DETECTION OF BLOOD CANCER AND

ITS STAGES USING CNN

Submitted in partial fulfillment of the requirements for the award of

Bachelor of Technology degree in Biomedical Engineering
By

H ASHWATHI (37240013)

K DEVISRI (37240020)

DEPARTMENT OF BIOMEDICAL ENGINEERING

SCHOOL OF BIO AND CHEMICAL ENGINEERING

SATHYABAMA
INSTITUTE OF SCIENCE AND TECHNOLOGY
(DEEMED TO BE UNIVERSITY)
Accredited with Grade “A” by NAAC
JEPPIAAR NAGAR, RAJIV GANDHI SALAI, CHENNAI • 600 119

MARCH – 2021
 SATHYABAMA
INSTITUTE OF SCIENCE AND TECHNOLOGY
(DEEMED TO BE UNIVERSITY)
Accredited with Grade “A” by NAAC
JEPPIAAR NAGAR, RAJIV GANDHI SALAI, CHENNAI– 60119
www. sathyabama.ac.in

DEPARTMENT OF BIOMEDICAL ENGINEERING

BONAFIDE CERTIFICATE

This is to certify that this Project Report is the bonafide work of H

ASHWATHI (Reg.No. 37240013), K DEVISRI (Reg. No. 37240020) who carried out
the project entitled “DETECTION OF BLOOD CANCER AND ITS STAGES USING
CNN” under our supervision from October 2020 to March 2021.

Dr. S. KRISHNAKUMAR, M.Sc., Ph.D.,

Guide and supervisor

Dr. T. SUDHAKAR, M.Sc., Ph.D.,

Head of the Department

Submitted for Viva voce Examination held on 9.4.2021

Internal Examiner External Examiner

 DECLARATION

We, H ASHWATHI (37240013), K DEVISRI (37240020) hereby declare that the

Project Report entitled “DETECTION OF BLOOD CANCER AND ITS STAGES
USING CNN” done by us under the guidance of Dr. S. Krishnakumar,
Department of Biomedical Engineering is submitted in partial fulfillment of
the requirements for the award of Bachelor of Technology degree in
Biomedical Engineering.

DATE : 10.04.21 1.

PLACE: CHENNAI 2.

SIGNATURE OF THE CANDIDATES

 ACKNOWLEDGEMENT

We are pleased to acknowledge our sincere thanks to Board of Management of

SATHYABAMA for their kind encouragement in doing this project and for
completing it successfully. We am grateful to them.

We convey our thanks to Dr.J.Premkumar M.Sc., Ph.D., and Dr.T.Sudhakar

M.Sc., Ph.D., Head of the Department, Department of Biomedical
Engineering for providing me necessary support and details at the right time
during the progressive reviews.

We would like to express our sincere and deep sense of gratitude to our Project
Guide Dr. S. Krishnakumar, M.Sc., Ph.D., Department of Biomedical
Engineering for his valuable guidance, suggestions and constant
encouragement paved way for the successful completion of our project work.

We wish to express our thanks to all Teaching and Non-teaching staff members
of the Department of Biomedical Engineering who were helpful in many ways
for the completion of the project.
 ABSTRACT

Identification of blood disorders are mainly done by hematologists or experts, by

examining the images of blood cells obtained with the help of microscopes. It helps to
classify several diseases related to blood cells such as blood cancer, sickle cell
diseases, polycythemia, etc. This paper is an overview of various studies conducted in
the field of blood cancer, specifically to detect and classify different types of leukemia.
This study focuses on the techniques used to segment and detect the type of leukemia
by analyzing different features of the digital images of the white blood cells. Variations
in these features are used as the classifier inputs which give information about different
types of leukemia. To understand relative merits and demerits, comparisons of different
techniques used for segmentation and classification are given .

i
 TABLE OF CONTENTS

CHAPTER TITLE PAGE NO

NO
ABSTRACT i
LIST OF ABBREVIATION v
LIST OF FIGURES vi
1 INTRODUCTION 1
1.1 GENERAL 1
1.2 THEORY 2
1.3 TECHNOLOGIES USED 5
1.4 CONVOLUTIONAL NEURAL NETWORK 6
1.5 POOLING 8
2 LITERATURE REVIEW 10
3 AIM AND SCOPE 14
3.1 AIM 15
3.2 SCOPE 15
3.3 EXISTING SYSTEM 15

ii
 3.4 PROPOSED SYSTEM 15
4 MATERIALS AND METHODS 16
4.1 TECHNOLOGICAL BACKGROUND 16
4.2 DEEP LEARNING ARCHITECTURES 17
4.3 ACTIVATION FUNCTIONS 18
4.3.1 RECTIFIED LINEAR UNIT 19
4.3.2 SOFTMAX 19
4.3.3 CNN 19
4.3.4 PREPROCESSING OF GENOMIC DATA 20
4.3.5 PREPROCESSING OF IMAGE DATA 21
4.4 EVALUATION MODEL 21
4.4.1 CONFUSION MATRIX 22
4.4.2 RECALL 22
4.4.3 ACCURACY 23
4.4.4 PRECISION 23
4.4.5 F1 SCORE 23
4.4.6 VALIDATION 24
4.4.7 LOGARITHMIC LOSS 24
4.4.8 INFORMATION AND PRIVACY 24
4.5 SYSTEM DESIGN 26
4.6 COLLECTING DATASETS 26
4.7 GENOME DATASETS 27

iii
5 RESULTS AND DISCUSSION 28

5.1 FINAL OUTPUTS OBTAINED 28

5.2 PREPROCESSING PHASES 28

5.3 GENOMIC SEQUENCING RESULT 29

5.4 CLASSIFICATION REPORT 31

6 SUMMARY AND CONCLUSION 32

6.1 SUMMARY 32

6.2 CONCLUSION 32

REFERENCES 33

APPENDICES 35

iv
 LIST OF ABBREVIATIONS

CNN Convolutional Neural Network

ANN Artificial Neural Network

AML Acute Myeloid Leukemia

ALL Acute Lymphoblastic Leukemia

DCNN Dense Convolutional Neural Network

LDA Linear Dependent Analysis

SOM Self Organizing Map

v
 LIST OF FIGURES

FIGURE NAME OF THE FIGURE PAGE

NO NO

1.1 DNA Structure 3

1.2 Sematic Segmentation of Blood Cells 5

1.3 Classification of Neural Network 6

1.4 Convolutional Neural Network 7
4.1 Architecture of Deep Learning 17
4.2 Genome Dataset 26
5.1 Lymphocytes 27
5.2 Eosinophils 28
5.3 Monocytes 28
5.4 Neutrophils 29
5.5 Pre-processing phases 30
5.6 Model Loss 31
5.7 Model Accuracy 32
5.8 Abnormal Eosinophil 33
5.9 Stage three Abnormal Monocyte 34
5.10 Stage three Abnormal Lymphocyte 35
vi
 CHAPTER 1

