Automated Detection of Acute Lymphoblastic

Leukemia Subtypes from Microscopic Blood Smear

Images using Deep Neural Networks

Submitted By
Md. Taufiqul Haque Khan Tusar
181472099
Roban Khan Anik
182482057

A project report submitted in partial fulfillment of the requirements for the

Degree of Bachelor of Science in Computer Science and Engineering

Supervised By
Md. Touhidul Islam
Lecturer
Department of CSE
City University

CITY UNIVERSITY
DHAKA, BANGLADESH
JULY 2022
 DECLARATION

We declare that this project report titled “Automated Detection of Acute Lymphoblastic
Leukemia Subtypes from Microscopic Blood Smear Images using Deep Neural
Networks” is the result of our own research except as cited in the references. This project
has been done by us under the supervision of Md. Touhidul Islam. This project is the partial
fulfillment of requirements for the degree awarded as Bachelor of Computer Science and
Engineering during the session of 2018-2022 in City University, Dhaka, Bangladesh.

Submitted By

---------------------------
Md. Taufiqul Haque Khan Tusar
ID: 181472099
Batch: 47
Dept. of CSE
City University,
Dhaka, Bangladesh

Approved By

-------------------------------
Supervisor

Md. Touhidul Islam

Lecturer
Department of Computer Science and Engineering (CSE)
City University, Dhaka, Bangladesh

i
 DECLARATION

We declare that this project report titled “Automated Detection of Acute Lymphoblastic
Leukemia Subtypes from Microscopic Blood Smear Images using Deep Neural
Networks” is the result of our own research except as cited in the references. This project
has been done by us under the supervision of Md. Touhidul Islam. This project is the partial
fulfillment of requirements for the degree awarded as Bachelor of Computer Science and
Engineering during the session of 2018-2022 in City University, Dhaka, Bangladesh.

Submitted By

---------------------------
Roban Khan Anik
ID:182482057
Batch: 48th
Dept. of CSE
City University,
Dhaka, Bangladesh

Approved By

-------------------------------
Supervisor

Md. Touhidul Islam

Lecturer
Department of Computer Science and Engineering (CSE)
City University, Dhaka, Bangladesh

ii
 ACKNOWLEDGEMENT

First, we express our heartiest thanks and gratefulness to almighty Allah for his divine
blessing in making us possible to complete this project successfully.

We would like to thank the following people for their help in the production of this project.
We are deeply indebted to our supervisor Md. Touhidul Islam, Lecturer, Department of
CSE. Without his help and support, throughout this project, it would not have been possible.
His endless patience, scholarly guidance, continual encouragement, constant and energetic
supervision, constructive criticism, valuable advice, and reading many inferior drafts and
correcting them at all stages have made it possible to complete this project.

Our special thanks go to the Head of the Department of CSE, Md. Safaet Hossain, who had
permitted us and encouraged us to go ahead. We are bound to the Honorable Dean of the
Faculty of Science and Engineering, Prof. Dr. Engr. Md. Huamaun Kabir, for his endless
support.

I am very grateful to all my faculty teachers who gave me their valuable guides to complete
my graduation. I am also very grateful to all those people who have helped me to complete
my project.

Finally, I must acknowledge with due respect the constant support and patients of my
parents.

Md. Taufiqul Haque Khan Tusar

ID: 181472099
Batch: 47
Dept. of CSE
City University,
Dhaka, Bangladesh

iii
 ACKNOWLEDGEMENT

First, we express our heartiest thanks and gratefulness to almighty Allah for His divine
blessing in making us possible to complete this project successfully.
We would like to thank the following people for their help in the production of this project.
We are deeply indebted to our supervisor Md. Touhidul Islam, Lecturer, Department of
CSE. Without his help and support, throughout this project, it would not have been possible.
His endless patience, scholarly guidance, continual encouragement, constant and energetic
supervision, constructive criticism, valuable advice, and reading many inferior drafts and
correcting them at all stages have made it possible to complete this project.

Our special thanks go to the Head of the Department of CSE, Md. Safaet Hossain, who had
permitted us and encouraged us to go ahead. We are bound to the honorable Dean of the
Faculty of Science and Engineering, Prof. Dr. Engr. Md. Huamaun Kabir, for his endless
support.

I am very grateful to all my faculty teachers who gave me their valuable guides to complete
my graduation. I am also very grateful to all those people who have helped me to complete
my project.

Finally, I must acknowledge with due respect the constant support and patients of my
parents.

Roban Khan Anik

ID:182482057
Batch: 48
Dept. of CSE
City University,
Dhaka, Bangladesh

iv
 Abstract
An estimated 300,000 new cases of leukemia are diagnosed each year which is 2.8 percent
of all new cancer cases and the prevalence is rising day by day. The most dangerous and
deadly type of leukemia is acute lymphoblastic leukemia (ALL) which affects people of all
age groups including both children and adults. In this study, we propose an automated
system to detect various-shaped ALL blast cells from microscopic blood smears images
using Deep Neural Networks (DNN). The system can detect multiple subtypes of ALL cells
with an accuracy of 98℅. Moreover, we have developed a telediagnosis software to provide
real-time support to diagnose ALL subtypes from microscopic blood smears images.

