Content uploaded by Chathurangi Shyalika
Author content
All content in this area was uploaded by Chathurangi Shyalika on Oct 04, 2020
Content may be subject to copyright.
Journal of Theoretical and Applied Information Technology
30th September 2020. Vol.98. No 18
© 2005 – ongoing JATIT & LLS
ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195
3740
STATISTICAL MORPHOLOGICAL ANALYSIS BASED
SUPERVISED CLASSIFICATION ALGORITHM FOR
DIAGNOSING ACUTE LYMPHOBLASTIC LEUKEMIA
1JKC SHYALIKA, 2PPNV KUMARA, 3DU KOTTAHACHCHI
1Graduate, Department of Information Technology, Faculty of Computing, General Sir John Kotelawala
Defence University, Sri Lanka
2Department of Computer Science, Faculty of Computing, General Sir John Kotelawala Defence
University, Sri Lanka
3Department of Medical Laboratory Sciences, Faculty of Allied Health Sciences, General Sir John
Kotelawala Defence University, Sri Lanka
E-mail: 1chathurangijks@gmail.com , 2nandana@kdu.ac.lk , 3darsha.uda@gmail.com
ABSTRACT
Leukemia is a fatal disease of the type “Blood Cancer”, where the White Blood Cells (WBC) increases in
human bone marrow and peripheral blood. Acute Lymphoblastic Leukemia (ALL) is a common types of
leukemia that affects young children of below 10 years and adults over 60 years, aroused by accumulation
and overproduction of immature and cancerous cells identified as lymphoblasts. At present, the diagnosis of
ALL includes measures alike performing a full blood count, bone marrow biopsy, blood picture,
immunophenotyping, cytochemical stain and cytogenetics. These medicinal techniques are highly tedious,
costly, requires expertise of hematologists and available only in few hospitals especially in developing
countries. Hence, as an alternative, use of image processing and machine learning to diagnose ALL would
become an effective solution. Even though, several research groups have used image processing to detect
and diagnose ALL, recognition and splitting of overlapping Red Blood Cells (RBC) with WBC has
however been a challenging issue. This paper is about a research study and an application that includes an
image processing and machine learning algorithm to diagnose ALL while attempting to solve the issue of
overlapping cells. The research is also extended to detect the quality devastation in blood films in terms of
storing them for prolonged period. The inputs for this application include microscopic peripheral blood
films of ALL patients and healthy individuals obtained from Department of Pathology Clinic at Faculty of
Medicine, University of Colombo, Sri Lanka. This research project has received verification of ethical
approval from Faculty of Medicine, General Sir John Kotelawala Defence University, Sri Lanka. In the
developed application, segmentation using morphological operations in OpenCV Python and supervised
learning based classification using K-Nearest Neighbour implementation has been proposed in detection
and diagnosing of ALL. As per the results, the proposed algorithm has led to a high accuracy of 88.8% in
diagnosing ALL. The end product includes a Python based QT GUI based development suite that performs
main targeted backend functionalities and a PHP based web application that serves hematologists, doctors
and patients to perform utility functions.
Keywords: Acute Lymphoblastic Leukemia, Image Processing, Segmentation and Feature Extraction,
Classification, K-Nearest Neighbour, Supervised Learning
1. INTRODUCTION
Blood is one of the most important materials of
the human body as it is the prime agent that make
humans live. Human blood consists of two major
parts; plasma and cells. Plasma consists of 90%
water and other compounds such carbohydrates,
proteins, hormones, lipids, electrolytes [1, 2]. In
adults, cells originate from the bone marrow in a
specialized cell type known as stem cells and
transferred to periphery when they become
matured. The peripheral blood cells consist of three
main components; Red blood cells (RBC), White
blood cells (WBC) and platelets. There are five
types of WBCs; Neutrophils (40-75%),
Lymphocytes (20-45%), Eosinophils (1-8%),
Journal of Theoretical and Applied Information Technology
30th September 2020. Vol.98. No 18
© 2005 – ongoing JATIT & LLS
ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195
3741
Monocytes (0-10%), Basophils (0.5-1%),
Neutrophils, Eosinophils and Basophils have
granules in their cytoplasm; hence they are named
as granulocytes. Monocytes and Lymphocytes do
not have granules, thus are termed agranulocytes
[1].
Leukemia, simple called “Blood cancer” in
which usually the number of WBC increase in the
bone marrow and peripheral blood. These leukemic
cells (usually immature) replace the other normal
blood cells causes malfunction of the bone marrow
as well as peripheral blood. Furthermore, excess
amount of these cells travels to other sites such as
liver, spleen to maintain normal cells production.
Later, the leukemia cells also invade other organs
causing them to malfunction [1].
There are two main types of leukemia according
to the morphology of cells in the bone marrow.
They termed as acute & chronic Leukemia.
Generally, acute leukemia involves the rapid
overgrowth of very immature blood cells whereas
chronic leukemia involves the overgrowth of
somewhat mature blood cells in the bone marrow
compared to acute type. With the introduction of
French-American-British (FAB) classification in
1976 [2], acute leukemia further categorized into
two groups based on the white blood cell from
which the malignancy originates from. They are
Acute Lymphoblastic Leukemia (ALL) is caused by
abnormal lymphoid cells, and Acute Myeloid
Leukemia (AML) is caused by abnormal myeloid
cells in the bone marrow [1]. The predominant
abnormal cells in the ALL are lymphoblasts.
Diagnosing leukemia usually begins with a
medical history and physical investigation. If
leukemia is suspected, the patient is made to
undergo a number of tests in order to detect and
diagnose leukemia and also to identify the sub type.
These tests include, performing Full Blood Count
(FBC), Blood Picture (BP) to identify abnormalities
in cell shape, a bone marrow aspiration and trephine
biopsy is then conducted to identify the type of
abnormal blood cells [3], cytochemical stains to
demonstrate enzymatic activity, carbohydrates or
lipids present or any other special characters present
in leukemia cells [2]. To clarify the subtypes and
also for comprehensive diagnosis, an advanced
technique such as immunophenotyping and
cytogenetics are employed [4, 5]. The whole
process takes about 3-4 days and also needs well-
trained experienced professionals to supervise.
However, early diagnosis of leukemia contributes to
early treatment and proper management of patients.
Furthermore, manual detection procedure stated
above is a highly tedious task that involves the
effort of hematologists and other supporting staff as
it is intensively slow, costly, time consuming. Even
though advanced techniques are being used, there
may be errors especially diagnosing subtypes.
