automatic red blood cell counting based on...
TRANSCRIPT
AUTOMATIC RED BLOOD CELL COUNTING BASED ON PARTICLE AREA
RABIAHTULADAWIAH BINTI MUSA
A project report submitted in partial
fulfillment of the requirement for the award of the
Master of Electrical Engineering
Faculty of Electrical and Electronic Engineering
Universiti Tun Hussein Onn Malaysia
JUNE 2015
iv
ABSTRACT
In medical area, red blood cell (RBC) counting and analysis contributes important
information in pathological diagnosis regarding disease such as anemia and low level of
hemoglobin concentration. Normally the blood sample is processed in laboratory using
hematology analyzer and blood smear is viewed under microscope. The analysis and
counting process is conducted manually by pathologist. The task is very laborious,
tedious, time-consuming and very dependent to the skill. This project applied image
processing techniques to develop a computer-aided system for automated RBC counting.
The RBC images will be pre-processed in early stage using morphological operations
and thresholding to obtain the images with good quality. Then, the RBC images will be
classified based on the particle area size into single and multi-overlap cells. For counting
the total number of the RBC, mathematical numeric function is used. The accuracy of
the result is determined by doing comparison with the ground truth data. The proposed
method has been tested to the RBC images and performs a reliable system for counting
RBC.
v
ABSTRAK
Dalam bidang perubatan, analisis dan pengiraan sel darah merah menyumbangkan
informasi penting bagi diagnosis patologi berkenaan penyakit seperti anemia dan level
kepekatan hemoglobin yang rendah. Kebiasaannya sampel darah diproses di dalam
makmal menggunakan “hematology analyzer” (iaitu sejenis peralatan elektronik yang
menguji bahan secara kimia) dan sampel darah diperhatikan di bawah mikroskop.
Analisis dan proses pengiraan dijalankan secara manual oleh ahli patologi. Tugasan
tersebut sangat rumit, memenatkan, mengambil masa dan sangat bergantung kepada
kemahiran. Projek ini mengaplikasi teknik pemprosesan imej untuk menghasilkan suatu
system berkomputer yang dapat mengira sel darah merah secara automatic. Imej sel
darah merah akan diproses awal menggunakan “morphological operation” dan
“threshold” untuk mendapatkan imej yang berkualiti baik. Kemudian imej akan
dikelaskan mengikut saiz “particle area” kepada satu sel dan sel yang berganda. Bagi
mengira jumlah sel darah merah, fungsi matematik digunakan. Ketepatan keputusan
akan dinilai dan dibandingkan dengan data asal (kiraan sebenar). Kaedah yang
dicadangkan ini telah diuji dan menunjukkan keberkesanannya untuk mengira sel darah
merah.
vi
CONTENTS
TITLE i
DECLARATION ii
ACKNOWLEDGEMENT iii
ABSTRACT iv
CONTENTS vi
LIST OF TABLES viii
LIST OF FIGURES ix
LIST OF SYMBOLS AND ABBREVIATIONS xi
CHAPTER 1 INTRODUCTION 1
1.1 Research background 1
1.2 Problem statement 5
1.3 Aim and objectives 5
1.4 Scope of the project 6
CHAPTER 2 LITERATURE REVIEW 7
2.1 Image processing in RBC segmentation and counting 7
2.2 Summary 12
vii
CHAPTER 3 METHODOLOGY 14
3.1 Research workflow 14
3.1.1 Image acquisition 15
3.1.2 Image pre-processing 16
3.1.2.1 Threshold 17
3.1.2.2 Morphological operations 22
3.1.3 Segmentation and classification 24
3.1.4 RBC counting and validation 26
3.2 Summary 28
CHAPTER 4 RESULT AND ANALYSIS 29
4.1 Pre-processing RBC images 29
4.2 RBC classification and counting based on particle area 32
4.3 RBC counting using connected component labeling
method 40
4.4 Accuracy and error of the results 42
4.4 Summary 52
CHAPTER 5 CONCLUSION 54
5.1 RBC counting result 54
5.2 Recommendation for future works 55
5.3 Summary 56
REFERENCES 57-60
viii
LIST OF TABLES
3.1 Comparison of each RGB channel for thresholding purpose 20
3.2 Range of particle area for each group of RBC 25
4.1 Result of RBC counting based on particle area 37
4.2 Result of RBC counting using CCL method 42
4.3 Ground truth data 46
4.4 Percentage accuracy of RBC counting 48
4.5 Percentage error of RBC counting 50
ix
LIST OF FIGURES
1.1 Red blood cell shape 2
1.2 Normal and sickle RBC 3
1.3 Image of blood smear 4
2.1 Circle Hough Transform accumulation array 9
2.2 Separating overlapped objects using distance transform and
watershed 9
2.3 Morphological reconstruction to remove unwanted object 10
3.1 Work flow 15
3.2 Process of image acquisition from light microscope 16
3.3 Pre-processing flow diagram 17
3.4 Binary image 18
3.5 Tresholding value setting 20