Keywords - ALL, Deep Neural Network, Digital Image Processing, Telediagnosis.

v
 Table of Contents

DECLARATION i
ACKNOWLEDGEMENT iii
Abstract v
Chapter 1 1
Introduction 1
1.1 Composition of Blood 1
1.2 Problem Statement 1
1.3 Prevalence of Leukemia 1
1.4 Difficulties in Existing Manual Diagnosis 2
1.5 Challenges in Automated Diagnosis using DNN 2
1.6 Proposed Approach 2
1.7 Contributions of the Proposed Approach 2
1.8 Organization of the Study 3
Chapter 2 4
Literature Review 4
Chapter 3 7
Dataset Description 7
Chapter 4 10
Proposed Method 10
4.1 Overview 10
4.2 Data Preparation Pipeline 10
4.3 Deep Neural Network Architecture 11
4.3.1 Applied Convolutional Neural Network (ConvNet) 11
4.3.2 Applied MobileNetV2 Architecture 11
4.3.3 Residual Neural Networks 50 (ResNet50) 11
4.3.4 Visual Geometry Group 2019 (VGG19) 11
Chapter 5 12
Results & Discussion 12
5.1 MobileNetV2 Model 12
5.2 Convolutional Neural Network (ConvNet) Model 13
5.3 Residual Neural Network 50 (ResNet 50) Model 14
5.4 Visual Geometry Group 2019 (VGG19) Model 16
5.6 Comparative Analysis with Related Works 17
5.7 Discussion 18
Chapter 6 19
vi
Telediagnosis Software Based on the Proposed Method 19
6.1 Telediagnosis Service of the WebApp 19
6.2 Web App Development Environments 20
Conclusion 21
Future Work 21
References 22
List of Abbreviation and Acronyms 24

vii
 List of Figures

Fig. 3.1.1 Benign (hematogone) cells 7

Fig 3.1.2 Early Pre-B ALL cells 8
Fig. 3.1.3 Pre-B ALL cells 8
Fig. 3.1.4 Pro-B ALL cells 9
Fig. 4.1.1 Workflow diagram of the proposed method 10
Fig. 5.1.1 Result of the MobileNetV2 model 12
Fig. 5.1.2 Training accuracy vs validation accuracy of the MobileNetV2 model 12
Fig. 5.1.3 Training Loss vs validation loss of the MobileNetV2 model 13
Fig. 5.2.1 Result of the ConvNet model 13
Fig. 5.2.2 Training accuracy vs validation accuracy of the ConvNet model 14
Fig. 5.2.3 Training Loss vs validation loss of the ConvNet model 14
Fig. 5.3.1 Result of the ResNet50 model 15
Fig. 5.3.2 Training accuracy vs validation accuracy of the ResNet50 model 15
Fig. 5.3.3 Training Loss vs validation loss of the ResNet50 model 15
Fig. 5.4.1 Result of the VGG19 the model 16
Fig. 5.4.2 Training accuracy vs validation accuracy of the VGG19 model 16
Fig. 5.4.3 Training Loss vs validation loss of the VGG19 model 17
Fig. 6.1.1 Architecture of the telediagnosis software 19
Fig. 6.1.2 Web Application home page 19
Fig. 6.1.3 Diagnosing blood smear image 20

viii
 List of Tables

Table I Testing accuracy and loss of the related works 17

Table II Testing accuracy and loss of the DNN models of the proposed 18
method

ix
 Chapter 1
Introduction
1.1 Composition of Blood
Blood consists of several elements. The major components of blood include plasma, red
blood cells, white blood cells, and platelets. Plasma is the major constituent of blood and
comprises about 55 percent of blood volume. It consists of water with several different
substances dissolved within [1]. Almost 45 percent of blood includes Red Blood Cells
(RBC), White Blood Cells (WBC), and platelets. The RBC rate ranges from 4,000,000 to
6,000,000 per microliter of blood, representing 40–45% of the total blood volume [2].
WBCs are the cells that defend against germs and give us immunity and resistance; they
range from 4500 to 11,000 per microliter of blood [3]. The platelets range from 150,000 to
450,000 per microliter of blood and are responsible for blood clotting [4]. Thus, an increase
or decrease in any of the basic blood components will cause problems to a person’s health,
such as leukemia, thalassemia, and anemia.

1.2 Problem Statement

Leukemia is one of the most dangerous types of malignancies affecting the bone marrow or
blood in all age groups, both in children and adults. The most dangerous and deadly type of
leukemia is acute lymphoblastic leukemia (ALL). Acute Lymphoblastic Leukemia (ALL)
is a cancer of the blood cells that are characterized by a large number of immature
lymphocytes, known as blast cells (myeloblasts) [5]. lymphocytes are classified into three
types. These are (I) normal lymphocytes, (II) atypical lymphocytes, and (III) reactive
lymphocytes. Normal lymphocytes are characterized by homogeneity and round, small, and
rough nuclei. Atypical cells, by a large size and nucleus and the fact that they have lumpy
chromatins. Reactive cells, by their heterogeneity and the fact that they are surrounded by
red cells. Microscopic examination is the method of diagnosing lymphocyte types. It
involves taking blood or bone marrow samples, which are diagnosed by a pathologist [6].
ALL is classified into three morphological types: (I) L1 acute lymphoblastic leukemia, (II)
L2 acute lymphoblastic leukemia, and (III) L3 acute lymphoblastic leukemia [7]. There, L1
cells are the smallest, with a uniform population and coarse chromatins. L2 cells have
nuclear heterogeneity and are larger than L1 cells. L3 cells have vacuoles protruding into
the cells and are larger than L1 cells [8]. About 90 % of patients can be cured, provided that
the disease is diagnosed at an early stage by adopting mass screening procedures with an
intelligent automated diagnosis system.