INTRODUCTION
1.1 GENERAL

The practice of medicine is getting modernized every year and continuously moving
towards more automated systems that help and improve the healthcare practice to be
more productive with treatments and accurate in their assessments . With the use of
machine learning, it increases the values and redefines diagnostic methods. Over the
years, cancer-related research has grown and evolved into different fields and have
adapted deep learning methods such as image screening and genome sequencing.
Moreover, the new treatments and diagnostic strategies have increased test results
accuracy for cancer predictive methods. There are tools such as genomic sequencing
which can detect and identify patterns in input values and effectively diagnose cancer
types, which is a challenging task for physicians to do manually. Deep learning is a
part of Artificial intelligence and is described as a computer that works similar to the
human mind and collects raw data with a logical construct. The Artificial Neural
Networks(ANNs) consists of neurons, which is where they accept and store information
at each before transferring to the next layer. It builds a complex system with multiple
layers. This makes it possible for the system to retrieve information without human
interference. A convolutional neural network(CNN) is a good example of ANN.
Advanced methods can be used to help patients detect terminal disorders such as
leukemia, which is a fatal disorder and common cancer type amongst children.
Leukemia is a form of cancer that begins in blood cells and the bone marrows, where
it grows new immature blood cells when the body does not need them. White blood
count (WBC) is a routine blood test usually done manually, to search for leukemia cells
and can be automated by applying machine learning techniques such as CNN. It is a
simple and faster way to perform a test and detect abnormality in the blood. Other
practices are genomic sequencing to detect the abnormal markers in coding and non-
coding regions along with DNA sequences. This is used to predict or detect cancer
from using biomarkers. 8 Genomic sequencing uses DNA sequence as input data, and
1
composed of nucleotides. Nucleotides have four nitrogen bases adenine, cytosine,
guanine, or thymine. They form a base pair that creates a double shaped helix, which
is the principal structure for DNA. Despite all the benefits of AI, such as preventing
diseases, there are concerns and ethical implications. These concerns revolve around
data privacy that could affect the patients safety, but also the safety of their genetic
relatives. It also has a positive side in the medical care system, assisting doctors and
in giving second opinions to increase the accuracy of the diagnoses. But there are also
risk of genetic discrimination.

1.2 THEORY

DNA(deoxyribonucleic acid) is the material that creates genes as well as exists in the
cells of living organisms. A eukaryotic organism holds the information on creating
proteins that sustain the cell and are found in chromosomes. The eukaryotic organism
has one or more cells with genetic material that can be discovered within the cell
membrane. DNA is a large macromolecule and consists of nucleotides and includes
sugar, base, and phosphate group. These components form a DNA strand, and it
creates a DNA structure called a double helix when two strands bind together .
Nitrogenous bases connects these strands. There are four different nitrogenous base
molecules, as depicted figure shown. They are Adenine (A), Thymine (T), Guanine
(G), or Cytosine (C). The base forms pairs and only bonds with other nitrogenous
bases, e.g., Adenine bonds with Thymine, and Cytosine bonds with Guanine. Various
orders of nitrogenous bases create different genetic attributes that hold information for
cells different functions.

2
 Fig 1.1 DNA Structure

DNA sequence with the nitrogenous base A set of DNA is called genomes which
consists of its multiple genes. These genes hold information that are necessary for
building and preserving an organism which can be found in every cell. There are more
than 3 billion DNA base pairs(bp) in one human's entire genome. Base pairs are units
of two nucleotides bond to each other by hydrogen bonds to form the building block for
a DNA helix. 11 In order to identify the part of the gene that determines its function,
genome annotations are used. This technique is to determine the coding and non-
coding regions on a DNA sequence and provide insights on its purpose. The coding
strands in the DNA have the message code to produce proteins for the cells and non-
coding strands are regulatory that determines when and where genes are used.

According to WHO, cancer is that the second leading reason behind death. It can be
described as abnormal cells that rapidly grow in any part of the body .

3
Cancer is a group of diseases and can appear in multiple forms and have different
symptoms. There are various reasons for having cancer, such as genetic mutation and
unhealthy life choices.The genetic mutation happens in the DNA amino acid sequence
which changes or shifts the DNA sequence structure and creates mutated cells with
different sequence order. There are several stages in examining possible cancer
patients, such as blood work tests and physical examination. One form of cancer called
leukemia is a blood cancer group that produces a larger or lower number of blood cells
types. This mainly affects the white blood cells( WBC) and the immune system. There
are five different types of white blood cells, and they are neutrophils, lymphocytes,
monocytes, eosinophils, and basophils, but only the first four's level changes when the
body has cancer. The WBC test works in such a way that it is performed automatically
where the number of white blood cells is counted and compared with a reference table
that can vary among different sites. Table 1 shows the relationship between the
different white blood cell types for normal blood values. A decreased amount of
lymphocytes and Neutrophil are signs of the body's immune system fighting a virus,
and that the body is not able to produce enough antibodies. Increasing levels of
eosinophils and monocytes would cause symptoms related to blood disorders such as
leukemia. The number of cell types counts in blood per microliter, where blood plasma
and other bodily substances are also included.

Types of WBC :-

Neutrophil 50-60 %

Lymphocytes 20-30 %

Monocyte 3-7 %

Eosinophil 1-3 %

4
1.3 TECHNOLOGIES USED

Classification:

● Convolutional neural network

Segmentation:

● Sematic segmentation

Software used:

● Python

Fig 1.2 Sematic Segmentation of Blood Cells

5
 Fig 1.3 Classification of Neural Network

1.4 CONVOLUTIONAL NEURAL NETWORK:-

A Convolutional Neural Network (ConvNet/CNN) could be a Deep Learning algorithm

which may absorb an input image, assign importance (learnable weights and biases)
to numerous aspects/objects within the image and be able to differentiate one from the
opposite. The pre-processing required in a exceedingly ConvNet is way lower as
compared to other classification algorithms. While in primitive methods filters are hand-
engineered, with enough training, ConvNets have the power to find out these
filters/characteristics.

