senior Scale)
Department of CSE
Daffodil International University

Submitted by:

Asif Mahmud
ID: -183-15-2315
Department of CSE
Daffodil International University

Raisul Islam
ID: -183-15-2319
Department of CSE
Daffodil International University

©Daffodil International University

ii
 ACKNOWLEDGEMENT

First we express our heartiest thanks and gratefulness to almighty God for His diving
blessing makes us possible to complete the final year project/internship successfully.

We really grateful and wish our profound our indebtedness to Dr. S M Aminul
Haque, Associate Professor, Department of CSE Daffodil International University,
Dhaka. Deep Knowledge & keen interest of our supervisor in the field of “Data
mining” to carry out this project. His endless patience ,scholarly guidance ,continual
encouragement, constant and energetic supervision, constructive criticism , valuable
advice ,reading many inferior draft and correcting them at all stage have made it
possible to complete this project.

We would like to express our heartiest gratitude to Md. Sabab Zulfiker, Lecturer
(Senior Scale), Department of CSE, for his kind help to finish our project and also to
other faculty member and the staff of CSE department of Daffodil International
University.

We would like to thank our entire course mate in Daffodil International University,
who took part in this discuss while completing the course work.

Finally, we must acknowledge with due respect the constant support and patients of
our parents.

©Daffodil International University

iii
 ABSTRACT

Brain Hemorrhage has gotten to be an extreme issue in the world. Brain Hemorrhage is also
called ‘Brain Stroke’ in our country. The amount of the stroke may be a particularly difficult
process among all of them; the brain stroke is the toughest one. There are three types of
strokes. Hemorrhagic Stroke, Ischemic Stroke and Transient Ischemic Attack (TIA).One of
the discrete methods that emerged as a first-line, radiation-free method of diagnosing brain
stroke is magnetic resonance imaging (MRI). Most of the people (87%) are affected by
Ischemic stroke. To distinguish this hemorrhage, we got to do a lab test at therapeutic. But
that's exceptionally costly for our country. So, we choose to do something for them. After
research, we developed a project which can detect this type of hemorrhage using deep
learning neural network based algorithm. When it comes to picture identification, deep
learning has made significant progress. To do that we collected raw data (CT Scan Copy)
from a number of hospital, we preprocessed it with a help of radiologist and finally trained
our model to nail our goal which is detecting hemorrhage stroke. Right now, we have been
found that VGG-16 achieves a high rate of accuracy with a minimal level of complexity.
It’s about 93 up percentage accuracy in total.Keywords: Brain Stoke, MRI, VGG-16.

©Daffodil International University

iv
 TABLE OF CONTENTS

CONTENTS PAGE NO

Board of examiners i

Declaration ii

Acknowledgment iii

Abstract iv

CHAPTER

1. INTRODUCTION 1-4

1.1 Introduction 1-2

1.2 Motivation 2

1.3 Rationale of the Study 2-3

1.4 Research Questions 3-4

1.5 Research Objective 4

1.6 Report Layout 4

2. BACKGROUND STUDY 5-8

2.1 Comparative Analysis 5-6

2.2 Research Summary 6-9

3. RESEARCH METHODOLOGY 10-19

3.1 Introduction 10

3.2 Data Collection Procedure 10

©Daffodil International University

v
 3.3 Data preparation 11

3.4 Research Subject and Instrumentation 11

3.5 Used Deep Learning Models 11-18

3.6 Implementation Procedure 19

4. EXPERIMENTAL RESULTS & DISCUSSION 20-23

4.1 Introduction 20

4.2 Confusion Matrix 20-22

4.3 Descriptive Analysis 22-23

5. IMPACT ON SOCIETY, ENVIRONMENT AND 24

SUSTAINABILITY
5.1 Impact on Society 24

5.2 Impact on Environment 24

5.3 Ethical Aspects 25

5.4 Sustainability Plan 25

6. CONCLUSIONS & FUTURE WORK 26

6.1 Conclusions 26

6.2 Future work 26

REFERENCES 27-28

©Daffodil International University

vi
 LIST OF TABLES
TABLES PAGE NO
Table 2.1.1: Comparative Analysis 05
Table 4.2.1: Confusion Matrix 20
Table 4.2.2: CNN confusion matrix 21
Table 4.2.3: VGG-16 confusion matrix 21
Table 4.2.4: Performance evaluation 22

©Daffodil International University

vii
LIST OF FIGURES
FIGURES PAGE NO
Figure 3.5.1: CNN RGB Image 12
Figure 3.5.2: CNN work procedure 13
Figure 3.5.3: Convolved feature 13
Figure 3.5.4: RGB color 14
Figure 3.5.5: Classification layer 15
Figure 3.5.6: Final layer 15
Figure 3.5.7: Polling 16
Figure 3.5.8: Polling image 17
Figure 3.5.9: VGG-16 Layer 18
Figure 3.6.10: Flow chart 19