1.3 Prevalence of Leukemia

According to the statistics from the World Health Organization’s (WHO) International
Agency for Research on Cancer (IARC), the number of leukemia cases in both sexes of all
age groups in 2018 was 4,37,033 and a total of 3,03,006 deaths [9]. Globally, the incidence
rate (per 1,00,000) was 5.2 and the mortality rate (per 1,00,000) was 3.5. A study was
conducted from January 2014 to December 2016 in the department of Pediatric Hematology
and oncology of Dhaka Shishu Hospital, Bangladesh. During this period all diagnosed
children with ALL (12 months to 16 years) were included in the study who were selected to
1
give chemotherapy. Out of the total of 597 children, 327 (54.7%) were males and 270
(45.3%) were female. Parents of a total of 199 (3.33%) children refused to start therapy and
the rest 398 children began chemotherapy. Among the total children who started
chemotherapy, 120 (20.10%) abandoned treatment in different phases of therapy. 66 (55%)
abandoned treatment during induction of remission. 23 (19.16%) during consolidation,
12(10%) re-induction, and 19 (1585%) children during maintenance phases of
chemotherapy [10]. There the 199 patient's parents refused treatment because of the high
cost and the treatment duration was very lengthy (about 10 years). The study shows the
severity of ALL.

1.4 Difficulties in Existing Manual Diagnosis

The evaluation of bone marrow morphology and blood smear in an electronic microscope
by experienced hematopathologists is essential in the diagnosis of Acute Lymphoblastic
Leukemia (ALL). However, it suffers from a lack of standardization and inter-observer
variability. Because diagnosing Acute Lymphoblastic Leukemia (ALL) is a tedious task that
is prone to human and machine errors. In several instances, it is difficult to make an accurate
final decision even after careful examination by an experienced pathologist. Manual
detection of ALL cells is conducted by pathologists and is typically subject to human error
and produces inaccurate results. This process is tedious, time-consuming, and subject to
inter and intra-class variations among pathologists. Only 76.6% of the cases showed
agreement between pathologists during leukemia diagnosis [11].

1.5 Challenges in Automated Diagnosis using DNN

The challenges in automated detection are associated with the complex nature of WBCs,
including irregular boundaries and the textural similarities between WBCs and other blood
components, which cause difficulties in separating blast cells (cancer cells) and healthy
cells. [4] Besides The Learnability of a Deep Learning Model strongly depends on the
effective and efficient data preparation pipeline. The accuracy and time complexity of ALL
detection is strongly associated with the quality of the extracted features used in training the
pixel-wise classification models.

1.6 Proposed Approach

In this thesis, we propose to develop multiple robust Deep Neural Network Models for
Medical Image Data with Acute Lymphoblastic Leukemia (ALL) cells, with enhanced
learnability, predictability, and accuracy. Besides, we develop a Web Application to provide
real-time support to diagnose ALL cells. We have also analyzed the performance of our
proposed method with multiple related works to assure the robustness of the proposed
model.

1.7 Contributions of the Proposed Approach

● Subtype classification of Acute Lymphoblastic Leukemia with almost 98%
accuracy.
● Robust Deep Neural Network model for multiclass imbalanced data sets.
● Real-time telediagnosis support through the web application which has been built
based on the proposed approach.

2
1.8 Organization of the Study
We organize the study as follows. In Chapter 2 Literature Review, we briefly describe the
related work’s methodology, outcomes, and limitations. In Chapter 3 Dataset Description,
we elaborately discuss the data set and illustrate the image samples. After that in Chapter 4,
the proposed method is detailed. Chapter 5 Result and Discussion elaborate on the
performance of the models. Finally, In Chapter 6 we discussed the developed telediagnosis
software. In the end we have concluded the research and mentioned the future works.

3
 Chapter 2
Literature Review
Researchers applied various methods which incorporate Deep Learning (DL), Deep Neural
Networks (DNN), and Medical Image Processing to detect leukemia cells from blood smear
images and bone marrow images.

Rohan et. al [5] developed an automation system to detect ALL blast cells in microscopic
blood smears using the You Only Look Once (YOLO v4) algorithm. They trained and
evaluated their method using the ALL-IDB1 and C_NMC_2019 data sets. The ALL-IDB1
dataset contained 108 images, which were collected in September 2005. The C_NMC_2019
dataset contained 10,661 images collected from 73 subjects. The images in this dataset were
single-celled and these were already pre-processed. In the preprocessing step they applied
Auto Orientation, Resizing, and Augmentation techniques. The preprocessing steps
generate 872 pre-process images after augmentation on the ALL-IDB1 dataset. Finally
achieved The mAP (Mean Average Precision) of 96.06 % for the ALL-IDB1 dataset and
98.7 % for the C_NMC_2019 dataset. Where the loss was seen to reduce exponentially and
reached an average of 0.57664 at the end of 6000 iterations on the ALL-IDB1 dataset and
also the loss was seen to reduce exponentially and reached an average of 0.502 at the end of
6000 iterations on the C_NMC_2019 dataset. But no complete automated screening or
diagnosing telehealth facility was established in this study.

The authors in [8] proposed 3 distinct systems which were built based on ALL-IDB1 and
ALL-IDB2. The first consists of the artificial neural network (ANN), feed-forward neural
network (FFNN), and support vector machine (SVM), all of which are based on hybrid
features extracted using Local Binary Pattern (LBP), Gray Level Co-occurrence Matrix
(GLCM) and Fuzzy Color Histogram (FCH) methods. The second proposed system consists
of the convolutional neural network (CNN) models: AlexNet, GoogleNet, and ResNet-18,
based on the transfer learning method, in which deep feature maps were extracted and
classified. The third proposed system consists of hybrid CNN–SVM technologies,
consisting of two blocks: CNN models for extracting feature maps and the SVM algorithm
for classifying feature maps. In the first system, ANN and FFNN reached an accuracy of
100%, while SVM reached an accuracy of 98.11%. In the second proposed system, all the
transfer learning models have achieved 100% accuracy and in the last proposed system
AlexNet + SVM achieved 100% accuracy, Goog-LeNet + SVM achieved 98.1% accuracy,
and ResNet-18 + SVM achieved 100% accuracy. Although the feature extraction methods
are well enough, the data preparation pipeline is not generalized and effective because of
the single Denoising technique (average and Laplacian filters) in the pre-processing step.
Where the robustness of the model comes into question.