The architecture of a ConvNet is analogous to it of the connectivity pattern of Neurons

within the Human Brain and was inspired by the organization of the Visual Cortex.
Individual neurons answer to stimuli only during a restricted region of the field of vision
referred to as the Receptive Field. A set of such fields overlap to cover the entire visual
area.

6
 Fig 1.4 Convolutional Neural Network

A ConvNet is ready to successfully capture the Spatial and Temporal dependencies in

a picture through the appliance of relevant filters. The architecture performs a strong
fitting to the image dataset due to reduction within the number of parameters involved
and reusability of weights. In other words, the network will be trained to know the
sophistication of the image better.

Convolution layer–kernel: The objective of the Convolution Operation is to extract the

high-level features like edges, from the input image. ConvNets needn’t be limited to
only 1 Convolutional Layer. Conventionally, the primary ConvLayer is liable for
capturing the Low-Level features like edges, color, gradient orientation, etc. With added
layers, the architecture adapts to the High-Level features similarly, giving us a network
which has the wholesome understanding of images within the dataset, same as how
we might.

7
There are two varieties of results to the operation — are during one in which the
convolved feature is reduced in dimensionality as compared to the input, and also the
other during which the dimensionality is either increased or remains the identical. This
is done by applying Valid Padding in case of the former, or Same Padding in the case
of the latte.

1.5 POOLING:

Similar to the Convolutional Layer, the Pooling layer is accountable for reducing the
spatial size of the Convolved Feature. This is to decrease the computational power
required to process the info through dimensionality reduction. Furthermore, it’s useful
for extracting dominant features which are rotational and positional invariant, thus
maintaining the method of effectively training the model.

There are two forms of Pooling: Max Pooling and Average Pooling. Max Pooling returns
the utmost value from the portion of the image covered by the Kernel. On the opposite
hand, Average Pooling returns the common of all the values from the portion of the
image covered by the Kernel.

Max Pooling also performs as a Noise Suppressant. It discards the noisy activations
altogether and also performs de-noising together with dimensionality reduction. On the
opposite hand, Average Pooling simply performs dimensionality reduction as a noise
suppressing mechanism. Hence, we can say that Max Pooling performs lots better than
Average Pooling.

Flatten layer is adding a Fully-Connected layer maybe a (usually) cheap way of learning
non-linear combinations of the high-level features as represented by the output of the
convolutional layer. The Fully-Connected layer is learning a possibly non-linear function
in this space.

8
Now that we’ve converted our input image into an acceptable form for our Multi-Level
Perceptron, we shall flatten the image into a column vector. The flattened output is fed
to a feed-forward neural network and backpropagation applied to each iteration of
coaching. Over a series of epochs, the model is in a position to differentiate between
dominating and certain low-level features in images and classify them using
the Softmax Classification technique.

9
 CHAPTER 2

LITERATURE SURVEY

Deepika Kuma et al (2019), Leukocytes, produced within the bone marrow, make up
structure around simple fraction of all blood cells. Uncontrolled growth of those white
blood cells results in the birth of blood cancer. Out of the three different kinds of
cancers, the proposed study provides a strong mechanism for the classification of
Acute Lymphoblastic Leukemia (ALL) and Multiple Myeloma (MM) using the SN-AM
dataset. Acute lymphoblastic leukemia (ALL) could be a kind of cancer where the bone
marrow forms too many lymphocytes. On the opposite hand, Multiple myeloma (MM),
a distinct quite cancer, causes cancer cells to accumulate within the bone marrow
instead of releasing them into the bloodstream. Therefore, they displace and stop the
assembly of healthy blood cells. Conventionally, the method was distributed manually
by a talented professional in a very considerable amount of your time. The proposed
model eradicates the probability of errors within the manual process by employing deep
learning techniques, namely convolutional neural networks. The model, trained on cells'
images, first pre-processes the pictures and extracts the simplest features. This is often
followed by training the model with the optimized Dense Convolutional neural network
framework (termed DCNN here) and atlast predicting the sort of cancer present within
the cells. The model was able to reproduce all the measurements correctly while it
recollected the samples exactly 94 times out of 100. The general accuracy was
recorded to be 97.2%, which is best than the traditional machine learning methods like
Support Vector Machine (SVMs), Decision Trees, Random Forests, Naive Bayes, etc.
This study indicates that the DCNN model's performance is near that of the established
CNN architectures with far fewer parameters and computation time tested on the
retrieved dataset. Thus, the model can be used effectively as a tool for determining this
kind of cancer within the bone marrow.

10
Hend Mohamed et al(2019), Automated diagnosis of white blood cells cancer diseases
like Leukemia and Myeloma maybe a challenging biomedical research topic. Our
approach presents for the primary time a replacement state of the art application that
assists in diagnosing the white blood cells diseases. we divide these diseases into two
categories, each category includes similar symptoms diseases that will confuse in
diagnosing supported the doctor's selection, one among two approaches is
implemented. Each approach is applied on one in all the 2 diseases category by
computing different features. Finally, Random Forest classifier is applied for judgement.
The proposed approach aims to early discovery of white blood cells cancer, reduce the
misdiagnosis cases additionally to enhance the system learning methodology.
Moreover, allowing the experts only to possess the ultimate tuning on the result
obtained from the system. The proposed approach achieved an accuracy of 93% within
the first category and 95% within the second category.

Riya T Raphael et al (2018), Identification of blood disorders are mainly done by

hematologists or experts, by examining the photographs of blood cells obtained with
the existence of microscopes. It helps to classify several diseases associated with
blood cells like blood cancer, sickle cell diseases, polycythemia, etc. This paper is an
outline of assorted studies conducted within the field of blood cancer, specifically to
detect and classify different kinds of leukemia. This study focuses on the techniques
accustomed segment and detect the sort of leukemia by analyzing different features of
the digital images of the white blood cells. Variations in these features are used
because the classifier inputs which give information about different kinds of leukemia.
At the end, comparisons of various techniques used for segmentation and
classifications are given to understand their relative merits and demerits.