©Daffodil International University

viii
 CHAPTER 1
Introduction

1.1 Introduction
One of the most widely studied topics today is brain stroke. Strokes are defined by the World
Health Organization which is an acute, focal or diffuse, dysfunction of the brain, originating for
vessels and lasting for a period longer than a day. The brain, which has billions of cells, is one
of the body's essential organs. Stroke is the third leading disease of death in the United States .
There are 15 million people suffer stroke worldwide every year (World Health Organization).
A high-grade stroke is also referred to as malignant. Cancerous strokes are not benign
strokes. As a result, it doesn't spread to other brain regions. The cancerous stroke, however,
is a malignant stroke. Because of this, it easily and quicklyspreads to other parts of the body
with arbitrary boundaries, making it dangerous. It results in instant death, according to
medical specialists. Clinical professionals must help patients access more effective e-health
care systems because it results in immediatemortality. Health care systems are useful in a
variety of medical fields. Biomedical imaging technologies based on computer vision are
becoming more important as they give radiologists recognition data for solving treatment-
related issues more effectively. There are numerous medical imaging techniques and
approaches, including X-rays, MRIs, ultrasounds, and computerized tomography (CT),
which have a significant impact on how patients are diagnosed and treated. In particular, an
MRI image can be used to track a stroke's progression using modeling. MRI images contain
more data thanCT or ultrasound images do. A magnetic resonance imaging (MRI) image
may be usedto detect irregularities in brain tissue and offer detailed information on brain
anatomy. It affects a lot of individuals right now, and therapy is pricey. People sometimes
want to avoid having a stroke because of the high cost of treatment, but a stroke can be fatal.
We are concerned about the rising number of stroke cases in Bangladesh and around the
world. There is a lot of research being done on brain stroke right now. Patients will find it
simpler to diagnose strokes thanks to our software, and doctors will find it simpler to keep
track of their patients. In order for our software to understand brain stroke, we used a lot of
images of the brain for brain stroke as well as a lot of photographs of what causes a brain
stroke. We believe that people will get far more from our project.

©Daffodil International University 1

1.2 Motivation
 We dream to bring wrong treatment into Level Zero

 We want to reduce pressure from the patient

 We want to see the effects of our innovation throughout the world.

 We want to gift the underworld a better healthy life

Stroke with intracerebral hemorrhage (ICH) is the most fatal and incapacitating type.
Hemorrhagic strokes, which are brought on by an immediate hemorrhage, account for 15%
of acute strokes. Intracerebral hemorrhage (ICH) and subarachnoid hemorrhage, which
account for around 5% of all strokes, are the two main kinds of hemorrhagic strokes. This
page deals with intracranial hemorrhage (ICH), a potentially fatal form of stroke that robs
the brain of oxygen and blood flow.

1.3 Rationale of the Study

This section reviews literature on brain stroke related work in computer science after
outlining research linked to addressing brain stroke.
Stroke happens when blood vessels in the brain burst and when the brain's blood supply is
cut off. Every year, 15 million people experience a stroke. In Europe, there are 650,000
stroke fatalities annually. In Bangladesh, a stroke is the third most common cause. The
World Health Organization ranks Bangladesh as having the 84th highest number of brain
stroke-related fatalities worldwide. Several recent research endeavors have concentrated
their emphasis on Brain Stroke issues in an effort to lessen the detrimental effects of brain
stroke globally. One of these is BCI (Brain-Computer Interface in Stroke Rehabilitation), a
computer-based system that converts brain impulses into commands for an output device to
carry out a desired action in the case of stroke.
Another project involves using angiography, MRI, and CT scans to diagnose strokes. Stroke
prevention and treatment methods for brain attack. Alteplase is currently the go-to treatment
for ischemic stroke, and computer technology and virtual reality help with brain
rehabilitation. Using new software, it is possible to determine the sort of brain damage and
the likelihood of recovery and recommend the most effective ways to observe the patient.

©Daffodil International University 2

1.4 Research Questions
 What is the Brain Stroke status globally?
 What are the associated factors with brain Stroke?
 Which algorithm perform well and why?
 What is image processing?
 How does CNN, VGG-16 work?

1.5 Research Objective

 To find out the Brain Stroke status.
 To identify the associated factors.
 Finding the best performing algorithm.
 Making awareness to reduce the percentage of brain stroke in every year.
 Giving message with the effect of lacking knowledge on brain stroke
causes to peoples.
 To increase the knowledge on brain stroke preventing factors.

©Daffodil International University 3

1.6 Report Layout
This research paper’s contents are:
I. In the first chapter we discuss about the motivation, rational study and
objectives.
II. In the second chapter we discus about the related work and research
summary.
III. Research methodology, data collection and preparation, Research
Subject and Instrumentation and discuss about the applied model in the
Chapter 3.
IV. The experimental evaluation and numerical result of the study are
discussed in Chapter 4.
V. The fifth chapter offer the summary, conclusion and future work.

©Daffodil International University 4

 CHAPTER 2
Background Study

2.1 Comparative Analysis

Brain stroke is a dangerous and acute cerebrovascular illness that is a significant cause of
mortality. It can be hemorrhagic, caused by blood flowing into brain tissue through a burst
intracranial arterial, or ischemic, caused by vascular blockage in the brain due to a blood
clot (thrombosis). Visible symptoms are identical in both situations, making differentiation
exceedingly difficult without modern imaging methods. Since human neural tissue is
rapidly destroyed as ischemic stroke advances, thrombolytic (or clot-dissolving) medicines
can be used to treat ischemic stroke during the first few hours. Contrarily, thrombolytics
can be hazardous or even deadly to individuals who are experiencing hemorrhagic stroke,
making it crucial to distinguish between the two. In order to expedite clinical choices on
therapies and to accelerate the recovery of patients suffering from acute stroke, a quick and
accurate diagnosis is essential for patient triage.

Table: 2.1.1 Comparative Analysis

Paper Name Working Idea What’s new?

Ischemic stroke detection Hemorrhagic stroke

Ischemic Stroke detection detection using a training
using MRI image, with a
using Image processing six-phase algorithm along data set and a test data set to
make the system learn and
and ANN with neural networking
then predict the result.
being applied at the ending
phase.