A multi-step DL approach was performed in [12] to automatically segment Leukemia cells

from bone marrow images. The authors analyzed 1251 patients who have been newly
diagnosed and treated with AML and 236 Healthy Cohort. Then build the model using 5204
augmented AML bone marrow smear images and 5428 augmented control bone marrow
smear images. The multi-step includes the following workflow. First, initial segmentation
was done with the VGG Image Annotator tool. Then, the FRCNN was trained with the
segmented images and created new segmentation proposals for unsegmented images which
4
were manually corrected by hematologists. Step 2: Feature extraction was performed
manually by hematologists. Step 3: For the distinction between AML and healthy control
samples based on segmented images, the authors trained a multitude of DL models for
binary predictions. In step 4: For NPM1 status prediction, transfer learning with a ResNet50
pre-trained on ImageNet was utilized on BMS images. Finally, the FRCNN achieved a cell
segmentation accuracy of 0.97 from BMS. The binary classification model showed an AUC
of 0.97 for both the ROC and the precision-recall curve and a micro-average accuracy of
0.91 distinguishing between AML and healthy bone marrow donor samples. For NPM1
prediction their DL model achieved a high accuracy of 0.86 in predicting mutation status.
Although the study has a good outcome, the manual feature extraction technique may lead
to inefficiency and error with the large dataset.

The AML_Cytomorphology_LMU dataset consisted of 18,365 expert-labeled single-cell

images obtained from peripheral blood smears of 100 AML patients and 100 controls at
Munich University Hospital between 2014 and 2017 used by the study [13]. The author
proposed a new hybrid feature extraction method using image processing and deep learning
methods. The proposed method consists of two steps: 1) a region of interest (ROI) is
extracted using the CMYK-moment localization method and 2) deep learning-based
features are extracted using a CNN-based feature fusion method. The method obtained
overall classification accuracies of 97.57%. There the image data preparation pipeline is not
generalized which can lead to a mass accuracy fall in terms of new image data.

Authors developed a method in [26] to detect ALL using transfer learning and later validate
the model using XAI. As the dataset was imbalanced with the more ALL class, the class
weight method was used to balance the weights of the two classes in preprocessing For the
used dataset, the class weight of normal cells was 1.57288, and the class weight of acute
lymphocytic leukemia (ALL) cells was 0.73301. Then, different pre-trained models, namely
InceptionV3, ResNet101V2, VGG19, and InceptionResNetV2, were used to train the
model. All the images were reduced to 299px * 299px because the Inception V3 input shape
must be (299, 299, 3) and a standard batch size of 32 is used while training. A segmentation
method was employed to divide the example image into separate sections to see if the model
could accurately read it. ResNet101v2 model Accuracy is 0.9861, Loss is 0.0333 and
validation accuracy is 0.9589, Validation loss is 0.1559 F1 score 0.9861, and Validation F1
score 0.9588.

The author of this paper [27] established a database, which consists of 1,732 images
obtained from the bone marrow smears of 89 children with leukemia from 2009 to 2019 at
the Shanghai Children’s Medical Center (SCMC). For multi-class WBC differentiation,
they developed three deep learning models (ResNext101_32∗ 8d swsl,
ResNext50_32∗ 4dswsl, and ResNet50). Finally, the Accuracy, Ap, F1 score, and AUC are
0.8149, 0.7982, 0.8073, and 0.8293 respectively. In this study, authors retrospectively
collected 1,732 bone marrow images containing 27,184 cells (including 24,165 cells and
2,983 cell debris) from 89 children with leukemia from the Shanghai Children’s Medical
Center. We randomly separated 70% of the cells in the training set, and the remaining cells
were used to form the validation set and test set.

In this study [28] 4,451 real images of bone marrow cells on the Hyde star HDS-BFS high-
speed micro scanning image system were collected, with a resolution of 4,000×3,000. The
object detection model is based on the Faster RegionConvolutional Neural Network (R-
CNN). Faster R-CNN is composed of the region proposal network (RPN) and Fast R-CNN,

5
in which RPN is used to select candidate target boxes and Fast R-CNN is used for accurate
target classification and regression. Compared to the baseline, although their method
resulted in a slight drop in recall (~4%), it improved the precision by 26.4%, the F1-score
by 12.1%, and the AP@50 by 3%. The model achieves a recall of 0.710, precision of 0.496,
AP@50 of 0.533, and F1-score of 0.575. They developed a new morphological diagnosis
system for bone marrow cells. The model is based on the Faster R-CNN.

The study [29] focuses on the C-NMC-2019 dataset and the application of DL and the
ensemble strategy. In their proposed model, the Hem class was over-sampled to 5822
images, and a total of 11644 images were trained during the training process. In this
circumstance, five pre-trained networks, VGG-16, Xception, MobileNetV2,
InceptionResNet-V2, and DenseNet-121, are adopted for transfer learning applications. The
image is transferred through a stack of convolutional layers, where the filters are used with
tiny receptive filters (3 × 3). The Kappa-based ensemble model has produced the best ALL
recognition results, with 89.72% WFS and 94.8% AUC.