Image Processing and Machine Learning fields
have provided fast, cost effective and accurate
solutions in fields such as medical image
management, image data mining, bio imaging,
neuroimaging and virtual reality in medical
visualization [6, 7]. Image processing techniques
includes Image acquisition, restoration, pre-
processing, segmentation, Feature extraction,
compression, wavelets, representation, recognition
etc. [7]. A digital image is a representation of a two-
dimensional image as a finite set of digital values
called pixels. Image processing is a type of signal
processing method that perform some operations on
these digital images in order to get an enhanced
image or to extract some useful information from it
[4]. Machine learning is the branch of Artificial
Intelligence that provides systems the ability to
automatically learn and improve from experience
that is without being programmed explicitly.
Machine learning focuses on the development of
computer programs that can access data and use it
learn for themselves. However, there are some
challenging issues in these areas. There are
unsolved problematic areas such as quality
degradations occurring in image compression and
enhancements, in recognizing generic objects,
visualization issues etc. [6, 7].
Researches have been conducted for the detection
and counting of RBC [8], white blood cells and to
diagnose diseases like anaemia, malaria and
deficiency of vitamin B12 using blood images [9].
Furthermore, Image Processing techniques have
been used for detecting cancer cells [10]. For
Image Classification, Supervised or Unsupervised
Machine learning techniques have been used [8, 10,
11]. The objective of this paper is to present the
results of a research project conducted in order to
diagnose a type of blood cancer, (ALL). In this
research project, an efficient and accurate image
processing algorithm to detect and diagnose ALL
cells using microscopic images obtained from
human blood peripheral blood films stained with
Leishman’s has been proposed. The proposed
system would use Digital Image Processing and
Machine learning techniques in order to complete
the task.
Journal of Theoretical and Applied Information Technology
30th September 2020. Vol.98. No 18
© 2005 – ongoing JATIT & LLS
ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195
3742
Section 2 of the paper discusses the early
research attempts on diagnosing ALL that
ultimately reveals the research gap for ALL
diagnosis. Section 3 is followed by the
methodology and the experimental design of the
solution developed by the authors. Section 4
presents the evaluation results incurred as the
research outcomes of the system testing. Section 5
concludes by indicating the research achievements
and identified research challenges for ALL
diagnosis.
2. EVOLUTION AND STATE-OF-THE ART
IN ALL DIAGNOSIS
Presently, a considerable contribution has
been done by researchers in the aim of ALL
diagnosis using image processing and machine
learning. The common flow of the image
processing techniques that is used in diagnosis can
be illustrated by Figure 1.
Figure 1: Flow of ALL Diagnosis
The researches done so far varies from the
segmentation methods and classification methods
they have used. A comparative review on the
segmentation methods and classification methods
deployed by early researchers in diagnosing ALL
have been elaborated in this section.
2.1. Segmentation Methods
Segmentation process partitions an image
to its constituent segments or objects known as
pixels. This locates objects and boundaries (curves
and lines etc.) of images and modifies the
representation of an image into somewhat that is
more meaningful and easier to analyze. In
literature, segmentation has been used to separate
the WBC from the cytoplasm, identify the
leucocytes and their nuclei, identify grouped
leucocytes and for image cleaning. Mostly used
segmentation methods used in leukemia detection
have been discussed in this section.
2.1.1. Watershed segmentation
“Watershed” refers to a ridge that splits areas
drained by different river systems [11]. Watershed
lines are defined on the nodes, edges, hybrid lines
on both nodes and edges and in continuous domain.
Watershed segmentation is an easy method for the
detection of WBC but requires best quality images
in order to achieve a better accuracy [11].
2.1.2. Fuzzy C Means clustering (FCM)
This data clustering technique groups a dataset into
n clusters with all data points in that dataset belong
to each and every cluster to a certain degree. FCM
result is much accurate and it’s able to measuring
nucleus boundaries with shape, colour and texture,
but it’s difficult in classification of lymphoblast in
to its sub types through this segmentation [11,12].
2.1.3. Fuzzy K-Means clustering in L*A*B*
colour space
K-Means method is a least squares partitioning
method and it divides a collection of objects to K
groups of clusters. It considers each object have a
location in the space and finds partitions in the
image such that objects within each cluster close to
each other as likely, and as far from the objects in
other clusters as possible. This method is not
applicable on incremental data and it cannot give
classification with labelled data [9]. Mohapatra and
the colleagues have used fuzzy based blood image
segmentation for separate out leucocytes from other
blood components [13].
2.1.4. Otsu’s method
This is a thresholding method and it’s the easiest
and fastest method used in segmentation.
Thresholding is based on a clip-level named a
threshold value used in converting a grayscale
image into a binary image. Fabio Scotti has used
Otsu’s method in nucleus and cytoplasm selection
in lymphoblasts and lymphocytes [14]. Their
experiments have showed a good performance of
this method in separating the nucleus from the
cytoplasm.
2.1.5. Shadowed C-Means clustering (SCM)
SCM is a method of partitive clustering developed
in the framework of shadowed sets. Unlike rough
clustering, in SCM, the choice of threshold
parameter is fully automated and the number of
clusters is optimized in terms of various validity
indices [15]. Shadowed clustering can handle
overlapping among clusters efficiently and also it
Journal of Theoretical and Applied Information Technology
30th September 2020. Vol.98. No 18
© 2005 – ongoing JATIT & LLS
ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195
3743
can model uncertainty in class boundaries [16]. The
algorithm is robust in the presence of outliers too.
However, fuzzy c-means clustering have problems
with high dimensional data sets and a large number
of prototypes [17].
2.1.6. HSI colour based segmentation
HSI (Hue, Saturation and Intensity) is a common
colour model used in image segmentation. HSI
colour model has a good capability of representing
the colours of human perception [18]. Nor Hazlyna
and the team have conducted a research for ALL
detection based on segmentation using HSI and
RGB colour space [19]. The results have shown
that the proposed segmentation technique based on
HSI has successfully segmented the acute leukemia
images while preserving significant features and
removing background noise. Singhal and Singh
[20] and Halim and his colleagues [21] are some
research groups who have used HSI colour based
segmentation in ALL diagnosis. They have used
HSI colour based segmentation as it provides better
performance than RGB colour segmentation.