Detection of different types Including machine learning

Detection of Brain Stroke of brain strokes by analyzing along with image
from CT Scan Image CT scan images manually, processing for system
with no machine learning. learning and better
prediction.

Medical image analysis Using image analysis Precise training process to

methods in MR/CT- method of 2D/3D images detect hemorrhagic stroke
imaged acute-subacute along with mathematical with the help of neural
ischemic stroke lesion: model for detecting networking.
Segmentation, ischemic stroke.
prediction and insights
©Daffodil International University 5
 into dynamic evolution
simulation models. A
critical appraisal

Ischemic Stroke Detection Ischemic stroke detection No manual enhancement

System with a Computer- system with a computer- method introduced
Aided Diagnostic Ability aided diagnostic ability
Using an Unsupervised using a four-step
Feature Perception unsupervised feature
Enhancement Method perception enhancement
method.

2.2 Research Summary

There is a wealth of knowledge available in the healthcare sector. With the right use of
dependable data mining categorization methodology, early disease prediction is achievable.
The fields of medicine can benefit from machine learning and data mining. The majority of
it is executed properly. The study examines a set of risk factors that are monitored by
systems that look out for brain strokes. The suggested method for identifying, classifying,
and segmenting brain strokes seems to be quite accurate and effective [16]. To reach this
level of precision, automatic or somewhat automatic processes were necessary. The
research claims that segmentation and classification can be carried out automatically
utilizing a technique that utilizes CNN and operates in compact 3x3 kernels.

CNN is a machine learning method that use neural networks with layers for result
classification. Feature extraction, segmentation, average filtering, preprocessing, CNN, and
VGG-16 were some of the processes used in these at various stages. Through categorization
and identification using data mining techniques, significant links and patterns in the data
can be found. Segmentation, classification, identification, data collection, pre-processing,
and average filtering are used in this data mining and feature extraction for CNN. The
approach utilized in this study is pre-processing, which eliminates noisy data from the data
received during data collecting. Following is the average filtering. After that, pixels-based
detection is performed using the segmentation technique. Then, many features including
PSNR, MEAN, ENTROPY, SD, etc., were retrieved. T1 and FLAIR are the MRI imaging
techniques used by CNN. They also employ Multi FD to extract the tissue-texture of brain
strokes. For greater performance, they employ support vector machines and neural
networks. segmentation, classification, identification, data gathering, pre-processing, and
average filtering for feature extraction in CNN. The technique utilized in this study is pre-

©Daffodil International University 6

processing, which eliminates noisy data from the data received during data collecting.
Following is the average filtering. After that, pixels-based detection is performed using the
segmentation technique. Then, many features including PSNR, MEAN, ENTROPY, SD,
etc., were retrieved. T1 and FLAIR are the MRI imaging techniques used by CNN. They
also employ Multi FD to extract the tissue-texture of brain strokes. For greater performance,
they employ support vector machines and neural networks.

The technique utilized in this study is pre-processing, (Table 2.1.1) which eliminates noisy
data from the data received during data collecting. Following is the average filtering. After
that, pixels-based detection is performed using the segmentation technique. Then, many
features including PSNR, MEAN, ENTROPY, SD, etc, were retrieved. T1 and FLAIR are
the MRI imaging techniques used by CNN. They also employ Multi FD to extract the tissue-
texture of brain strokes. For greater performance, they employ support vector machines and
neural networks. The technique utilized in this study is pre-processing, which eliminates
noisy data from the data received during data collecting. Following is the average filtering.
After that, pixels-based detection is performed using the segmentation technique. Then,
many features including PSNR, MEAN, ENTROPY, SD, etc., were retrieved. T1 and
FLAIR are the MRI imaging techniques used by CNN. They also employ Multi FD to
extract the tissue-texture of brain strokes. For greater performance, they employ support
vector machines and neural networks. The technique utilized in this study is pre-processing,
which eliminates noisy data from the data received during data collecting. Following is the
average filtering. After that, pixels-based detection is performed using the segmentation
technique. Then, many features including PSNR, MEAN, ENTROPY, SD, etc., were
retrieved. T1 and FLAIR are the MRI imaging techniques used by CNN. They also employ
Multi FD to extract the tissue-texture of brain strokes. For greater performance, they employ
support vector machines and neural networks. The technique utilized in this study is pre-
processing, which eliminates noisy data from the data received during data collecting.
Following is the average filtering. After that, pixels-based detection is performed using the
segmentation technique. Then, many features including PSNR, MEAN, ENTROPY, SD,
etc., were retrieved. T1 and FLAIR are the MRI imaging techniques used by CNN. They
also employ Multi FD to extract the tissue-texture of brain strokes. For greater performance,
they employ support vector machines and neural networks. The technique utilized in this
study is pre-processing, which eliminates noisy data from the data received during data
collecting. Following is the average filtering. After that, pixels-based detection is performed
using the segmentation technique. Then, many features including PSNR, MEAN,