The authors in [31], applied feature extraction using PCA 2 to reduce dimensionality in a
heart disease dataset. There the proposed method integrates the K-means Clustering and
Genetic Algorithm and finally achieved 94.06% of clustering accuracy. In [32], the authors
proposed a novel feature selection algorithm in a chronic kidney disease dataset that
aggregated wrapper, filter, and ensemble methods and achieved 100% accuracy in
prediction.

6
 Chapter 3
Dataset Description

We have used the Acute Lymphoblastic Leukemia (ALL) Image dataset. The images of this
dataset [24] [25] were prepared in the Bone Marrow Laboratory of Taleqani Hospital
(Tehran, Iran). This dataset consisted of 3256 peripheral blood smear (PBS) images from
89 patients suspected of ALL, whose blood samples were prepared and stained by skillful
laboratory staff, including 25 healthy individuals with a benign diagnosis (hematogone) and
64 patients with a definitive diagnosis of ALL subtypes, Early Pre-B, Pre-B, and Pro-B
ALL. This dataset is divided into two classes benign and malignant. The former comprises
hematogenous and the latter is the ALL group with three subtypes of malignant
lymphoblasts: Early Pre-B, Pre-B, and Pro-B ALL. All the images were taken by using a
Zeiss camera in a microscope with a 100x magnification and saved as JPG files. A specialist
using the flow cytometry tool made the definitive determination of the types and subtypes
of these cells. After color thresholding-based segmentation in the HSV color space, dataset
collectors also provide segmented images. Figure 3.1.1 - 3.1.4 depicts the images of Benign
(hematogone) Cells, Early Pre-B ALL Cells, Pre-B ALL Cells, and Pro-B ALL Cells
respectively.

Fig. 3.1.1 Benign (hematogone) cells

7
Fig. 3.1.2 Early Pre-B ALL cells

Fig. 3.1.3 Pre-B ALL cells

8
 Fig. 3.1.4 Pro-B ALL cells

We have used the Acute Lymphoblastic Leukemia (ALL) Dataset to train and validate our
DNN model. It is a multiclass dataset that contains a benign cell and 3 subtypes of blast
cells named Early Pre-B, Pre-B, and Pro-B. A DNN model trained with ALL datasets may
aid in detecting the subtypes of Acute Lymphoblastic Leukemia by classifying all types of
subtypes.

9
 Chapter 4
Proposed Method
4.1 Overview
We have selected the Acute Lymphoblastic Leukemia (ALL) public datasets to build our
Deep Neural Network models. Figure 4.1.1 illustrates the workflow of the proposed method.

Fig. 4.1.1 Workflow of the proposed method

In the beginning, we called the Kaggle API to fetch the dataset. Then we split the dataset
into Training, Validation, and Testing Set to avoid data leakage and overfitting. Then we
preprocessed the Training Set using our Data Preprocessing Pipeline. Then we have only
normalized the Validation and Testing Set’s images. It will keep the Global and Local
Minima-Maxima different, which is another essential procedure to avoid data leakage and
overfitting. Then the Training and Validation images have passed to the DNN architectures.
Now, Multiple DNN Algorithms have been trained and validated from the Training and
Validation Set. The best-fitted models depending on validation accuracy at this stage have
been selected as optimized models. Now the normalized Test images have been tested with
the Multiple Optimized DNN Models. Finally, the images in the Test set will be classified
into Benign, Early Pre-B, Pre-B, and Pro-B to detect various subtypes of acute
lymphoblastic leukemia cells.

4.2 Data Preparation Pipeline

The performance of a DNN depends largely on suitable data preparation. Our data
preparation pipeline contains the following techniques to pre-process the microscopic
images for ALL detection.
● Resizing (224px width, 224px height, 3 channels)
● Padding
● Augmentation
● Normalization
There the Augmentation technique contains the following techniques.

10
 ● Sheared Range ( 0.2)
● Zooming (0.2)
● Flipping (Horizontally)

4.3 Deep Neural Network Architecture

We have implemented the following DNN algorithms for our thesis work.
● Convolutional Neural Network (ConvNet)
● MobileNetV2
● Residual Neural Network 50 (ResNet50)
● Visual Geometry Group 2019 (VGG19)

4.3.1 Applied Convolutional Neural Network (ConvNet)

We have built an 11-layer Convolutional Neural Network with 4 Convolutional Layers, 3
Max Pooling layers, 2 Dropout layers, a single Flatten, and a Dense layer with SoftMax
activation function for multi-class classification. The ConvNet model contains 5,638,440
trainable parameters.

4.3.2 Applied MobileNetV2 Architecture

We have built a MobileNetV2 architecture with 81 layers. There are 13 Conv2D layers, 24
Convolutional layers for Batch Normalization, 26 layers for Relu activation function, 12
Depth wise Convolutional layers, 4 Zero padding layers, a single Flatten layer, and finally
a Dense layers with 4 unit and SoftMax activation function. There are total parameters of
3,429,572 where 200,708 parameters are trainable and 3,228,864 non-trainable parameters.

4.3.3 Residual Neural Networks 50 (ResNet50)

We have built a ResNet50 with 149 layers, where 42 Activation layers, 46 Convolutional
Layers, 45 layers for Batch wise Normalization, 16 Add layers, 2 Zero Padding layers, a
Flatten layer, and a Dense layer. There is a total of 23,989,124 parameters, 401,412 trainable
parameters, and 23,587,712 Non-trainable parameters.

4.3.4 Visual Geometry Group 2019 (VGG19)

The VGG19 architecture contains a total of 20,124,740 parameters, where 100,356
parameters are trainable and 20,024,384 non-trainable parameters. There are 16
Convolutional layers, 5 Max Pooling layers, a single Flatten layer, and a single Dense layer.