2.1.7. K-Means clustering
K-means clustering is an unsupervised learning
algorithm which involves two simple processes as
relegating the given data set and classifying the
colligated data sets to the centroid nearest to them.
K-means clustering segmentation have been used in
identifying the leukemia sub types [22,23] and in
AML screening systems [24]. K-means clustering
does not give classification with labelled data and
also not applicable on incremental data [9].
2.1.8. Morphological operations (shape-based)
Segmentation using morphological operations is a
technique considering the processing of geometrical
structure based on set hypothesis, topology, lattice
hypothesis and arbitrary functions etc. This is the
most successful segmentation method that has been
used so far. Through this method it is very easy for
detecting white cells, overlapping of cells and
shape of cells. Thus this is based on statistics so can
get approximate results.
Bhattacharjee and Saini [25], Vaghela et al.[9]
and Raje and Rangole [26], Shyalika, Kumara and
Kottahachchi [27] are among the researchers who
have used morphological based image segmentation
in leukemia diagnosis. They have discovered that
the morphological operators used for the extraction
of features have resulted in high segmentation
accuracy. Segmentation using morphological
operations has been used in morphological
classification of Leucocytes by microscopic images
[13,28]. In these researches, the researchers have
focused on reducing the problem of identification
and classification of WBC types in microscope
images using morphological operations.
Mostly used segmentation techniques in ALL
diagnosis and the advantage/merits and
disadvantages/demerits of the varied techniques
have been highlighted in Table 1. As per the
evaluation results in earlier researches,
Morphological/shape-based segmentation can been
verified as the best method for image segmentation
in ALL detection.
Table 1: Merits and Demerits of Segmentation
Techniques.
Method Merits Demerits
Watershed
transform
Easy method
for detection
of white cells
It cannot give
accurate result
and cannot
implement on
each and
every image
K-means
clustering
It is used for
clustering and
separate the
data based on
value of K.
It does not
give
classification
with labeled
data and also
not applicable
on
incremental
data.
Edge detection
using histogram
equalizing
method and
linear contrast
stretching
This is very
useful method
to detect white
cell and for
contrast
enhancement.
It is hard to
define
boundary of
overlapping
cell.
Shadowed
C-means
clustering(SCM)
SCM can
handle
overlapping
among
clusters
efficiently and
can model
uncertainty in
Class
boundaries.
Robust in the
presence of
outliers.
Have
problems with
high
dimensional
data sets and a
large number
of prototypes.
Shape based
features
Very easy for
the detection
of white and
overlapping
cell and shape
of cell
This is based
on statistics so
can get
approximate
result.
Journal of Theoretical and Applied Information Technology
30th September 2020. Vol.98. No 18
© 2005 – ongoing JATIT & LLS
ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195
3744
2.2. Classification Methods
Classification in Machine Learning and
Statistics is a supervised learning approach in
which the program learns from the input data and
then uses this learning to classify new observations.
It is in charge of assigning to the unknown test
vector for new observations which is a label from
one of the known classes [29]. Mostly used
classifiers are discussed in this section.
2.2.1. Support vector machine (SVM)
SVM is a discriminative classifier that is formally
defined by a separating hyper plane. When labelled
training data is given (supervised learning), the
algorithm outputs an optimal hyperplane which
categorizes new examples. Dodandeniya, Kumara
and Kulasekara in their research on Automated
Blood Counter [8] have used SVM to separate the
white cells and form red cells. Patel and Mishra
[30] is a research group who presented an
automatic approach for leukemia detection using
microscopic images. Colour, geometric, shape and
statistical features have been analyzed and
classified under the SVM classifier in the intention
of grouping the normal and abnormal cells. SVM
has been used to classify leukemia types too. A
three-layered framework consists of feature
extraction, coding, and classification for the
detection of leukemia from blood smear images has
been proposed by Faivdullah and his colleagues
[31] leukemia types. They have employed a one-vs-
all technique to convert SVM, which is a binary
classifier in to a multi-class classifier.
2.2.2. Artificial neural network (ANN)
This is a statistical learning algorithm defined by an
interconnected set of nodes that are similar to the
network of neurons found in brain. ANNs are
capable of pattern recognition and machine
learning, thus is mainly used in generating and
estimating the output from a large number of input
data set [25]. Mohapatra and the colleagues [32]
have engaged in another project in Lymphocyte
image segmentation using Functional Link Neural
Architecture for ALL detection [32]. Fatma and
Sharma [22] have tried on a system to identify and
classify sub types of acute leukemia using neural
network.
2.2.3. CART (classification and regression
trees)
CART (Classification and Regression Trees)
statistical method has been used in automatic
leukemia diagnosis in investigating the
classification power of cell markers extracted in
segmentation [33]. This method generates
classification tree diagrams with complete splitting
information at each node and then produces a
classification matrix, splitting cost and probability
matrix for both the learning sample and the cross
validation. The classification trees can be saved and
used in classifying unknown specimens. Serbouti
and the research team has employed CART in their
research done in automatic leukemia diagnosis [33].
2.2.4. K-Nearest neighbour (KNN)
This is considered to be the best classifier in the
family of nonparametric method with a good
scalability. In leukemia detection kNN=1 is
considered to classify between blast cells and
normal lymphocytic cells [25, 27]. Bhattacharjee
and Saini [25] in their research in diagnosing ALL
have discovered that KNN is the best classifier that
produced high specificity and also have the lowest
computational complexity which has produced a
specificity of 95.23%.
2.2.5. Ensemble of classifiers (EOC)
Ensemble methods are machine learning algorithms
that construct a set of classifiers and then classify
new data points by taking a weighted vote of their
predictions [34]. EOC improves of the performance
of individual classifiers. The ultimate goal of
classification result integration algorithms is to
generate more certain, precise and accurate system
results. But EOC possess some limitations also
such as increased storage, increase the number of
computations and decreased comprehensibility.
EOC is been an efficient classification model used
in leukemia diagnosis so far. An ensemble classifier
system for early diagnosis of ALL has been
developed by Mohapatra and group in 2014 [16].
As the results they have obtained more accuracy in
EOC in comparison with other classifiers
employed. Scotti and Piuri [28] have used ensemble
of classifiers on their research done in
Morphological Classification of Blood Leucocytes
by Microscope Images. The classification accuracy
has been tested and a proper classifier has been
chosen from a set of candidates of different
classifiers.