©Daffodil International University 7

ENTROPY, SD, etc., were retrieved. T1 and FLAIR are the MRI imaging techniques used
by CNN. They also employ Multi FD to extract the tissue-texture of brain strokes. For
greater performance, they employ support vector machines and neural networks. The
technique utilized in this study is pre-processing, which eliminates noisy data from the data
received during data collecting. Following is the average filtering. After that, pixels-based
detection is performed using the segmentation technique. Then, many features including
PSNR, MEAN, ENTROPY, SD, etc., were retrieved. T1 and FLAIR are the MRI imaging
techniques used by CNN. They also employ Multi FD to extract the tissue-texture of brain
strokes. For greater performance, they employ support vector machines and neural
networks. The technique utilized in this study is pre-processing, which eliminates noisy data
from the data received during data collecting. Following is the average filtering. After that,
pixels-based detection is performed using the segmentation technique. Then, many features
including PSNR, MEAN, ENTROPY, SD, etc., were retrieved. T1 and FLAIR are the MRI
imaging techniques used by CNN. They also employ Multi FD to extract the tissue-texture
of brain strokes. For greater performance, they employ support vector machines and neural
networks. The technique utilized in this study is pre-processing, which eliminates noisy data
from the data received during data collecting. Following is the average filtering. After that,
pixels-based detection is performed using the segmentation technique. Then, many features
including PSNR, MEAN, ENTROPY, SD, etc., were retrieved. T1 and FLAIR are the MRI
imaging techniques used by CNN. They also employ Multi FD to extract the tissue-texture
of brain strokes. For greater performance, they employ support vector machines and neural
networks. The technique utilized in this study is pre-processing, which eliminates noisy data
from the data received during data collecting. Following is the average filtering. After that,
pixels-based detection is performed using the segmentation technique. Then, many features
including PSNR, MEAN, ENTROPY, SD, etc., were retrieved. T1 and FLAIR are the MRI
imaging techniques used by CNN. They also employ Multi FD to extract the tissue-texture
of brain strokes. For greater performance, they employ support vector machines and neural
networks. The technique utilized in this study is pre-processing, which eliminates noisy data
from the data received during data collecting. Following is the average filtering. After that,
pixels-based detection is performed using the segmentation technique. Then, many features
including PSNR, MEAN, ENTROPY, SD, etc., were retrieved. T1 and FLAIR are the MRI
imaging techniques used by CNN. They also employ Multi FD to extract the tissue-texture
of brain strokes. For greater performance, they employ support vector machines and neural
networks.

©Daffodil International University 8

In AI, they also employ the Anaconda framework. TensorFlow, too. They also experiment
with multi-model brain stroke segmentation, which combines a variety of segmentation
algorithms to achieve high efficiency. But it is quite difficult. They understand that data
clustering, approximation, pattern matching, optimization methods, and classification
approaches all benefit greatly from neural networks. They exclaim that there are three
different types of neural networks. These networks are feed-forward, recurrent, and
feedback. The feed forward neural network has a single layer and a multilayer architecture.
In a single layer network, the hidden layer is not visible. However, it simply has input and
output layers. There are a total of three layers in the multilayer. These are the output layer,
hidden layer, and input layer. The CNN-based brain stroke classification system divides the
technique into two phases: training and testing. Pre-processing, feature extraction, and
classification using the Loss function are carried out during the training phase in order to
develop a classification model. The author of this article studies, reviews, and describes the
methods used to detect brain strokes using MRI and artificial neural networks (ANN). After
gathering the MRI image data, the Computer Aided Detection System (CAD) processed it
in a number of ways. The initial stage of processing MRI images involved pre- and post-
processing to improve and prepare them for analysis. In the second stage, the MRI images
were segmented using a threshold utilizing the mean gray level technique. In the second
stage, when features are extracted from images using statistical feature analysis, the
Particles based on the dependency matrix for the spatial gray levels (SGLD) were selected
as the most important characteristic to determine stroke localization. This classification is
used to classify both abnormal and typical MR images of various types of brain strokes.
Using this differential deep-CNN for image analysis of a pixel-directed pattern with contrast
computations has a benefit. A dataset of 25,000 brain MRI scans, both abnormal and
normal, was used. Their proposed design generates the best accurate feature map to
differentiate between LG and HG gliomas in comparison to the 2D CNN version. Their
Deep- CNN offers high-speed and precise detection and classification skills when
compared. Convolutional layers, pooling layers, and fully convolutional layers are the three
main building blocks of a CNN.

©Daffodil International University 9

 CHAPTER 3
Research Methodology

3.1 Introduction
In the field of cancer research, deep learning (DL) algorithms for deciphering magnetic
resonance imaging (MRI) pictures and forecasting brain strokes have advanced quickly. We
found that DL models' high accuracy, interpretability, and explanation varied significantly.
As a result, we describe a convolutional neural network-based explanation-driven deep
learning model.

3.2 Data Collection Procedure

Kaggle has provided data. Kaggle is the location where data scientists spend their evenings
and weekends. It is a crowdsourced platform that recruits, teaches, trains, and evaluates a
number of data scientists around the world to handle issues utilizing predictive analytics,
machine learning, and data science. It has roughly 536,000 active members, and each month
it receives about 150,000 entries. began in Melbourne, Australia. When Kaggle first came
in Silicon Valley in 2011, investors Max Levchin, PayPal, Index, Hal Varian, Google's
Chief Economist, and Khosla Ventures helped the company raise roughly $11 million. In
the end, Google bought Kaggle in March 2017. On Kaggle, data science enthusiasts from
all around the world compete for prizes and rankings advancement. There are currently only
94 Kaggle Grandmasters [20] left. In every area of employment, data always comes first in
order to produce any kind of output. We required the datasets to finish this deep learning
process, so we downloaded them from the Kaggle website. Finding the most appropriate
datasets for our needs was the biggest challenge using Kaggle, a massive data warehouse.
Therefore, we carefully examined the website and other reference documents that we had
read for our study summary. We also investigated the UCI machine learning repository,
which has one of the largest datasets and a paper warehouse. However, UCI was unable to
help us with the datasets, and we unsuccessfully sent more resources from the articles we
had found to Kaggle Data Warehouse before they were able to help us find the datasets.