The complete code of the study is available in the following GitHub repository.
https://github.com/Muhammad-Taufiq-Khan/Automated-Detection-of-Acute-
Lymphoblastic-Leukemia-Subtypes-from-Microscopic-Blood-Smear-Images

11
 Chapter 5
Results & Discussion

5.1 MobileNetV2 Model

The Optimized MobileNetV2 Model provides Training, Validation, and Testing accuracy
of 0.9312, 0.9799, and 0.9742 respectively. There the Training, Validation, and Testing Loss
are 0.7753, 0.1792, and 0.2351 respectively. Figure 5.1.1 shows the result of the
MobileNetV2 Model. Figure 5.1.2 shows the Training accuracy vs validation accuracy of
the MobileNetV2 Model and Fig. 5.1.3 shows the Training loss vs Validation loss of the
MobileNetV2 Model.

Fig. 5.1.1 Result of the MobileNetV2 model

Fig. 5.1.2 Training accuracy vs validation accuracy of the MobileNetV2 model

12
 Fig. 5.1.3 Training loss vs validation loss of the MobileNetV2 model

5.2 Convolutional Neural Network (ConvNet) Model

The accuracy and loss of the Optimized ConvNet Model in the Training, Validation, and
Testing has been shown in Fig. 5.2.1. Figure 5.1.2 shows the Training accuracy vs
Validation accuracy of the model and Fig. 5.1.3 shows the Training loss vs validation loss
of the model.

Fig. 5.2.1 Result of the ConvNet model

13
 Fig. 5.2.2 Training accuracy vs validation accuracy of the ConNet model

Fig. 5.2.3 Training loss vs validation loss of the ConvNet model

5.3 Residual Neural Network 50 (ResNet 50) Model

The accuracy and loss of the Optimized ResNet Model on the Training, Validation, and
Testing has been shown in Fig. 5.3.1. Figure 5.3.2 shows the Training accuracy vs validation
accuracy of the model, and Fig. 5.3.3 shows the Training loss vs validation loss of the model.

14
 Fig. 5.3.1 Result of the ResNet50 model

Fig. 5.3.2 Training accuracy vs validation accuracy of the ResNet50 model

Fig. 5.3.3 Training loss vs validation loss of the ResNet50 model

15
5.4 Visual Geometry Group 2019 (VGG19) Model
The accuracy and loss of the Optimized VGG19 Model on the Training, Validation, and
Testing has been shown in Fig. 5.4.1. Figure 5.4.2 shows the Training accuracy vs validation
accuracy of the model, and Fig. 5.4.3 shows the Training loss vs validation loss of the model.

Fig. 5.4.1 Result of the VGG19 model

Fig. 5.4.2 Training accuracy vs validation accuracy of the VGG19 model

16
 Fig. 5.4.3 Training loss vs validation loss of the VGG19 model

5.6 Comparative Analysis with Related Works

TABLE I
TESTING ACCURACY AND LOSS OF THE RELATED WORKS
Title & Year Algorithm Accuracy Loss

Automated blast YOLOv4 for

cell detection for ALL_IDB1 92.0% 8.0%
Acute
Lymphoblastic YOLOv4 for
Leukemia diagnosis C_NMC-2019 96.0% 4.0%
(2021) [5]

Feature Extraction FCL 94.92% 5.08%

of White Blood
Cells Using CMYK RF 95.47% 4.53%
-Moment
localization SVM 96.41% 3.59%
andDeep Learning
in Acute Myeloid XGBoost 95.18% 4.82%
Leukemia Blood
Smear Microscopic
Image (2022) [13]

Executing Spark GoogleNet 96.06% 3.995%

BigDl for Leukemia
Detection from
Microscopic Image CNN 94.69% 5.32%
using Transfer
Learning (2021)
[30]

17
 Hybrid Inception v3 AlexNet 89.4% 10.6%
XGBoost Model for
Acute DenseNet121 86.9% 13.1%
lymphoblastic
Leukemia ResNet18 91.7% 8.3%
Classification
(2021) [7] VGG16 92.4% 7.65%

SqueezeNet 93.2% 6.8%

MobileNetV2 95.8% 4.2%

Development and ResNet101 81.49% 18.51%

Evaluation of a
Leukemia Diagnosis ResNext5 79.82% 20.18%
System Using Deep
Learning in Real ResNet50 80.73% 19.27%
Clinical Scenarios
(2021) [27] Ensemble 82.93% 17.07%

5.7 Discussion
Analyzing the above results from Fig. 5.1.1 - 5.4.3 we can conclude that except for
ResNet50 the other models provide desirable outcomes. There the MobileNetV2 model
performs higher in terms of Training Accuracy and VGG19 provides the lowest loss. Table
II shows the Testing accuracy and loss of the optimized models below. Also, compared to
Table I the MobileNetV2 model provides higher accuracy.
TABLE II
TESTING ACCURACY AND LOSS OF THE DNN MODELS OF THE PROPOSED METHOD
Accuracy Loss

MobileNetV2 0.9742 0.2351

VGG19 0.9613 0.099

ConvNet 0.9128 0.2309

ResNet50 0.8526 0.8412

18
 Chapter 6
Telediagnosis Software Based on the Proposed Method

6.1 Telediagnosis Service of the WebApp

We have developed a Realtime Tele Diagnosis Web App to aid in the detection of ALL
subtypes. Fig 6.1.1 illustrates the architecture of the Telediagnosis software. The App
collects ALL Images from the user-end (Frontend) and passes the image data to the server-
end (Backend). The server processes the images and calls the best DNN model. Finally, the
model predicts the subtype of the ALL and passes the result to the user-end. Fig 6.1.2
illustrates the home page of the Web Application and fig 6.1.3 illustrates how the diagnosis
result is shown to the user-end.