When analysing the overall results of early
researches in terms of the classification techniques
used, it was understood that K-Nearest Neighbor,
Support Vector Machines (SVM) and using
Ensemble of classifiers have resulted better results
than other identified classification metrics. The
identified precipitate is summarized as in Table 2.
The classification technique and the feasibility of
Journal of Theoretical and Applied Information Technology
30th September 2020. Vol.98. No 18
© 2005 – ongoing JATIT & LLS
ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195
3745
diagnosing ALL has been evaluted and has been
presented in Table 2 as a good or weak classifcation
algorithm to employ in the developing solution.
Table 2: Summary on Classification Techniques.
Classification Technique Good Weak
KNN
SVM X
ANN X
NB X
RFBN X
CART X
EOC
The accuracy of the ALL diagnosing applications
incredibly depend on the segmentation and
classification methods used by the researchers. As
per the literature review it was revealed that
segmentation using morphological operations has
produced good results and in image classification,
K-Nearest Neighbor (KNN), Support Vector
Machines (SVM) and using Ensemble of Classifiers
(EOC) has given more accurate results. Most of the
research groups have tried on detecting leukemia in
isolated lymphoblasts. Recognition and splitting of
overlapping red blood cells (RBC) with WBC has
yet been a challenging issue in diagnosing
leukemia. Furthermore, blood films lead to quality
devastation when storing them for a long period.
When blood films are being transported to foreign
countries and long distances the quality of them
degraded and important features of them get lost.
Current researches have not yet been extended to
detect this effect in diagnosing leukemia.
Researches must be proposed in order to achieve
these challenging issues which are the current live
problems for the hematologists in diagnosing
leukemia.
3. METHODOLOGY AND EXPERIMENTAL
DESIGN
An automated diagnosing application
would be a beneficial tool in diagnosing of ALL in
peripheral blood samples efficiently and accurately.
This section presents the methodological approach
and the system design of the proposed solution
developed in order to address the identified
research limitations in ALL diagnosis. Ethical
clearance regard to this research has been obtained
prior to the initiation of study from the Ethical
Review Committee (ERC) of Faculty of Medicine,
Kotelawala Defence University, Sri Lanka. The
basic method proposed for diagnosis proposed can
be divided into steps as in Figure 2.
Figure 2: Basic Diagram of the System
The input for the system is the Leishman’s
stained blood slide image, and the lymphocytes in
the image are cropped and individuated manually.
Firstly, in the pre-processing module, the image
acquisition noise and background non-uniformities
are removed. Secondly, image segmentation is
performed using proper segmentation techniques.
This is done using four consecutive steps of
background removal, detection of overlapped cells,
separating the lymphocytic cell and separating the
nucleus region which has been described in this
paper. In the feature extraction module, various
morphological features are being sorted differently
using the segmented regions of the lymphocytic cell
and the nucleus. Combining the features of both
cell and nucleus, some new features are also
calculated. In the lymphocyte classification module,
the tested cells are labelled as blast or normal using
the implemented supervised learning classification
method. The block diagram of the proposed
algorithm is shown in Figure 3.
Journal of Theoretical and Applied Information Technology
30th September 2020. Vol.98. No 18
© 2005 – ongoing JATIT & LLS
ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195
3746
3.1. Image Acquisition
The inputs for this automation process are
microscopic images obtained from peripheral blood
films which stained by Leishman’s that has been
obtained from Department of Pathology Clinic at
Faculty of Medicine, University of Colombo, Sri
Lanka. All the obtained images are affected from
B-ALL precursor which is a major type of ALL.
The images are captured from two different camera
sources as Huawei GR5 2017 smartphone camera
and Canon camera in the same lightning conditions,
resolution and magnification. The slides are placed
under a MicroTech XSZ-N207 microscope in X100
magnification. 142 of the chosen images are taken
into the experiment.
3.2. Image Pre-Processing
Pre-processing is essential as normal
images consist of excessive staining and shadows.
Image enhancement, which is used to bring out the
image details that are obscured is the main task of
this stage. Following three main tasks are
performed in this stage.
3.2.1 FastNlMeansDenoisingColoured
technique
This is done to remove noise and excess blurriness
that is presented in coloured blood images.
3.2.2 Edge enhancement
Done in order to sharpen the image by cleaning the
cell/cell segments in the boundary of the blood
images.
3.2.3 RGB splitting
The RGB image is split in to three channels; green,
red and blue in order to identify the red blood cells
and white blood cells separately.
3.2.4 Removing the green channel
Green channels are mostly sensitive to red blood
cells. Thus, it is removed from the image in this
step. After removing green channel, red and blue
channels are merged.
3.3. Image Segmentation
Segmentation process partitions an image
to its constituent segments or objects known as
pixels. This locates objects and boundaries (curves
and lines etc.) of images and modifies the
representation of an image into somewhat that is
more meaningful and easier to analyze. This is a
crucial step as the following feature extraction and
classification results are much related with the
result of the segmentation module. In this stage the
following four steps are under gone.
3.3.1 Background removal
In this stage, canny filter is first used to reconstruct
the border of the cells present in the image. Then
morphological operation ‘dilation’ is done using a
prepared structuring element. Then ‘closing’ is
done. Combing the images obtained from dilation
and closing, a new image is obtained. Next,
threshold to Zero and Inverted thresholding is
performed to the image obtained from pre-
processing. Then the resulted image is combined
with the new image obtained from morphological
segmentation and the background is now removed.
3.3.2 Detection of overlapped cells
The overlapped cells of the image are detected
using a combination of grab cut algorithm and
circles detection algorithm.
3.3.2.1 Grab cut algorithm
Grab Cut is an image segmentation method based
on graph cuts. Grab Cut Algorithm has been used in
this context for separating the background in the
Figure 3. Block Diagram of the Proposed Algorithm
Journal of Theoretical and Applied Information Technology
30th September 2020. Vol.98. No 18
© 2005 – ongoing JATIT & LLS
ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195
3747
overlapped cell regions. Here firstly, a user-
specified bounding box around the blood plasma
region to be segmented was initiated. Then, the
color distribution of the target nucleus and the
background plasma were estimated using a
Gaussian mixture model.
3.3.2.2 Circles detection algorithm
An advancement of Circle Detection Algorithm has
been developed for detecting overlapped cells.
Thresholding, closing, template matching and
finding the local maxima were carried out in this
step.