©Daffodil International University 10

3.3 Data preparation
One of our study project's most difficult tasks is data collection. But the main challengewas
to applied the algorithm. As we are not using framework, we have used the raw python
coding to do that the data must need to prepare and specified folder to access the datasets.
There are numerous numbers of cluster datasets which we have cleaned and integrate for
our purpose at last the datasets was ready to apply the algorithms.

3.4 Research Subject and Instrumentation

More sophisticated and complex instruments are required due to the dramatically larger
amount of data generated by technology breakthroughs. There is a constant need for new
approaches and tools that can assist in turning massive data into useful information and
knowledge, even while advancements in Deep Learning technologies have made large data
collecting much more prudently necessary. The idea and methods uphold the tradition of
giving users access to the theory and practice of finding hidden patterns in huge data sets
through knowledge and application. In this study, we use the CNN model, the VGG-16
model, and Python 3.5 to investigate the topic of " Disease Diagnosis by X-ray or CT-scan
Using Deep Learning Models".

3.5 Used Deep Learning Models:

To classify brain strokes, we used two different deep learning models. Convolutional Neural
Network (CNN) and Visual Geometry Group-16 are two examples (VGG-16). Here is a
quick explanation of those.

a) Convolutional Neural Network (CNN):

Convolutional neural networks, often known as CNNs, are a deep learning method that are
extensively employed for the classification and recognition of objects and images [21].
Deep Learning uses a CNN to identify things in a picture as a result. Compared to several
other classification models, ConvNet requires a lot less pre-processing. Simple sorting
methods are still effective, but with enough training, they can learn about new features like
ConvNet [22]. The usage of CNNs is widespread in a variety of tasks and applications, such
as voice recognition in natural language processing, motion detection, image analysis
problems, computer vision, and auto vehicle sensing system.

©Daffodil International University 11

Because of its significant contribution to all of these quickly evolving and increasing
disciplines, CNNs are widely utilized for deep learning.
Let's first talk about the fundamentals, such as what an image is and how it is shown, before
we examine how CNN functions. A grayscale image and an RGB picture both have a single
plane, but an RGB image is nothing more than a matrix of pixel values. See this picture.

Figure 3.5.1: CNN RGB Image

Figure 3.5.1. Illustrates the preceding discussion that any color image contains three
channels: red, green, and blue. An image can exist in a variety of color spaces, including
grayscale, CMYK, and HSV. To keep things simple, let's only use grayscale images to show
how CNNs operate.

©Daffodil International University 12

 Figure 3.5.2: CNN work procedure

Figure 3.5.2. There is a convolution, as shown in the graphic above. A filter or kernel (3x3
matrix) is applied to the input picture to produce the convolved feature. This convolved
function is being passed across to the layer following table.

Figure 3.5.3: Convolved feature

Figure 3.5.3. The green box is an image matrix, and the yellow box is a filter, which is a
matrix of 0s and 1s that defines a transformation. These features are combined into a feature
map as the filter passes through the picture pixels in each subregion of the input volume
using a unique type of matrix multiplication.

©Daffodil International University 13

 See this animation to learn more about how the RGB color channel functions.

Figure 3.5.4: RGB color

Figure 3.5.4. Convolutional neural networks are made of artificial neurons that are stacked
in layers. Similar to their biological counterparts, artificial neurons are mathematical
constructs that compute the activation value from the weighted sum of a set of inputs. When
an image is fed into a ConvNet, each layer generates a set of activation functions, which are
subsequently transferred to thefollowing layer.
Horizontal or diagonal edges are fundamental characteristics that are frequentlyretrieved by
the first layer. Following receipt of this output, the following layer searches for more
intricate features like corners and multiple edges. As we dig deeper, the network may
someday be able to distinguish objects, faces, and other fine characteristics.

©Daffodil International University 14

 Figure 3.5.5: Classification layer

Figure 3.5.5. Based on the activation map of the final convolution layer, the classification
layer generates a number of optimism ratings (numbers between 0 and 1), indicating how
likely it is for the image to belong to a "class". The last

©Daffodil International University 15

layer's output, for example, can be the likelihood assuming the input image contains any of
the cats, dogs, or horses detected by the ConvNet.

Figure 3.5.6: Final layer

(https://www.ijraset.com/researc
h-paper/suspicious-human-
activity-recognition-and-
alarming-system)

Similar to the convolutional layer (figure 3.5.6), the pooling layer is in charge of reducing
thespatial size of the convolutional feature. By reducing the size, less CPU processing power
will be required to process the data. There are two types of pooling: average pooling and
maximal pooling. I've only used Max Pooling once, but so far there have been no problems.

Figure 3.5.7: Polling

©Daffodil International University 16

As a result, in Max Pooling (Figure 3.5.7), we select a portion of the image that the kernel
hascovered in order to find the maximum value of a pixel. Max Pooling also acts as a sound
deadener. Along with entirely rejecting the noisy activations, de- noising and
dimensionality reduction are also carried out.
On the Figure 3.5.8 average pooling yields the mean of all the values in the region of the
image that the Kernel has covered. With average pooling, noise can only be reduced by
reducing dimensionality. Therefore, we may say that Max Pooling significantly
outperforms Average Pooling.

Figure 3.5.8: Polling image

b) Visual Geometry Group-16 (VGG-16):
VGG-16 refers to a 16-layer convolutional neural network. On over a million images from
ImageNet's collection, a pre-trained version of the network has been trained . VGG is
composed of two fully connected layers, which are recommended by a softmax activation
function for the output layer [24]. Many animals, a keyboard, a mouse, and a pencil are
among the 1000 possible item categories that may be assigned to photos using a trained
network. As a result, the network has accumulated a wide range of feature-rich picture
representations. With 97.1% accuracy, the image segmentation and object identification
system VGG16 classifies1000 images into 1000 distinct categories.