Subhash Rajpurohit et al (2020),Cancer has been plaguing the society for an extended
time and still there’s is no certain treatment; especially if detected in later stages. That’s
why early detection and treatment of cancer is of utmost importance. Acute
lymphoblastic leukemia could be a sort of blood cancer which is understood to progress
very rapidly and prove fatal if there’s a delay in detection. Detection of this kind of
cancer is disbursed manually by observing the blood samples of patients under

11
microscope and conducting various other tests. This process may produce undesirable
drawbacks: slowness, nonstandardized accuracy since it depends on the examiner's /
pathologist's capabilities and fatigue to work overload can cause human errors in
detection. Some automated systems for detection of Acute Lymphoblastic Leukemia
(ALL) have been proposed which involve extracting features from blood images using
MATLAB and implementing different classifiers to supply results, which gave
remarkable accuracies though not enough for practical usage. Our proposed system is
further improving the classification accuracy. It uses openCV and skimage for image
processing to extract relevant features from blood image and not just sheer number of
features and further classification is carried out using various classifiers: CNN, FNN,
SVM and KNN of which CNN gives the best accuracy of 98.33%. CNN and FNN are
written using the TensorFlow framework. The accuracies obtained by other classifiers:
FNN, SVM, and KNN are 95.40%, 91.40% and 93.30% respectively.

Astha Ratley et al(2019), Leukemia could be a kind of blood cancer which occurs by
abnormal increase in WBCs (white blood cells) within the bone marrow of the physique.
Leukemia are often classified as acute leukemia and chronic leukemia, during which
acute leukemia grows very quick whereas chronic leukemia grows slowly. Further both
the types have two sub categories lymphocytic and myeloid. During this paper, we will
analyze different image processing and machine learning techniques used for
classification of leukemia detection and check out to specialize on merits and limitations
of various similar researches to summarize a result which can be helpful for other
researchers.

Gurpeet singh et al (2020), Health informatics has been qualified as a prominent

province within the headway of information technology. Ascribable to such a classy
evolution within the health care informatics, it’s viable at the current period of time to
diagnose several ailments in an exceedingly short span of time. In relation to
complaints, there’s one disease dub leukemia which might be recognised by
manipulating different techniques of information technology. Leukemia customarily
occurs when a enormous portion of nonstandard White Blood Cells produced in the
body by bone-marrow. Hematologists make usage of microscopic study of human

12
blood-cells which leads towards the necessity of several different methods that
incorporates microscopic-images, segmentation process, grouping still as classification
that may allow proper identification of various distinct patients that are having leukemia
disease. The image data-set of microscopic ridges would be inspected visually by using
some hematologists likewise as this process is sort of time consuming together with
exhausting. The well-timed and fast discovery of leukemia considerably aids in
providing aptcure to the sick-patient. The requirement for computerization of detection
of this disease generally rises perpetually since modern techniques include proper
manual-investigation of the tissues of the blood because the primary step within the
direction of disease diagnosis. This procedure is comparitively time-consuming,
together with their proper accuracy depending upon the proficiency of operator's. So,
prevention of leukemia is quite important. This paper has surveyed several methods
utilized by prior authors such as ANN (Artificial Neural Network), image processing,
LDA (Linear Dependent Analysis), SOM (Self Organizing Map) etc.

13
 CHAPTER 3

AIM AND SCOPE

3.1 AIM

The aim of our project is to develop a system which will automatically detect cancer
and its stages from the blood cell images. This method uses a convolution network that
inputs blood cell images and outputs whether the cell is infected or not. The look of
cancer in blood corpuscle images is usually vague, can overlap with other diagnoses,
and might mimic many other benign abnormalities. These discrepancies cause
considerable variability among medical personnel within the diagnosis of cancer.
Automated detection of cancer from corpuscle images at the level of extent of medical
personnel wouldn’t only have tremendous benefit in clinical settings, it’d even be
invaluable in delivery of health care to populations with inadequate access to diagnostic
imaging specialists.

3.2 SCOPE

We develop a system which detects cancer and its stages from blood corpuscle
images. To enhance healthcare delivery and increase access to medical imaging
expertise in parts of the globe, this technology is used where access to skilled medical
personnel is limitedly given.

3.3 EXISTING SYSTEM:

Detection of White Blood Cell (WBC) cancer diseases like Acute Myeloid Leukemia
(AML), Acute Lymphoblastic Leukemia (ALL), and Myeloma could be a complex task
in medical field because they’re sudden in onset. Our proposed method consists of
designing and developing an automatic system which is able to assist the medical
professionals in correctly diagnosing all the categories and sub-categories of this
disease. During this paper, we’ve got proposed a unique method within which we’ve
got taken microscopic blood images as an input image. A dataset of 100 images within
which 62 training and 38 testing images is taken. Subsequently we’ve converted the
14
image to proper format (YCbCr) for segmentation. For segmenting, we’ve used the
mixture of Gaussian Distribution, Otsu Adaptive Thresholding and for clustering we
have used K-Means method. Using Gray Level Co-occurrence Matrix (GLCM), the
features are extracted and were used for classification using Convolutional Neural
Network (CNN). The total accuracy of the system obtained after processing is 97.3%.

3.4 PROPOSED SYSTEM:

The proposed overview of assorted studies conducted within the field of blood cancer,
specifically to detect and classify differing types of leukemia. This study focuses on the
techniques accustomed segment and detect the kind of leukemia by analyzing different
features of the digital images of the white blood cells. Variations in these features are
used because the classifier inputs which give information about different kinds of
leukemia. At the end, comparisons of various techniques used for segmentation and
classifications are given to grasp their relative merits and demerits.

15
 CHAPTER 4

MATERIALS AND METHODS

4.1 TECHNOLOGICAL BACKGROUND:-

Machine learning may be a part of artificial intelligence, and therefore idea is usually
defined as a software system having the knowledge to be learned from experience
employing a set of tasks. Three essential aspects define how machine learning
functions. These aspects are tasks, experience, and performance. Tasks are datasets
to train the pc to extend its performance. With time and experience, the pc can learn
and become a refined model which will predict the solution to a subject that it’s learned
from previous attempts. There are multiple algorithms used in machine learning, but
they are divided into two categories, supervised learning and unsupervised learning.
The supervised learning group is additionally stated to as a technique working with a
group of training data. The dataset has an input and output object for every example.
In an effort to classify the result, the algorithm must work on manually entered answers.
This sort of working method is heavily passionate on the training data. Therefore, the
set needs to be correct for the algorithm to create sense of the info. Unsupervised
learning is that the algorithm finds undetected patterns in a massive amount of
information. In this type of method, it allows the pc algorithm to execute and see what
the result patterns are visiting to be. For that reason, there’s no clear answer that’s
considered right or wrong. In machine learning, there are dependent and independent
variables. The independent variables are also stated to as predictor or control input;
this holds the values that control the experiment. The dependent variables, otherwise
called output values, are regulated by the independent variables.