3.3.3 Isolating the lymphocytic cell
In the resulted image, the largest contour area is
considered to be the area of the cell region. The
image is then subjected to a combination of binary
thresholding and Otsu’s thresholding and a binary
image of the cell is produced. The total blood cell’s
binary image is now ready for feature extraction.
3.3.4 Isolating the nucleus
In this step, firstly the intensity of the original
cropped blood image is increased such that only the
nucleus will be visible in the image. Then
thresholding is done in order to separate the
nucleus. Here a combination of binary thresholding
and Otsu’s thresholding is done. Then the nucleus
region is segmented by subjecting the image to the
background removal step described earlier. Then
the segmented nucleus is converted to binary and it
is now ready for feature extraction. The image
processing techniques listed in Image processing
and segmentation stages have been used after
applying and testing the image visibility and
processing functionality of them with the
acquisitioned blood images.
3.4. Feature Extraction
In feature extraction, the acquired data
from the image is transformed and labelled to a
particular set of features, which is going to be used
for further classification. The binary equivalent
images produced by the segmentation technique of
blood cell and cell nucleus are used to extract those
morphological features. Using the extracted
features of blood cells and nucleus, combined
features also have been acquired. The features
extracted have been explored in Table 3. Feature
parameters were gathered with respect to four
categories; colour features, geometric features,
texture features and statistical features as shown in
Table 3 in detailed. As for the color features; mean
color values of blood cell, cytoplasm and nucleus
were extracted. Geometric features; area, perimeter,
circularity, diameter, roundness, compactness, form
factor, major axis length, minor axis length,
convexity, solidity, ratio of area of cytoplasm to
nucleus and ratio of area of nucleus to cell were
extracted with respect to the cell and the nucleus.
Entropy, energy, correlation, homogeneity was
extracted features regard to the Texture of the cell
and nucleus. Several statistical parameters;
skewness, variance, mean, gradient matrix were
extracted for the cell and nucleus separately. The
final feature set includes 20 features including; cell
area, cell perimeter, cell circularity, cell diameter,
cell roundness, cell compactness, nucleus area,
nucleus perimeter, nucleus circularity, nucleus
diameter, nucleus roundness, nucleus compactness,
nucleus form factor, nucleus to cell ratio,
convexity, solidity, major axis length, minor axis
length, ratio of area of cytoplasm to nucleus and
ratio of area of nucleus to cell.
Table 3: Parameters Obtained in Feature Extraction.
Feature Parameters Extracted
Colour
features
Mean colour values
Geometric
features
Area, Perimeter, Circularity,
Diameter, Roundness, Compactness,
Form Factor, Major axis length,
Minor axis length, Convexity,
Solidity, Ratio of area of cytoplasm
to nucleus, Ratio of area of nucleus
to cell.
Texture
features
Entropy, Energy, Correlation,
Homogeneity etc.
Statistical
features
Skewness, Variance, Mean,
Gradient matrix etc.
Contour detection was applied in feature
extraction stage for the binary images of cell and
nucleus resulted in segmentation. Generally,
contours are curves that join all the continuous
points along the boundary that have same color or
intensity. This concept is very useful for shape
analysis and object detection and recognition which
was used as guideline for feature extraction step.
The features of final dataset are extracted as
follows. The features were gathered separately for
cell and nucleus.
Area- Area is the total number of non-zero (white)
pixels available within the image region. To
calculate this, the contours were selected, sorted
according to area, the largest contour of the
segmented cell was taken and contour area was
calculated using OpenCV function;
Journal of Theoretical and Applied Information Technology
30th September 2020. Vol.98. No 18
© 2005 – ongoing JATIT & LLS
ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195
3748
cv2.contourArea and determining its bounding
rectangle.
Perimeter- Perimeter is the distance (major and
minor both) between successive boundary pixels. It
was obtained by using cv2.arcLength function on
the first largest contour.
Circularity- This was calculated by area and
perimeter.
Circularity = (4 * pi * Area / Perimeter2)
Diameter- Was calculated using cv2.
minEnclosingCircle function on the first largest
contour.
Roundness- Is a feature representing the degree to
which a shape is round.
Roundness= (4 * Area) / pi * Diameter*Diameter
Compactness- It is a numerical measure
representing the degree to which a shape is
compact.
Compactness = ((4/pi) *Area)1/2/Maximum
Diameter
Form Factor- It is calculated by area and
circularity of both blood cell and cell nucleus.
Form Factor = ((CN Area / BC Area) * (BC
Circularity / CN Circularity)
Where, CN = Cell Nucleus; BC = Blood Cell;
Major axis length- Length in pixels of the major
axis of the ellipse containing the nucleus.
Minor axis length- Computed as the length in
pixels of the minor axis of the ellipse containing the
nucleus.
Convexity- When calculating the convexity, firstly
a convex hull was obtained using cv2.convexHull
(firstlargestcontour_nuclues) function. Then
cv2.contourArea function was used to obtain the
convexity.
Solidity- Area / Convex Area
Ratio of area of cytoplasm to nucleus.
Ratio of area of nucleus to cell.
The selected feature set was visualized using
Seaborn and matplotlib plot and statistically
analyzed, which has been explained under
evaluation section.
3.5. Image Classification
The features extracted in feature extraction
stage were directed to the image classification stage
which employs a supervised learning classification.
For classification, K-Nearest Neighbour (k-NN)
algorithm is used. k-NN algorithm is a non-
parametric method used in pattern recognition for
classification and regression. This classifier has
been used in classification of the cells as ALL
affected or healthy. In k-NN classification, the
output is generally a class membership. In this
context, the data points were classified by a
majority vote of its neighbors, with the points being
assigned to the class most common among its k
nearest neighbors. The label 1 has been assigned for
detected acute leukemic cells and label 0 has been
assigned for healthy cells respectively. Statistical
data visualization on classified results was done
using Seaborn and matplotlib plot analysis.
Figure 4: ALL Diagnosing Window
Journal of Theoretical and Applied Information Technology
30th September 2020. Vol.98. No 18
© 2005 – ongoing JATIT & LLS
ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195
3749
4. EVALUATION AND RESULTS
There are two main objectives that this
research is focused on. One is to develop an
algorithm to compare the image qualities of purely
isolated cells and overlapping cells. Second one is
to detect the quality devastation in blood films in
terms of storing them for prolonged period. The
proposed algorithm was implemented using python
programming language using the OpenCV package
for python. Figure 4 depicts the design of a user
interface in the GUI suite done using Qt. Figure 8
and Figure 9 are the results that were obtained for
one instance in the feature extraction respectively
for the cell and the nucleus. Figure 5 presents an
image of the microscopic image used for testing.