©Daffodil International University 17

It is a tried-and-true technique for image classification that is easy to incorporate into
transfer learning. Although this form is incredibly straightforward, tasteful, and simple to
use, it has shortcomings. The total number of parameters in this model was 138M, and it is
more than 500MB. The model's employment is significantly constrained, especially in edge
computing, because the compared to similar is longer. Second, the problem of vanishing or
expanding gradients is not specifically addressed. Both GoogleNet modules and ResNet's
skip connections were used to overcome this issue.

VGG Architecture:

VGG offers two models: the VGG-16 and the VGG-19. In this blog, our dataset will be
classified using VGG-16. VGG-16 is primarily composed of convolution, pooling, and fully
linked layers. The input size of thecorresponding network is 224 by 224 [25].

● Convolution layer: In this layer, filters are used to extract features from photos.
The most important parameters are the stride and kernel size.
● The pooling layer's task is to shrink a network's spatial extent in orderto scale
back on processing and parameterization.
● Fully Connected: These connections to the preceding levels are all fully coupled
in a simple neural network.
Given figure shows the architecture of the model:

Figure 3.5.9: VGG-16 Layer

Source Import a pre-trained model using PyTorch, Depending on our goals, we can either
skip the final fully connected layer or add a new completely connected layer at the end
before running the model to include transfer learning.

©Daffodil International University 18

 3.6 Implementation Procedure:

Figure 3.6.10: Flow chart

To begin this procedure, we create a clear diagram that lays out the steps we will take to
complete our task. The process was started in the order shown in the diagram, and then we
obtained data from Kaggle. Why Kaggle? Since we were unable to locate the real dataset
from any institution, we made the decision to focus on the demo datasets in order to gain
expertise before moving on to the real datasets. we keep working as per the flowchart
diagram.
Figure 3.6.10. Preprocessing of data is the process of cleaning and filtering datasets before
applying algorithms to them once they have been acquired. The data sets were then divided
into two portions: testing and training, respectively. The training (80%) percentage of the
data and the rest (20%) of the data as testing dataset. After completing the process, we have
applied two algorithms one is CNN and another one is VGG -16 to find out the associated
factors that is causes the brain stroke and the accuracy of the result. Between these two
algorithms VGG-16 perform better which isshown in the flowchart.

©Daffodil International University 19

 CHAPTER 4
Experimental Results & Discussion

4.1 Introduction
In the part before, we spoke about the dataset and the methods for processing it. This section
will provide an explanation of various models' findings that make use of the prepared data.
The results of CNN and VGG-16 are being evaluated to ascertain which method offers the
most accuracy.

4.2 Confusion Matrix

A technique for measuring a classification algorithm's effectiveness is the confusion matrix
(Table 4.2.1). Classification accuracy alone may be deceptive if our dataset has more than
two classes or fewer observations in certain classes than others. We can better understand
the successes and shortcomings of the categorization model by developing a confusion
matrix. This is the table of confusion matrix:

Table 4.2.1: CONFUSION MATRIX

Predicted

No Stroke Stroke
Actual

No Stroke True Negative (TN) False Positive (FP)

Stroke False Negative (FN) True Positive (TP)

The following performance metrics are obtained from the segmentation and
classifiedresults.

©Daffodil International University 20

 Table 4.2.2: CNN CONFUSION MATRIX

Predicted

No Stroke Stroke
Actual

No Stroke 527 51

Stroke 14 223

Table 4.2.3: VGG-16 CONFUSION MATRIX

Predicted

No Stroke Stroke
Actual

No Stroke 557 24

Stroke 17 221

Accuracy is defined as the percentage of correctly predicted data points among all data
points. The official definition of accuracy is as follows:
𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑐𝑜𝑟𝑟𝑒𝑐𝑡 𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛𝑠
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = × 100%
𝑇𝑜𝑡𝑎𝑙 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑖𝑜𝑛𝑠

Binary classification accuracy can also be assessed in terms of positives andnegatives.as

shown below:

𝑇𝑃 + 𝑇𝑁
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = × 100%
𝑇𝑃 + 𝑇𝑁 + 𝐹𝑃 + 𝐹𝑁

©Daffodil International University 21

A machine-learning algorithm's sensitivity reveals how well it can identify
favorable traits. The sensitivity, also referred to as the true positive rate, can be
computed as follows:

𝑇𝑃
𝑆𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦 = × 100%
𝑇𝑃 + 𝐹𝑁

Specificity is the proportion of genuine negatives that the model successfully

detects. Calculating specificity involves doing the following:

𝑇𝑁
𝑆𝑝𝑒𝑐𝑖𝑓𝑖𝑐𝑖𝑡𝑦 = × 100%
𝑇𝑁 + 𝐹𝑃

Precision is defined as the ratio of True Positives to All Positives. A formula for
calculating precision is provided below:

𝑇𝑃
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = × 100%
𝑇𝑃 + 𝐹𝑃

The F1 score is a machine learning statistic that can be used in classification

models. The following can be used to calculate the F1 score:

2𝑇𝑃
𝐹1 𝑠𝑐𝑜𝑟𝑒 = × 100%
2𝑇𝑃 + 𝐹𝑃 + 𝐹𝑁

Table 4.2.4. The performance of each algorithm is described in Table 1. Based on the
efficiency and validity of different algorithms, we determined the optimal solution for our
model. Based on this F1-Score, it is obvious that VGG-16 performs the best in terms of
accuracy, sensitivity, specificity, and precision. After considering everything, this approach
can be used to optimize model performance.
Table 4.2.4: PERFORMANCE EVALUATION

No Model Name Accuracy Sensitivity Specificity Precision F1-score

1 VGG-16 97.21 % 93.22 % 96.02 % 90.53 % 91.85 %

4.3 Descriptive Analysis

After going through the entire application algorithm procedure, we finally get a result that
shows how important the F1 score, precision, sensitivity, and specificity are.