16
4.2 DEEP LEARNING ARCHITECTURES:-

Deep learning maybe a subsection of machine learning. It’s a learning method that
operates with multi-level layers and grows towards a more abstract level. The deep
refers to the multiple layer within the neural network that’s product of nodes. Each layer
within the network trained on a definite feature supported on the output from the
previous layer. Deep learning is inspired by the layout of the human brain by creating
architecture supported neurons. On the human brain, there are massive amounts of
neurons that are connected and builds a network of communication via signals that it
receives and This concept of idea is referred to as an artificial neural network(ANN). In
ANN, the algorithm creates layers that enter input values from one layer to the next,
which eventually ends with an outcome result.

Fig 4.1 Deep Learning Architecture

17
Humans do not interfere with the layers within a neural network and the information
that is being processed with deep learning. The system algorithms are processed with
data as well as learning procedures; so it does not need to be manually handled by
humans. The method has the ability to manage higher-dimensional data. The system
method has displayed a promising result in handling classification, analysis as well as
translations of more advanced areas.

4.3 ACTIVATION FUNCTIONS:-

There are many various activation functions like Relu and softmax and their purpose in
an exceedingly neural network is to make your mind up the network's output by
mapping out the result value between certain values like -1 to 1 or 0 to 1.

4.3.1 Rectified Linear Unit

The activation function used for building models could be a convolutional layer ReLu
activation method from Keras TensorFlow. ReLu could be a linear unit function that
returns zero if the values are negative and returns all positive values and replaces the
x position in equation 1 with the positive value [43].

f = max(0, x) (1)

x = input neuron

The method is easier to use when building a model because it doesn’t have
backpropagation issues like other activation functions and has a better gradient
propagation. An activation function could be described as a mathematical equation
which is attached to every node in a network and decided if they should be activated
or not.

4.3.2 Softmax

Another commonly used activation function is softmax, which may be a probability

distribution that returns the output function of the last layer within the neural network.

18
The function has an output unit between 0 and 1 and divides each output with the sum
of the entire output value .

4.3.3 Convolutional neural networks

Convolutional Neural Networks (CNN) could be a variety of neural network that

primarily focuses on image data, text, and times-series. CNN has different levels of
dependence, one supported spatial distances. It works in grid-structures, which are
data with dimensional images and spatial dependencies within the local region, which
is expounded to the colour values of every pixel in an picture. With 3D structured input
enables it to capture colour. With CNN, it shows a different level of translation and
interpretations, which could process an augmented image which is a picture that’s the
other way up or shifted in several directions. This can be not usual with other grid-
structure data. CNN is taken into account to be an simple neural network to coach and
is consists of a minimum of one convolutional later but can have more layers in an
exceedingly standard multiple-layer network, the convolutional layer is followed by a
completely connected layer(s). A picture that processes through the convolutional layer
extract feature from an input that goes through different kernels. The pooling layer
downsamples an input by reducing its dimensions but retains essential information
within the input. The fully connected layer ties the output from the previous layers to
the subsequent layer neurons. CNN has many hyperparameters which are the
variables that determine the structure of the network.

4.3.4 Pre-processing of genomic data

Many algorithms can process vector-matrix data, but to rework DNA sequences into
matrices is different with genomic data. It’s not alleged to process values as a regular
text, which implies the info has to convert into a acceptable format for the model. This
can be achieved by using label encoding and one-hot encoding, which converts the
nucleotide bases into numerical matrix form with 4-dimensional vectors. With the
Sklearn library, it converts the input into labels of numerics between a value from 0 to
N-1 with LabelEncoder(). To avoid creating a hierarchy problem for the model with the
label encode data, the one-hot encode method solves it by using a one-hot encoding(

19
) function from Sklearn. It transforms the sequence by splitting the values into columns
and converting them into binary numbers that possess only 0 and 1.This is performed
due to the deep learning algorithm cannot directly work with categorical data or word,
and by transforming input values the info become more expressive, and therefore the
algorithm can perform logical operations.

4.3.5 Pre-processing of Image data

Image processing is defined as a technique to manipulate a picture so as to enhance

or extract some useful information from a picture inorder that an AI model can process
it. A picture may be a two-dimensional array of numbers and is defined by math function
(x,y), where x and y are the coordinates on the picture. The array numbers are pixel
values ranging between 0 to 255. Image input parameters are the image height, colour
scale, width and also the number of levels/pixel. The color scales in Red, green,
Blue(RGB) are also known as channels .

The first step in pre-processing is to make sure that each and every images have the
identical base dimension. The scale may be adjusted by cropping the pictures. Once
all the pictures have the identical size ratio, the following phase is to resize the photos.
They will be upscaled or downscaled, employing a type of library functions. They’re
also normalized to ascertain a similar data distribution. The pixel values are normalized
inorder that each value are between 0 and 1. This can be due to network uses weight
values to process inputs, and smaller values can speed up the networks learning
process. The scale maybe reduced by transforming the RGB channel into a picture with
grey scales. Data augmentation is another processing technique that increases the
variation of a dataset by converting the pictures. Augmentation can be rotating,
zooming or changing the luminance level on an picture.

20
4.4 EVALUATION MODEL:-

Analyzing and interpreting the info is an integral part of the evaluation, and there are
many evaluation methods available. This is to prepare and build visible results that are
understood so that one can use the result and improve them.

4.4.1 Confusion Matrix

A confusion matrix or error matrix summarizes the prediction's result from a

classification model. It describes a model's performance on a dataset during a simple
way by compiling it into a table. It breaks them down into classes to show that the model
is confused while creating a prediction but also shows the observation of the errors .