Figure 6 elaborates a results of an isolated
lymphocyte image used for segmentation stage of
the cell and nucleus. In the visualization, cell
segmentation steps and nucleus segmentation steps
are shown separately.
In cell segmentation steps, the pre-processed
image shows the result of the steps
FastNlMeansDenoisingColoured technique, Edge
enhancement, RGB Splitting and removing the
green channel respectively. Next canny, dilation,
closing and combination of dilated and closed
images, thresholding for pre-processed image and
binary image of the cell has been shown. In nucleus
segmentation step firstly, original image is shown.
Then the thresholding mask image shows the result
of the combination of the binary and Otsu’s
segmentation performed. Next canny, dilation,
closing and combination of dilated and closed
images, thresholding for pre-processed image and
binary image of the cell has been shown.
Figure 5: Sample Microscopic Image used for Testing.
In the feature extraction module, chosen
geometric features that are appropriate for the data
set with regard to the cell and nucleus had been
extracted separately. The feature set includes 20
features including; cell area, cell perimeter, cell
circularity, cell diameter, cell roundness, cell
compactness, nucleus area, nucleus perimeter,
nucleus circularity, nucleus diameter, nucleus
roundness, nucleus compactness, nucleus form
factor, nucleus to cell ratio, convexity, solidity,
major axis length, minor axis length, ratio of area of
cytoplasm to nucleus and ratio of area of nucleus to
cell. Exploratory data analysis was performed to
analyze relation between each and every feature
variable. Seaborn pairs plotting was used to
discover patterns in the feature dataset by
considering distribution of single variable
(univariate analysis) and relationships between two
variables (bivariate analysis). The plotted results on
a selected sample for a selected feature set is shown
in Figure 7.
Figure 6: Segmentation Results of the Cell and Nucleus
Journal of Theoretical and Applied Information Technology
30th September 2020. Vol.98. No 18
© 2005 – ongoing JATIT & LLS
ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195
3750
The features extracted with respect to a one
lymphocytic cell and nucleus are depicted in Figure
8 and Figure 9. The features have been saved in a
csv file and continued for use in classification
stage.
The proposed algorithm has been tested and
developed to detect features of overlapping cells
and to detect the quality devastation that occurs in
old blood films when they are kept for 3-6 months.
In detecting overlapping cells, firstly the
overlapping property was identified from the
acquisitioned images. Then overlapped cells were
isolated using the implemented algorithm, then they
were treated as separate individual cells and
continued for ALL diagnosing algorithm proposed.
For convenience, overlapping cells that have two
overlapped cells were taken into consideration in
the first phase. These two blood cells were
observed in two layers in the blood image when the
3D bilateral layers were onserved. The summary of
overlapped cells and isolated cells contributed to
the evaluation is added to Table 4. Figure 10
depicts how the overlapping cells were detected and
identified in the developed algorithm. In order to
provide a solution for the quality devastation of
blood films, images of blood films were taken in 2
attempts keeping a time gap of 6 months in
between. The algorithm was modified taking
normalized values that match with both the fresh
and old blood films.
The details of the blood images used are as
follows in Table 4. The microscopic images were
obtained from two ALL affected patients and one
healthy person respectively in consecutive two
rounds as shown in Table 4.
Figure 7: Exploratory Analysis on Selected Sample
Figure 8: Feature Extraction Results of Cell Figure 9: Feature Extraction Results of Nucleus
Journal of Theoretical and Applied Information Technology
30th September 2020. Vol.98. No 18
© 2005 – ongoing JATIT & LLS
ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195
3751
The number of overlapped cell instances and
isolated cell instances obtained from the same are
included in Table 4. Total number of 2184 images
(ALL-1274, Healthy-910) were used as training
dataset and 936 (ALL-546, Healthy-390) images
were reserved for testing purpose as represented
detailed in Table 5. Evaluation matrix / confusion
matrix was built to evaluate the performance of the
classification model. The evaluation matrix resulted
from the evaluation is presented in Table 6 where
the actual class and predicted class is weighed
accordingly. The results were categorized into four
partitions; correct diagnosis of healthy, error
diagnosis of healthy, error diagnosis of ALL and
correct diagnosis of ALL. Table 7 includes the
legend of the evaluation matrix.
Table 5: Number of Images used in Training and Testing.
Number of
microscopic images
obtained
Patient 1
(ALL affected)
Patient 2
(ALL affected)
Patient 3
(Healthy)
Isolated
cells
Overlap
cells
Isolated
cells
Overlap
cells
Isolated
cells
Overlap
cells
First Round 350 100 300 180 420 220
Second Round 300 150 240 200 410 250
Total Images 650 250 540 380 830 470
ALL Affected vs
Healthy
1820 1300
Training &
Testing Datasets
Total
Number
of images
used
ALL
affected
Healthy
Training Data set 2184 1274 910
Testing Data set 936 546 390
Figure 10: (a) Overlapped cells (b) Isolation of overlapped cells
Table 4: Summary of Number of Images Obtained.
(a) (b)
Journal of Theoretical and Applied Information Technology
30th September 2020. Vol.98. No 18
© 2005 – ongoing JATIT & LLS
ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195
3752
Table 6: Evaluation Matrix with the Total Results
Obtained.
Table 7: Legend of Evaluation matrix.
As per the evaluation results, it was able to achieve
an average accuracy of 88.8% in diagnosing ALL.
The accuracy was calculated as ratio between the
total number of correctly diagnosed ALL or healthy
results and the total samples tested.
Accuracy = TP + TN x 100
TP+TN+FP+FN
= 88.8%
The error rate/misclassification rate was calculated
as the ratio between the total incorrectly diagnosed
samples to the total tested samples as follows.
Misclassification Rate = FP + FN x 100
TP+TN+FP+FN
= 11.11%
Overall, the algorithm holds 11.11% of
misclassification rate. Accordingly, there seem to
have some place for improve, when considering the
quality of blood images obtained, environment
conditions, normalizing indicators etc.
To get the value of precision, the total number of
correctly diagnosed ALL effected positive samples
was divided by the total number of predicted
positive examples. High Precision resulted with a
less number of error diagnosis of healthy, indicates
the results labelled as positive ALL diagnosed is
indeed positive.