©Daffodil International University 22

Whereas the f1 score is composed of two fundamental components, recall and precision,
both expressed as percentages and Vgg-16 achieved the best result when coupled as
harmonic mean to assign a single number that is simple to understand and performs well in
the accuracy matrix. As you can see in the table, CNN performs less because it contains a
smaller or no fraction of the larger section. We tested various folding methods in epochs
and determined that vgg-16 is the best option for our dataset. VGG-16 shows the highest
Accuracy 95.21%, in Specificity VGG-16 shows 96.02%, in Precision VGG-16 shows
90.53%, and in F1-score VGG-16 shows 97.1%. All of those are greater than CNN. On the
other hand, in Sensitivity CNN shows 93.64%.

©Daffodil International University 23

 CHAPTER 5
Impact on Society, Environment and Sustainability

5.1 Impact on Society

When it comes to identifying brain strokes, deep learning models have the potential to
significantly affect society in a number of ways.
 Enhanced diagnostic precision: These models can quickly and accurately evaluate
medical images like CT and MRI scans, minimizing human error and enhancing
patient outcomes.
 Faster treatment: Patients can receive care more quickly after being diagnosed with
strokes, reducing brain damage and enhancing recovery.
 Reduced workload for radiologists: Deep learning models can automate the
diagnosis process, reducing the effort for radiologists and other medical
professionals and allowing them to concentrate on other duties.
 Cost-effectiveness: Compared to conventional approaches, using deep learning
models to detect brain strokes may be more economical since it eliminates the need
for human labor and enhances the effectiveness of the healthcare system.
 Better health outcomes: Deep learning models have the potential to save lives by
identifying brain strokes earlier and more accurately.
Overall, using deep learning to brain stroke diagnosis can result in better patient care,
quicker recovery times, and better health outcomes, making it an important tool in the
medical profession.

5.2 Impact on Environment

By minimizing the need for intrusive procedures and physical examinations, deep learning
models for detecting brain stroke can have a beneficial environmental impact by lowering
the quantity of medical waste created and the carbon footprint of healthcare. Deep learning
models can also assist to increase the precision and effectiveness of diagnoses, perhaps
lowering the need for follow-up tests and treatments. But it's important to remember that
the environmental effect of deep learning models will rely on the precise deployment and
training infrastructure employed, as well as the system's total energy usage.

©Daffodil International University 24

5.3 Ethical Aspect
When utilizing deep learning models to detect brain strokes, there are various ethical issues
to keep in mind. The likelihood of bias in the training data is a significant factor to take into
account. The model may give unreliable or biased results if the data used to train it is not
representative of the population it will be applied to. Furthermore, there is a chance of
overdiagnosis or underdiagnosis, which may result in unneeded therapy or a failure to offer
treatment that is required. It's crucial to take into account any potential unintended
outcomes, such as prejudice towards particular groups of individuals. It's crucial to make
sure that the data used to train the model is representative and varied in order to reduce these
risks, and to periodically assess and update the model to enhance performance.

5.4 Sustainability plan

Several crucial elements would be included in a sustainability strategy for deep learning
models used to diagnose brain strokes:
 Regularly updating the model with new data to ensure that it remains accurate and
up-to-date. This could include new imaging techniques, patient data, and
advancements in the understanding of brain strokes.
 Keeping an eye on the model's performance and making changes as necessary. This
can entail modifying the training data, the model's parameters, or the model's
architecture.
 Ensuring that model users, such as radiologists or neurologists, receive continual
training and assistance so they can correctly interpret the model's output and make
appropriate diagnoses.
 Exploring ways to make the model more accessible to a wider range of users, such
as through cloud-based deployment or mobile app integration.
 Constantly looking for methods to reduce the model's environmental effect, for as
by adopting more energy-efficient hardware or looking at alternative deep learning
architectures.
 To ensure that the model is producing accurate and trustworthy results, its
performance is compared to other diagnostic tools, such as human experts.
Additionally, the model is continuously seeking feedback from end users, such as
radiologists or neurologists, in order to identify areas for improvement and make
necessary adjustments.

©Daffodil International University 25

 CHAPTER 6
Conclusions & Future work
6.1 Conclusions
Segmenting medical images is a challenging challenge because of because of how
complicated the images are and because there aren't any anatomical models that can
accurately represent all of the deformations in each component. The initial cluster size and
cluster centers can be successfully managed with the help of the suggested technique. In
order to segment, BWT techniques are utilized, which have lesser accuracy and processing
speed. This research offers an almost entirely automated method for dividing brain tissue.
This suggested method's main goal is to enable neurosurgeons or human experts to quickly
identify the patients. The experimental findings show that the accuracy is 98.5 percent
greater than that of cutting-edge technologies. The processing time, system complexity, and
memory space requirements of the techniques can all be further decreased. The similar
approach can be used to recognize and investigate diverse conditions present in other body
areas (kidney, liver, lungs, etc.). Future studies can combine greater segmentation and
extraction efficiency algorithms with real time photos, clinical cases, a larger data set with
a range of settings, and different classifiers that apply optimization methodology to boost
accuracy.