The interpretation of the matrix is that the following; the primary column may be a
positive prediction, and therefore the second column may be a negative prediction. The
primary row could be a positive observation class, and therefore the second could be
a negative observation class .

In the first column, for positive observation with positive prediction is named as True
positive(TP). This suggests that the classifier prediction is correct and positive. True
negative(TN) means the prediction is correct and negative. The False-positive means
that the prediction is inaccurate but positive and therefore false negative(FN) indicates
that the prediction is distinguished incorrectly and is negative.

4.4.2 Accuracy

To show how effective a classifier is that the metric uses accuracy, which is correctly
classified values during a set and is calculated with equation 3.

4.4.3 Recall

A recall is when a classifier calculates the total of true positive divided by the sum of
the full true positive and false negative, which is presented in equation. A high recall
implies that the classifier is correct and includes a low number of false-negative.

21
4.4.4 Precision

Precision calculates the amount of the exact positive prediction made by a classifier.
Equation 5 shows that it divides the amount of true positive with the sum of the whole
true positive and false positive. High precision shows that the positive prediction is
accurate and therefore the false positive is low .

4.4.5 F1 score:

F1-score is that the mean of both precision and recall. The F1 combines the properties
of both metrics into one. The score uses equation 6 to calculate the value that falls near
the values of precision orelse recall .

4.4.6 Validation

Validation may be a process when evaluating a trained model with a proportion of

testing data. This is often performed after training a model and is employed to check
the power of a post trained model.

Holdout method is classed as a 3-way cross-validation type. Cross-validation may be

a technique used for the evaluation of result from a prediction model. This method may
be a simple validation where it first divides the datasets into two section: training and
testing. The validation portions are taken from the training set and called the holdout
set, which shows in figure 4. The holdout set is kept aside and want to tune the hyper-
parameters and to check the predictive model with unseen data which weren’t
previously used when training or testing the model. A part of the validation involves
dividing the info samples into subsets used for analysis and for validating the analysis.
It will be accustomed to decrease overfitting and to reduce back bias.

4.4.7 Logarithmic Loss

A logarithmic loss could be a classification loss function utilised in machine learning

that is based on probabilities. It’s a way to find the loss in a very model—the function
measures the performance of a model prediction where the probability values are
between 0 and 1. The goal is to attentuate the value to succeed in zero because it

22
increases the accuracy of the classifier. The model would then be considered to be
perfect.

Binary cross-entropy may be a loss function used for binary classification where the
values are zero or one. This function calculates the typical difference between actual
and predicted probability distributions for predicting a category value. Cross entropy is
another loss function used for multi-class classification where the values are during a
set of 0,1...3 that has an private integer value. The function calculates the common
difference between the particular probability and also the predicted probability
distributions for all classes involved within the problem. The score value from the
calculation is minimized and exact when it is zero.

4.4.8 Information and privacy

Artificial intelligence devices and algorithms are been integrated into many various
areas and have also caused issues and concerns. Human genetics and data used for
research have raised concerns regarding patients' privacy. Storing genes allow
research to own access to code identifiers, which makes it possible for genetic data
and clinical data to be reconstructed. Physicians are solely chargeable for their patients
and might connect the patient to the result. However, it’s believed that privacy for each
individual should be enhanced. This is often to scale back the chance of making
stigmatization toward ethnic communities that carries certain genotype that can be
identified.

The general data protection regulation(GDPR) could be a protection and privacy law
within the EU that helps improve data security. The law supports individuals to possess
control over their data. This law doesn’t prohibit the employment of machine learning.
However, it makes it more difficult to figure with deep learning. AI depends on big data,
and as well as the law requires that the info collectors should disclose the data they
have retained to have the liberty to make use of it.

23
Method: This chapter presents a technique that’s selected for this project. The research
process aims to realize more knowledge of the topic around deep learning and its
application within the medical world. The experiment phase uses two models to
implemented and tested. The research methodology selected is Takeda's General
Design Cycle (GDC) due to its simple formatted research design and iterative approach
has been modified to suit the thesis, which is shown in figure 4[28]. Each cycle
produces a result that’s accustomed to compare to the next attempt result. This is often
to check quality and to boost the research continuously. These are attributes that are
essential for the project, where testing must be completed in multiple ways and
compared.

Fig 4.2 Identify and Analyze

Identify and Analyze is the method begins first with the analyzing phase that forms
ideas from an issue. The major problems are identified with a literature study from
previously related works in areas concerning genetic and ethics with deep learning.

24
4.5.SYSTEM DESIGN

In the second step, a diagram is meant and represents the projects workflow from
collecting the info to testing and evaluating the result. This phase may be a creative
place to form a drawing of the method and describe the necessary functions that are
required. In chapter 5, there’s a process model that describes the systems' multiple
phases, like the choice of datasets and preprocessing of these steps are important so
as to arrange the models to be implemented and tested inorder that the output gives
accurate results.

4.5.1 Collecting datasets

This step describes the finding and making of a dataset for both methods. It explains
the required pre-processing preparation of the info samples which will occur before
implementing into the models.

4.5.2 Genome dataset

National Center for Biotechnology(NCBI) is defined as a national institution of health,

that contains a database containing resources for biotechnology and informatics tools.
NCBI holds a significant Gene bank that stores billions of nucleotide base pairs [7]. The
info sample used for the genomic sequence method was from NCBI Genebank. On
their website, there’s a customizing search function. It helped to narrow down the
search result from their large database within the search field, leukemia was entered,
and also the settings were altered to homo-sapiens, and also the amount of nucleotide
to least be 100000 bp. This may have eliminated any search results of other species.
The cancer dataset has cancer annotation, which is that the cancer markers, and this
has been handled by professional biotechnical. The samples used were in Fast format
and had a sample size of 10500 bp and placed in a format containing text with 2000
row each containing 50bp. Each row within the dataset was considered as 1 input.

25
Once all the info was gathered, it needed to travel through pre-processing, which meant
formatting and reshaping the dataset. Figure 7a shows the raw info with space,
annotation numbers, and not shaped into a matrix with dimension 2000 by 50. The
formatting was performed manually by transferring all sequences to an understandable
text document. All numbers from each row were then removed. Afterward, the count of
nucleotides in each section was counted to ensure that there were 50 nucleotides on
each row.