Precision = TP
TP+FP
= 0.93
The Recall was defined as the ratio between the
total number of correctly diagnosed ALL effected
positive examples to the total number of positive
samples obtained by the addition of correctly
diagnosed ALL affected and error diagnosis of
ALL. This resulted in considerable high recall
(0.93%) which results in the success of the
classifier model.
Sensitivity / Recall = TP
FN+TP
=0.93
The end product of the research comprises dual
components interconnected; a Python based Qt GUI
enabled development suite and a PHP based web
application. Python suite performs main targeted
backend functionalities and web application that
serves hematologists, doctors and patients to
perform utility functions. The end system includes
modules for login and authentication, user
registration, main window, ALL diagnosing, patient
information recording, records viewing, diagnosed
results plotting, database functionalities, report
generation and diagnosed reports sending for
patients via email.
5. CONCLUSION
The developed system has been extended
to have good results in automatic diagnosis of the
disease in the acquired human blood samples. It has
obtained an average accuracy of 88.8% in
diagnosing ALL in human blood samples. The
proposed algorithm can also be further developed to
detect the granules and intra cellular components
inside the cell. Using more than one classifier in the
aim of increasing the accuracy of the proposed
algorithm has been identified as a further work in
this research. As per the statistical data published in
Evaluation
Matrix
Predicted
Class
Actual
Class
Class=ALL
Class=Healthy
Class=ALL 702
(TP)
52
(FN)
Class=Healthy 52
(FP)
130
(TN)
Name Description
TN True Negative-correct diagnosis of
healthy
FP False Positive- error diagnosis of
healthy
FN False Negative- error diagnosis of
ALL
TP True Positive- correct diagnosis of
ALL
Journal of Theoretical and Applied Information Technology
30th September 2020. Vol.98. No 18
© 2005 – ongoing JATIT & LLS
ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195
3753
future spreading of a cancer like leukemia in the
world, automat ion procedures to detect leukemia
has become an urgent need. Governments,
especially in developing countries like South Asia
would find these automatic leukemia diagnosing
systems as cost effective solutions to implement in
hospitals.
ACKNOWLEDGEMENTS
The research is financially supported by
Mobitel (Pvt) Ltd, Sri Lanka. Authors would
greatly appreciate and acknowledge their
contribution, whose dedicated assistance for the
publication that should be highly recognized.
The corresponding author would
acknowledge the supervisors Mr. PPNV Kumara
and Dr. Darshana Kottahachchi whose insight and
expertise that greatly assisted the research. Special
thanks goes to all the lecturers and staff of Faculty
of Computing and Department of Medical
Laboratory Sciences of Faculty of Allied Health
Sciences at General Sir John Kotelawala Defence
University, Sri Lanka whose loyal support for the
research project was highlighted in all stages of the
research. Authors are also grateful to the staff of
Department of Pathology Clinic at Faculty of
Medicine, University of Colombo, Sri Lanka who
supported by providing research materials and
Faculty of Medicine at General Sir John Kotelawala
Defence University, Sri Lanka for granting ethical
approval for the research.
AVAILABILITY OF DATA AND MATERIAL
The dataset used to support the findings,
source code and the executable file developed in
this research project are available from the
corresponding author upon request.
REFRENCES
[1] Catovsky D, Hoffbrand AV. Acute leukemia.
In: Hoffbrand AV, Lewis SM, Tuddenham
EDG (eds), “Postgraduate Haematology”, 4th
edition, Arnold: London, 2001, pp 373–404.
[2] Bennett JM, Catovsky D, Daniel MT, Flandrin
G, Galton DA, Gralnic HR, Sultan C,
“Proposals for the Classification of Acute
Leukemia”, Br J Haematology, 33: 451-8,1976.
[3] Hughes-Jones NC, Wickramasinghe SN, Hatton
CSR, “Lecture notes on Hematology”, Seventh
Edition, Wiley Blackwell Publishing 2004
[4] Purohit SS, “Biotechnology Fundamentals and
Applications”, Third Edition,2000.
[5] Cui J, Wang J, He K, Jin B, Wang H, Li W,
Kang L, Hu Li MH, Yu M, Shen B, Wang G
and Zhang X, “Proteomic analysis of Human
Acute Leukemia Cells: Insight into Their
Classification, Clinical Cancer Research”
Volume. 10, October 15,2004, pp. 6887 – 6896.
[6] Scholl I, Aach T, Deserno TM, Kuhlen T,
“Challenges of Medical Image Processing.”
Computer Science - Research and
Development, Volume 26, Issue 1-2, 2011, pp.
5–13, doi:10.1007/s00450-010-0146-9.
[7] Hegadi RS, “Image Processing: Research
Opportunities and Challenges.”, National
Seminar on Research in Computers (NSRC),
13-12-2010, Bharathiar University, Coimbatore,
India December, 2010.
[8] Dodandeniya JMDGCM, Kumara PPNV, and
Kulasekara DMR, “Automated Blood Counter
(ABC): An Image Processing Solution”,
Proceeding of International Symposium of
General Sir John Kotelawala Defence
University, September, 2016
[9] Vaghela HP, Modi H, Pandya M, Pandya M,
“Leukemia Detection using Digital Image
Processing Techniques”, International Journal
of Applied Information Systems (IJAIS)–ISSN
2249-0868, Volume 10, Issue 01, November,
2015.
[10] Patil BG, Jain SN, “Cancer Cells Detection
Using Digital Image Processing Methods.”
International Journal of Latest Trends in
Engineering and Technology (IJLTET), Volume
3, Issue 4, March, 2014.
[11] Mahaja S, Golait SS, Meshram A, Jichlkan N,
“Review: Detection of Types of Acute
Leukemia”, International Journal of Computer
Science and Mobile Computing, Volume 3
Issue 3, March- 2014, pp. 104-111.
[12] Viswanathan P, “Fuzzy C Means Detection of
Leukemia Based on Morphological Contour
Segmentation”, Procedia Computer Science 58
(2015) 84 – 90, Second International
Symposium on Computer Vision and the
Internet (VisionNet’15), Volume 58, pp. 84–90.
doi: 10.1016/j.procs.2015.08.017.
[13] Mohapatra S, Samanta SS, Patra D, Satpathi S,
“Fuzzy Based Blood Image Segmentation for
Automated Leukemia Detection”, IEEE
Journal of Theoretical and Applied Information Technology
30th September 2020. Vol.98. No 18
© 2005 – ongoing JATIT & LLS
ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195
3754
International Conference on Devices and
Communications (ICDeCom), 2011, pp. 1–5.