Fig. 6.1.1 Architecture of the telediagnosis software

Fig. 6.1.2 Web Application home page

19
 Fig. 6.1.3 Diagnosing blood smear image

The complete code of the web application is available in the following github repository.
https://github.com/Muhammad-Taufiq-Khan/ALL_Subtype_Detector_WebApp

6.2 Web App Development Environments

● Programming Language: Python
● Backend: Flask
● Frontend: Bootstrap5, HTML, CSS
● Deep Learning Library: Tensorflow-Keras
● Digital Image Processing Library: OpenCV
● Cloud Platform: Heroku.

20
 Conclusion
A lot of people are losing lives because of the severity of Acute Lymphoblastic Leukemia
and the number is increasing day by day. The patients who survive lose their vitality. The
current Artificial Intelligence-based automated systems are also facing challenges due to the
various shape, patterns, and textures of ALL blast cells. We have developed optimized
Multi-DNN models and a telediagnosis web application, which will fight against the
severity of ALL by providing efficient diagnosis assistance.

Future Work
According to our research many researchers have applied DNN to detect Acute
Lymphoblastic Leukemia but a thorough method and generalized model are still missing.
In the future, we would like to develop a robust method, which will be able to perform
segmentation, ROI detection, estimated blast cell calculation, and generalized classification.

21
References
[1] R. Bailey, “What’s in your blood?,” ThoughtCo. [Online]. Available:
https://www.thoughtco.com/blood-373480. [Accessed: 16-Apr-2022].

[2] F. Al-Hafiz, S. Al-Megren, and H. Kurdi, “Red blood cell segmentation by thresholding
and Canny detector,” Procedia Comput. Sci., vol. 141, pp. 327–334, 2018.

[3] I. Rahadi, M. Choodoung, and A. Choodoung, “Red blood cells and white blood cells
detection by image processing,” J. Phys. Conf. Ser., vol. 1539, no. 1, p. 012025, 2020.

[4] P. E. J. van der Meijden and J. W. M. Heemskerk, “Platelet biology and functions: new
concepts and clinical perspectives,” Nat. Rev. Cardiol., vol. 16, no. 3, pp. 166–179, 2019.

[5] R. Khandekar, P. Shastry, S. Jaishankar, O. Faust, and N. Sampathila, “Automated blast

cell detection for Acute Lymphoblastic Leukemia diagnosis,” Biomed. Signal Process.
Control, vol. 68, no. 102690, p. 102690, 2021.

[6] B. J. Bain, “Diagnosis from the blood smear,” N. Engl. J. Med., vol. 353, no. 5, pp. 498–
507, 2005.

[7] C. Spier, “Wintrobe’s atlas of clinical hematology,” Am. J. Surg. Pathol., vol. 32, no. 9,
p. 1428, 2008.

[8] I. Abunadi and E. M. Senan, “Multi-method diagnosis of blood microscopic sample for
early detection of acute lymphoblastic leukemia based on deep learning and hybrid
techniques,” Sensors (Basel), vol. 22, no. 4, p. 1629, 2022.

[9] S. Hosseini et al., “Menstrual blood contains immune cells with inflammatory and anti-
inflammatory properties: T and NKT cells in menstrual blood,” J. Obstet. Gynaecol. Res.,
vol. 41, no. 11, pp. 1803–1812, 2015.

[10] S. Tahura and M. Hussain, “Treatment refusal and abandonment in pediatric patients with
Acute Lymphoblastic Leukemia in Bangladesh.” IJSR, ISSN (Online): 2319-7064, pp. 1-
3, 2017

[11] L. Bigorra, A. Merino, S. Alférez, and J. Rodellar, “Feature analysis and automatic
identification of leukemic lineage blast cells and reactive lymphoid cells from
peripheral blood cell images,” J. Clin. Lab. Anal., vol. 31, no. 2, p. e22024, 2017.

[12] J.-N. Eckardt et al., “Deep learning detects acute myeloid leukemia and predicts
NPM1 mutation status from bone marrow smears,” Leukemia, vol. 36, no. 1, pp. 111–
118, 2022.

[13] T. A. M. Elhassan, M. S. M. Rahim, T. T. Swee, S. Z. M. Hashim, and M. Aljurf,

“Feature extraction of white blood cells using CMYK-moment localization and deep
learning in acute myeloid leukemia blood smear microscopic images,” IEEE Access,
vol. 10, pp. 16577–16591, undefined 2022.

[14] R. Donida Labati, V. Piuri, F. Scotti, "ALL-IDB: the acute lymphoblastic leukemia
image database for image processing", in Proc. of the 2011 IEEE Int. Conf. on Image
Processing (ICIP 2011), Brussels, Belgium, pp. 2045-2048, September 11-14, 2011.
ISBN: 978-1-4577-1302-6.

[15] F. Scotti, "Robust Segmentation and Measurements Techniques of White Cells in

Blood Microscope Images", in Proc. of the 2006 IEEE Instrumentation and

22
 Measurement Technology Conf. (IMTC 2006), Sorrento, Italy, pp. 43-48, April 24-
27, 2006. ISSN: 1091-5281.

[16] F. Scotti, "Automatic morphological analysis for acute leukemia identification in

peripheral blood microscope images", in Proc. of the 2005 IEEE Int. Conf. on
Computational Intelligence for Measurement Systems and Applications (CIMSA
2005), Giardini Naxos - Taormina, Italy, pp. 96-101, July 20-22, 2005

[17] V. Piuri, F. Scotti, "Morphological classification of blood leucocytes by microscope

images", in Proc. of the 2004 IEEE Int. Conf. on Computational Intelligence for
Measurement Systems and Applications (CIMSA 2004), Boston, MA, USA, pp. 103-
108, July 12-14, 2004. ISBN: 0-7803-8341-9.