6.2 Future Work

Due of several restrictions, we are unable to completely clear the dataset or add more
algorithms at this time. But since we haven't encountered it. In order to use additional
algorithms and achieve the best accuracy, we must look further. And this can assist
humanity in finding a solution to the "Brain Stroke" disease problem not just in this country
but throughout the world. And this firmly maintains that we can use the algorithm to
accomplish much greater accuracy before we can primarily serve it over the world.

©Daffodil International University 26

 References
1. Zhang, Jiachi, et al. "Brain stroke segmentation based on refined fully convolutional neural networks with
a hierarchical dice loss." arXiv preprint arXiv:1712.09093 (2017).
2. Christ, MC Jobin, S. Sivagowri, and P. Ganesh Babu. "Segmentation of brain strokes using meta heuristic
algorithms." Open J. Commun. Softw 2014.1 (2014): 1-10.
3. Khilariwal, Swapnil, Richa Verma, and Madhur Deo Upadhayay. "Design of notch antenna for5 GHz
high speed LAN." 2016 International Conference on Wireless Communications, Signal Processing and
Networking (WiSPNET). IEEE, 2016.
4. Chato, Lina, and Shahram Latifi. "Machine learning and deep learning techniques to predict overall
survival of brain stroke patients using MRI images." 2017 IEEE 17th international conference on
bioinformatics and bioengineering (BIBE). IEEE, 2017.
5. Goo, Hyun Woo, and Young-Shin Ra. "Advanced MRI for pediatric brain strokes with emphasison clinical
benefits." Korean journal of radiology 18.1 (2017): 194-207.
6. Apriantoro, Nursama Heru, and Christianni Christianni. "Analisis Perbedaan Citra Mri Brain Pada
Sekuent1se Dan T1flair." Sinergi: Jurnal Teknik Mercu Buana 19.3 (2015): 206-210.
7. Kaur, Hardeep, and Jyoti Rani. "MRI brain image enhancement using Histogram equalization
Techniques." 2016 international conference on wireless communications, signal processing and
networking (WiSPNET). IEEE, 2016.
8. Siar, Masoumeh, and Mohammad Teshnehlab. "Brain stroke detection using deep neural network and
machine learning algorithm." 2019 9th international conference on computer and knowledge engineering
(ICCKE). IEEE, 2019.
9. Van Eycke, Yves-Rémi, et al. "Segmentation of glandular epithelium in colorectal tumours to
automatically compartmentalise IHC biomarker quantification: A deep learning approach." Medical
image analysis 49 (2018): 35-45.
10. Menze, Bjoern H., et al. "The multimodal brain stroke image segmentation benchmark (BRATS)." IEEE
transactions on medical imaging 34.10 (2014): 1993-2024.
11. Huda, Shamsul, et al. "A hybrid feature selection with ensemble classification for imbalanced healthcare
data: A case study for brain stroke diagnosis." IEEE access 4 (2016): 9145-9154.
12. Seetha, J., and S. Selvakumar Raja. "Brain stroke classification using convolutional neural networks."
Biomedical & Pharmacology Journal 11.3 (2018): 1457.
13. Anitha, V., and S. J. I. C. V. Murugavalli. "Brain tumour classification using two‐tier classifier with
adaptive segmentation technique." IET computer vision 10.1 (2016): 9-17.
14. Siar, Masoumeh, and Mohammad Teshnehlab. "Brain stroke detection using deep neural network and
machine learning algorithm." 2019 9th international conference on computer and knowledge engineering
(ICCKE). IEEE, 2019.
15. Abdalla, Hussna Elnoor Mohammed, and M. Y. Esmail. "Brain stroke detection by using artificial neural
network." 2018 International Conference on Computer, Control, Electrical, andElectronics Engineering
(ICCCEEE). IEEE, 2018.

©Daffodil International University 27

16. Hemanth, G., M. Janardhan, and L. Sujihelen. "Design and implementing brain stroke detection using
machine learning approach." 2019 3rd international conference on trends in electronics and informatics
(ICOEI). IEEE, 2019.
17. Shankaragowda, et al BB, M. Siddappa, and M. Suresha. "A novel approach for the brain strokedetection
and classification using support vector machine." 2017 3rd International Conference on Applied and
Theoretical Computing and Communication Technology (iCATccT). IEEE, 2017.
18. Gaur, Loveleen, et al. "Explanation-Driven Deep Learning Model for Prediction of Brain Tumour Status
Using MRI Image Data." Frontiers in Genetics (2022): 448.
19. Usmani, Z. "What is Kaggle, Why I Participate, what is the Impact." Getting Started (2018).
20. Collect information about Brain stroke, available at <<https://www.cancer.net/cancer- types/brain-
stroke/statistics>>, last accessed on 13-03-2022 at 09:15 PM.
21. Tamm, Markus-Oliver, Yar Muhammad, and Naveed Muhammad. "Classification of vowels from
imagined speech with convolutional neural networks." Computers 9.2 (2020): 46.
22. Wang, Xinggang, et al. "Searching for prostate cancer by fully automated magnetic resonance imaging
classification: deep learning versus non-deep learning." Scientific reports 7.1 (2017): 1-8.
23. Girshick, R., and R. C. N. N. Fast. "IEEE Int. Conf. Comput. Vis." Santiago, Chile, December(2015): 7-
13.
24. Van Grinsven, Mark JJP, et al. "Fast convolutional neural network training using selective datasampling:
Application to hemorrhage detection in color fundus images." IEEE transactions on medical imaging 35.5
(2016): 1273-1284.

©Daffodil International University 28