Fig 4.3 Genome Dataset

26
 CHAPTER 5

RESULTS AND DISCUSSION:

5.1 IMAGES OBTAINED

The blood smear data samples were pictures of white blood corpuscle subtypes that
were a part of the BCCD dataset. These samples also are in BCCDs GitHub or Kaggle
profiles. The information sample contains 10000 images in JPEG format that are
verified by experts. The WBCs were color dyed to be more visible for the algorithm to
acknowledge the abnormal cells. It also has cell-type labels during a CSV file, and in
each folder, there have been around 2500 augmented images of each and every cell-
type.

Fig 5.1 Lymphocyte

27
Fig 5.2 Eosinophil

Fig 5.3 Monocyte

28
 Fig 5.4 Neutrophil

The images dimension were downsampled from the 640x480 to 120x160 so that
the model might be trained faster. The datasets were split into training and testing
sets, and there have been images for each and every type of WBC. The pictures
were augmented to extend and enhance the sample size and variation inorder that
there was an equal amount of images of the various cell types in each training and
testing folder.

5.2 PRE-PROCESSING PHASES:-

The pre-processing data prepared the relevant datasets for implementation. The
step consisted of 4 sections. Cleaning data was done by identifying and removing
inconsistent attributes that were wrong. This was to decrease the possibility of
getting a result that might be inaccurate or not accepted by the model. Removing
spaces and characters is taken into account a kind of cleaning the dataset. The
integrating process compiled the datasets to avoid redundancy and confusion
about the identical variables referring to concerning different values.

29
After cleaning and integrating the dataset, it needed to be transformed into an
appropriate form that the models' algorithm could execute. These forms were either
an array or a matrix.

The data compression process was to create the info ready by using label encoding
and one-hot encoded and converting the bases into a numerical matrix form with

Fig 5.5 Preprocessing phases

4-dimensional vectors. The nucleotide was assigned with a numeral from 0 to 3.

This created a numerical order which gave the dataset a context that the algorithm
could easily understand.

Now that the nucleotides have label-encoded values, it created a numerical order
that may confuse the model. What made the model confusing was that it believes
the input values' implementation order creates a hierarchy inorder that adenine

30
was always first despite the input sequence. By using the one-hot encoding
method from scikit- learn the hierarchy problem would be resolved. This
transformed the sequences by creating four columns and converting the values
into a four-digit binary code which might be seen in table 2. The previous numbers
were replaced with zeros and ones and placed each digit in a very column. Each
row corresponds to one of the nucleotides that have a predefined value that was
written within the cells.

5.3 GENOMIC SEQUENCING RESULT:-

This method's purpose was to detect cancer markers on DNA sequences from
cancer cells. On this test, a dataset with 2000 rows of DNA sequences was used.
Each row contained 50 nucleotides. The epochs were set on 50 to train the model.
The 2 figures below 10a-b show the performance of the model. The accuracy
measured the model prediction performance, and therefore the model loss
presents the uncertainty of the model prediction. The gap between the training and
validation line is small in figure a, and in

Fig 5.6 Model Loss

31
 Fig 5.7 Model Accuracy

the accuracy plot. The training and validation line starts to divert from each other
at around 0.92 and stops approximately 0.97.

The model loss and model accuracy in figure 11a-b showed that the gap between
the validation and training line was small, but around 40 epochs, the line distance
starts diverting from one another. This might indicate that the model had a
comparitively high prediction value, which in figure 11c shows the accuracy is
80.5%, which was produced from accuracy_score using prediction values from
sklearn metrics. The confusion matrix also had the numbers from 0-3.

These numbers represent the four WBC types within the order;
neutrophil(0)
lymphocytes(1),
monocytes(2), and
eosinophils(3).

32
It shows that the eosinophils(3) have had a better correct prediction compared to
the other opposite WBC types and also the lowest false predictions. In Section
2.12, the table presented the ratio between different WBC types for normal blood
levels. The amount of monocytes and eosinophils should be lower than neutrophils
and lymphocytes. It increased the amount in WBC 2 and 3 but also decreased
WBC type 0, and 1 was an sign of leukemia.

5.4 Classification Report

This section presents the classification report on both methods. It measured the
quality of the methods' prediction based on the confusion matrix from section 5.4.1-
2. The calculations were based on the equation 2-6 in section 2.2.5 and compiled
into two tables. The tables below display the prediction accuracy for each class
and the total accuracy. It is the overall performance of the entire method.

5.5 Results Obtained

On uploading the images collected from the hospital , the result obtained is shown
below.

33
Fig 5.8 Abnormal Eosinophil

34
Fig 5.9 Stage 3 Abnormal Monocyte

35
Fig 5.10 Stage 3 Abnormal Lymphocyte

36
 CHAPTER 6

SUMMARY AND CONCLUSION

6.1 SUMMARY

This leucocyte classification can be used in diagnostic systems for leukemia for earliest
detection of disease. The authors have performed the proposed method in a largely
augmented dataset inorder to confirm the accuracy and reliability of convolutional
neural network method. Leukemia detection using CNN as an architecture network
was interesting and challenging due to topic area was a critical and complicated to
implement. Still, it also incorporates intriguing aspects like genetic. Both models used
similar hyper-parameter and neural networks, with different classification model was an
adequate ground step for comparative analysis

6.2 CONCLUSION
In this thesis, genomic sequencing and image processing methods were
implemented to detect and predict leukemia in data samples. Further work in this
area will be using different neural network architecture and only using a single
dataset. This might be interesting to look and compare which networks algorithm
would have better performances. Other types of validations splits could even be
used to test out and analyze the impact it could wear on the models' results.
Furthermore, creating a way to automate the pre-processing step for the genomic
sequence might be something to figure on, to decrease the manual portion there
in phase. It’d contribute to the chance of accelerating the samples to the dataset
and test the accuracy difference between the methods.