[18] Gupta, A., & Gupta, R. (2019). ALL Challenge dataset of ISBI 2019 [Data set]. The
Cancer Imaging Archive. https://doi.org/10.7937/tcia.2019.dc64i46r

[19] Rahul Duggal, Anubha Gupta, Ritu Gupta, and Pramit Mallick, "SD-Layer: Stain
Deconvolutional Layer for CNNs in Medical Microscopic Imaging," In: Descoteaux
M., Maier-Hein L., Franz A., Jannin P., Collins D., Duchesne S. (eds) Medical Image
Computing and Computer-Assisted Intervention − MICCAI 2017, MICCAI 2017.
Lecture Notes in Computer Science, Part III, LNCS 10435, pp. 435–443. Springer,
Cham.

[20] Rahul Duggal, Anubha Gupta, Ritu Gupta, Manya Wadhwa, and Chirag Ahuja,
“Overlapping Cell Nuclei Segmentation in Microscopic Images UsingDeep Belief
Networks,” Indian Conference on Computer Vision, Graphics and Image Processing
(ICVGIP), India, December 2016.

[21] Rahul Duggal, Anubha Gupta, and Ritu Gupta, “Segmentation of

overlapping/touching white blood cell nuclei using artificial neural networks,” CME
Series on Hemato-Oncopathology, All India Institute of Medical Sciences (AIIMS),
New Delhi, India, July 2016.

[22] Ritu Gupta, Pramit Mallick, Rahul Duggal, Anubha Gupta, and Ojaswa Sharma,
"Stain Color Normalization and Segmentation of Plasma Cells in Microscopic Images
as a Prelude to Development of Computer Assisted Automated Disease Diagnostic
Tool in Multiple Myeloma," 16th International Myeloma Workshop (IMW), India,
March 2017.

[23] Anubha Gupta, Rahul Duggal, Ritu Gupta, Lalit Kumar, Nisarg Thakkar, and
Devprakash Satpathy, “GCTI-SN: Geometry-Inspired Chemical and Tissue Invariant
Stain Normalization of Microscopic Medical Images,”, under review.

[24] Mehrad Aria, Mustafa Ghaderzadeh, Davood Bashash, Hassan Abolghasemi,

Farkhondeh Asadi, and Azamossadat Hosseini, “Acute Lymphoblastic Leukemia
(ALL) image dataset.” Kaggle, (2021). DOI: 10.34740/KAGGLE/DSV/2175623.

[25] M. Ghaderzadeh, M. Aria, A. Hosseini, F. Asadi, D. Bashash, and H. Abolghasemi,

“A fast and efficient CNN model for B‐ALL diagnosis and its subtypes classification
using peripheral blood smear images,” Int. J. Intell. Syst., 2021.

[26] W. H. Abir et al., “Explainable AI in diagnosing and anticipating leukemia using

transfer learning method,” Comput. Intell. Neurosci., vol. 2022, p. 5140148, 2022.

[27] M. Zhou et al., “Development and evaluation of a leukemia diagnosis system using

23
 deep learning in real clinical scenarios,” Front. Pediatr., vol. 9, p. 693676, 2021

[28] J. Liu et al., “A deep learning method and device for bone marrow imaging cell
detection,” Ann. Transl. Med., vol. 10, no. 4, p. 208, 2022.

[29] C. Mondal et al., “Ensemble of Convolutional Neural Networks to diagnose Acute

Lymphoblastic Leukemia from microscopic images,” Inform. Med. Unlocked, vol.
27, no. 100794, p. 100794, 2021

[30] Aftab, M. O., Javed Awan, M., Khalid, S., Javed, R., & Shabir, H. (2021). Executing
spark BigDL for leukemia detection from microscopic images using transfer learning.
2021 1st International Conference on Artificial Intelligence and Data Analytics
(CAIDA), 216–220.

[31] Islam, M.T., Rafa, S.R., Kibria, M.G.: Early prediction of heart disease using PCA
and hybrid genetic algorithm with k-means. In: 2020 23rd International Conference
on Computer and Information Technology (ICCIT). IEEE (2020).

[32] Haque Khan Tusar, M.T., Islam, M.T., Raju, F.I.: Detecting chronic kidney
disease(CKD) at the initial stage: A novel hybrid feature-selection method and robust
data preparation pipeline for different ML techniques. In: 2022 5th International
Conference on Computing and Informatics (ICCI). IEEE (2022).

List of Abbreviation and Acronyms

AML Acute Myeloid Leukemia

ALL Acute Lymphoblastic Leukemia

DNN Deep Neural Networks

NPM1 Nucleophosmin Gene

FRCNN Faster R-CNN Object Detection Network

WHO World Health Organization

VGG19 Visual Geometry Group 2019

ResNet50 Residual Neural Network 50

ConvNet Convolutional Neural Network

YOLOV4 You Only Look Once

FCL Fully Connected Layers

RF Random Forest

SVM Support Vector Machine

XGBoost Extreme Gradient Boosting

FFNN Feed Forward Neural Network

FCH Fuzzy Color Histogram

24
LBP Local Binary Pattern

PBS Peripheral Blood Smear

ANN Artificial Neural Network

25