[14] Scotti F, “Automatic morphological analysis for
acute leukemia identification in peripheral
blood microscope images”, IEEE International
Conference on Computational Intelligence for
Measurement Systems and Applications, Italy,
July, 2005, pg. 96–101.
[15] Mitra S, Pedrycz W, Barman B, 2010.
“Shadowed c-means: Integrating fuzzy and
rough clustering”. Pattern Recognition, Volume
43, Issue 4, pg. 1282–1291, doi:
10.1016/j.patcog.2009.09.029
[16] Mohapatra S, Patra D, Satpathy S, An ensemble
classifier system for early diagnosis of acute
lymphoblastic leukemia in blood microscopic
images”, Neural Computing and Applications,
Volume 24, Issue 7-8, 2014, pg.1887–1904.
doi:10.1007/s00521-013-1438-3
[17] Olivera JV, Pedrycz W (Editors), Winkler R
and Klawonn F and Kruse R (Writers),
“Advances in fuzzy clustering and its
applications”, Wiley, Chichester, 2007.
[18] Cheng HD, Jiang XH, Sun Y and Wang JL,
“Color Image Segmentation: Advances &
Prospects”, Department of Computer Science,
Utah State University, Logan, 2006.
[19] Nor Hazlyna H, Mashor MY, Mokhtar NR,
Aimi Salihah AN, Hassan R, Raof RAA,
Osman MK, “Comparison of Acute Leukemia
Image segmentation using HSI and RGB color
space”. 10th International Conference on
Information Science, Signal Processing and
their Applications (ISSPA 2010) (IEEE),2010,
pp. 749–752, doi:10.1109/ISSPA.2010.5605410
[20] Singhal V, Singh P, “Local Binary Pattern for
automatic detection of Acute Lymphoblastic
Leukemia”, 2014 Twentieth National
Conference on Communications (NCC), IEEE,
2014, pp. 1–5.
[21] Halim NHAbd, Mashor MY, Nasir ASA,
Mokhtar NR, Rosline H, “Nucleus segmentation
technique for acute leukemia”, 2011 IEEE 7th
International Colloquium on Signal Processing
and Its Applications (CSPA), IEEE, 2011, pp.
192–197.
[22] Fatma M, Sharma J, “Identification and
Classification of Acute Leukemia using Neural
Network”, 2014 International Conference on
Medical Imaging, m-Health and Emerging
Communication Systems (MedCom), IEEE,
2014, pp.142–145.
[23] Rejintal A, Aswini N, “Leukemia Cancer Cell
Detection using Image Processing”,
International Journal of Advanced Research in
Electrical, Electronics and Instrumentation
Engineering, 4th National Conference on
Design Innovations for 3Cs “Compute-
Communicate-Control”, Volume 5, Special
Issue 6, Dept. of ECE, MVJ College of
Engineering, Bangalore-560067, India, July
2016.
[24] Agaian S, Madhukar M, Chronopoulos AT,
“Automated Screening System for Acute
Myelogenous Leukemia Detection in Blood
Microscopic Images”, IEEE Systems Journal,
Volume 8, Issue 3, 2014, pp. 995–1004, doi:
10.1109/JSYST.2014.2308452
[25 Bhattacharjee R, Saini LM, “Robust Technique
for the Detection of Acute Lymphoblastic
Leukemia”, 2015 IEEE Power, Communication
and Information Technology Conference
(PCITC), IEEE, Siksha ‘O’ Anusandhan
University, Bhubaneswar, India, 2015, pp. 657–
662.
[26] Raje C, Rangole J, “Detection of Leukemia in
Microscopic Images using Image Processing”,
2014 International Conference on
Communications and Signal Processing
(ICCSP), IEEE, India, April 3-5, 2014, pp. 255–
259.
[27] Shyalika JKC, Kumara PPNV and,
Kottahachchi DU, “An Image Processing
Application for Diagnosing Acute
Lymphoblastic Leukemia (ALL)”, 10th
International Research Conference, General Sir
John Kotelawala Defence University, Sri Lanka.
General Sir John Kotelawala Defence
University, Sri Lanka, August, 2017, p.7.
Available at: http://ir.kdu.ac.lk/handle/345/1695
[28] Piuri V, Scotti F, “Morphological
Classification of Blood Leucocytes by
Microscope Images”, CIMSA 2004 - IEEE
International Conference on Computational
Intelligence for Measurement Systems and
Applications, IEEE, Boston, MA. USA, 14-16
July 2004, pp. 103–108.
[29] Rege MV, Dr. Gawli BW, “Detection of
Leukemia in Human Blood Sample through
Image Processing: A Review”, International
Journal of Modern Trends in Engineering and
Research, Volume 02, Issue 10, October 2015,
pp.363-369.
[30] Patel N, Mishra A “Automated Leukaemia
Detection Using Microscopic Images”, Procedia
Computer Science, Volume 58, Second
Journal of Theoretical and Applied Information Technology
30th September 2020. Vol.98. No 18
© 2005 – ongoing JATIT & LLS
ISSN: 1992-8645 www.jatit.org E-ISSN: 1817-3195
3755
International Symposium on Computer Vision
and the Internet (VisionNet’15), pp. 635–642,
doi: 10.1016/j.procs.2015.08.082.
[31] Faivdullah L, Azahar F, Htike ZZ, Naing
WYN, “Leukemia Detection from Blood
Smears”. Journal of Medical and
Bioengineering, Engineering and Technology
Publishing, Volume 4, No. 6, December 2015,
pp. 488–491, doi:10.12720/jomb.4.6.488-491.
[32] Mohapatra S, Patra D, Kumar S, Satpathy S,
“Lymphocyte Image Segmentation using
Functional Link Neural Architecture for Acute
Leukemia Detection”. Biomedical Engineering
Letters, Volume 2, 2012, pp. 100–110,
doi:10.1007/s13534-012-0056-9.
[33] Serbouti S, Duhamel A, Harms H, Gunzer U,
“Image Segmentation and Classification
Methods to Detect Leukemias”, Proceedings of
the Annual International Conference of the
IEEE Engineering in Medicine and Biology
Society, Volume 13, Issue 01, France, 1991, pp.
260–261.
[34] Dietterich TG, “Ensemble Methods in Machine
Learning”, Oregon State University, Corvallis,
Oregon, USA, 2011.