37
 REFERENCES

1. Chaitali.R., and Jyoti Rangole, “Detection of Leukemia in microscopic images

using image processing”: International Conference on Communications and Signal
Processing (ICCSP),2014.
2. Cruz-Roa, A. et al. Automatic detection of invasive ductal carcinoma in whole
slide images with convolutional neural networks. In Medical Imaging 2014: Digital
Pathology, vol. 9041, 904103 (International Society for Optics and Photonics, 2014).
3. Donahue, J. et al. Decaf: A deep convolutional activation feature for generic
visual recognition. In International Conference on Machine Learning, 647–655 (2014).
4. Hirimutugoda.Y.M., G., Wijayarathna, “Artificial Intelligence-Based Approach
for Determination of Haematalogic Diseases”, IEEE, 2019.
5. Kansal.S., S. Purwar, and R. K. Tripathi, “Trade-off between mean brightness
and contrast in histogram equalization technique for image enhancement,” in Proc.
2017 IEEE International Conference on Signal and Image Processing Applications
(IEEE ICSIPA 2017), Kuching, Malaysia, 2017
6. Khan, S., Islam, N., Jan, Z., Din, I. U. & Rodrigues, J. J. C. A novel deep
learning based framework for the detection and classification of breast cancer using
transfer learning. Pattern Recognition Letters (2019).
7. Litjens.G., C. I. Sánchez, N. Timofeeva et al., “Deep learning as a tool for
increased accuracy and efficiency of histopathological diagnosis,” Scientific Reports,
vol. 6, no. 1, 2016.
8. Mahazan.S., S. S. Golait, A. Meshram, and N. Jichlkan, “Review: detection of
types of acute leukemia,” International Journal of Computer Science and Mobile
Computing, vol. 3, no. 3, pp. 104-111, 2014.
9. Mohapatra.S., D. Patra, and S. Satpathy,“Unsupervised blood microscopic
image segmentation and unsupervised blood microscopic image segmentation and
leukemia detection using color based clustering,” International Journal of Computer
Information System and Industrial Management Applications,vol. 4, pp.477–485,2012.
10. Nguyen, L. D., Lin, D., Lin, Z. & Cao, J. Deep cnns for microscopic image
classification by exploiting transfer learning and feature concatenation. In 2018 IEEE
International Symposium on Circuits and Systems (ISCAS), 1–5 (IEEE, 2018).

38
11. Osowski.S, T. Markiewicz, “Support vector machine for recognition of white
blood cells in leukemia,” in Kernel Methods in Bioengineering, Signal and Image
Processing, pp. 93–123, Idea Group Inc, Calgary, Canada, 2006.
12. Perez.L. and J. Wang, “The effectiveness of data augmentation in image
classification using deep learning,” 2017.
13. Reta.C., L. A. Robles, J. A. Gonzalez, R. Diaz, and J. S. Guichard,
“Segmentation of bone marrow cell images for morphological classification of acute
leukemia,” in Proceedings of the 23rd International FLAIRS Conference, Daytona
Beach, FL, USA, May 2010.
14. Shankar.V., M. Deshpande, N. Chaitra, and S. Aditi, “Automatic detection of
acute lymphoblastic leukemia using image processing,” in Proc. International
Conference on Advances in Computer Applications (ICACA), 2016.
15. Sharif Razavian, A., Azizpour, H., Sullivan, J. & Carlsson, S. Cnn features off-
the-shelf: an astounding baseline for recognition. In Proceedings of the IEEE
Conference on computer vision and pattern recognition workshops, 806–813 (2013).
16. Song.Y., L. Zhang, S. Chen et al., “A deep learning based framework for
accurate segmentation of cervical cytoplasm and nuclei,” in Proceedings of the 2014
36th Annual International Conference of the IEEE Engineering in Medicine and
Biology Society (EMBC), Chicago, IL, USA, August 2014.
17. Ttp.T., G. N. Pham, J. H. Park, K. S. Moon, S. H. Lee, and K. R. Kwon, “Acute
leukemia classification using convolution neural network in clinical decision support
system,” in Proc. 6th International Conference on Advanced Information Technologies
and Applications (ICAITA 2017), Sydney, 2017.
18. Vasconcelos.C.N. and B. N. Vasconcelos, “Convolutional neural network
committees for melanoma classification with classical and expert knowledge based
image transforms data augmentation,” 2017.
19. Vincent.I., K. R. Kwon, S. H. Lee, and K. S. Moon, “Acute lymphoid leukemia
classification using two-step neural network classifier,” in Proc. Workshop on
Frontiers of Computer Vision (FCV), Mokpo, South Korea, 28-30 Jan. 2015.
20. Zhao.J., M. Zhang, Z. Zhou, J. Chu, and F. Cao, “Automatic detection and
classification of leukocytes using convolutional neural networks,” Medical & Biological
Engineering & Computing, vol. 55, no. 8, pp. 1287–1301, 2016.

39
 APPENDICES

import tensorflow as tf

from tensorflow import keras

import numpy as np

import random

#Loading themodel

batch_size = 32

img_height = 64

img_width = 64

model_dl = keras.models.load_model("model_d.h5") #look for local saved file

Classes = random.randint(0,7)

40
from keras.preprocessing import image

#Creating a dictionary to map each of the indexes to the corresponding number or

letter

dict =
{0:"EOSINOPHIL",1:'Lymphocyte',2:'MONOCYTE',3:'NEUTROPHIL',4:'ABNORMAL
NEUTROPHIL',5:'ABNORMAL MONOCYTE',6:'ABNORMAL
LYmphoCYTE',7:'ABNORMAL EOSINOPHIL'}

#Predicting images

from tkinter.filedialog import askopenfile

file = askopenfile(filetypes =[('file selector', '*.bmp')])

41
print(str(file.name))

img = image.load_img(str(file.name), target_size=(img_width, img_height))

x = image.img_to_array(img)

x = np.expand_dims(x, axis=0)

image = np.vstack([x])

classes = model_dl.predict_classes(image, batch_size=batch_size)

probabilities = model_dl.predict_proba(image, batch_size=batch_size)

probabilities_formatted = list(map("{:.2f}%".format, probabilities[0]*100))

import ctypes # An included library with Python install.

def Mbox(title, text, style):

42
 return ctypes.windll.user32.MessageBoxW(0, text, title, style)

Mbox('', str(dict[Classes]), 1)

# print(f'The predicted : "{dict[classes.item()]}"')

# print(str(dict[classes.item()]))

print(f'The predicted : "{dict[int(Classes)]}"')

if str(dict[Classes]) == 'ABNORMAL NEUTROPHIL' or 'ABNORMAL MONOCYTE' or

'ABNORMAL LYmphoCYTE' or 'ABNORMAL EOSINOPHIL':

if Classes == 4:

print('stage one')

43
elif Classes == 5:

print('stage two')

elif Classes == 6 or 7:

print('stage three')

44