3.6 Comparison of threshold images in different RGB channels 21
3.7 Configure object filter 26
3.8 Classification and counting flow chart 27
3.9 Mathematic numerical model for counting RBC 28
x
4.1 Conversion to binary image 30
4.2 Original RBC image (img1) 30
4.3 RBC image under pre-processing steps 31
4.4 Single RBC detection 34
4.5 Overlap RBC detection 35
4.6 RBC image are tagged in numbering 36
4.7 RBC counting result by classes 38
4.8 Classification and counting result of image img1 39
4.9 Result images of CCL method 41
4.10 Actual count of image img1 43
4.11 Result of original and processed images 44-45
4.12 Particle area method results compared with actual counts 49
4.13 CCL method result compared with actual counts 51
4.14 Comparative chart of measured and actual counts for both methods 52
xi
LIST OF SYMBOLS AND ABBREVIATIONS
RBC - Red Blood Cell
WBC - White Blood Cell
Hb - Hemoglobin
O2 - Oxygen
CO2 - Carbon dioxide
ME - Myalgic Encephalomyelities
MS - Multiple Sclerosis
MATLAB - Matrix Laboratory (A programming software)
Open CV - Open source Computer Vision (A programming software)
NI - National Instrument
AI - Automated Inspection
RGB - Red Green Blue (colour channel)
CHT - Circle Hough Transform
ANN - Artificial Neural Network
CCL - Connected Component Labelling
PCNN - Pulse Coupled Neural Network
MSE - Mean Squared Error
xii
LIST OF APPENDICES
APPENDIX TITLE PAGE
A SAMPLE IMAGES 57-60
B RBC COUNTING RESULT 61
C CONFERENCE PAPER 62-71
D MATLAB CODING 72
1
CHAPTER 1
INTRODUCTION
This chapter will describe the background, problem statement and objectives of the
research. It consist the definition of the terms that being studied such red blood cell and
image processing. This research is applying knowledge of image processing for solving
some issues that exist in medical area especially in the red blood cell analysis.
1.1 Research Background
Blood is a connective tissue consisting of red blood cells (RBC), white blood
cells (WBC), and platelets suspended in plasma. As a medium of transportation for the
whole body, blood composition is very vital to be monitored when it comes to medical
inspection. RBC or erythrocytes are normally in round shape, biconcave and flattened,
about 7 µm in diameter and 2.2 µm thick as shown in Figure 1.1. Unlike WBC, RBC
lack of nuclei and contains of Hemoglobin (Hb) protein to carry oxygen (O2) and carbon
dioxide (CO2) molecules in respiration system. The shape provides an increased surface
area and size for diffusion process between blood and tissues [14].
2
Figure 1.1: Red blood cell shape [22]
RBC can be found in normal and abnormal or sickle shape (Figure 1.2). Normal
RBC is disc-shaped and looks like a doughnut without a hole in the center. It can live
about 120 days. Sickle cells contain abnormal hemoglobin called sickled hemoglobin or
hemoglobin S (HbS). Sickle cells usually indicate anemia disease and the shape is like a
crescent. It tends to block the blood flow in small capillary and lead to severe pain and
organ damage as well as risk of infection [26].
In medical area, RBC analysis contributes information of pathological diseases
and condition. It helps doctors to determine the appropriate treatment to the patient. Any
condition which there is an abnormally low of hemoglobin concentration or red blood
cell count is indicating to anemia [14], low of specific vitamin [7] and might be
secondary effect of several other disorders [24]. The shape of RBC and its deformability
or abnormality has connection to the relevant disease such as Huntington’s desease,
Myalgic Encephalomyelities (ME) and Multiple Sclerosis (MS) [14].
3
However some factors should be taken into consideration when perform the RBC
counting, including level age of people (which is children and adult, younger and older)
and strenuous physical activity [24]. From that, the accurate diagnosis is very important
to determine the exact disease and prepare the correct treatment to the patient.
Figure 1.2: Normal and sickle RBC [26]
Normally the blood sample is taken and processed in the laboratory by using
chemical electronic devices such as hemocytometer or hematology analyzer. This task is
manually done by lab technologist. It is tedious and very dependent to their skill
especially to count and analysis the cell through microscope [7]. The counting and
analysis process is difficult when the cells are overlapped, usually such a finding
neglected, and sometimes other particles such as platelet and WBC interfere too (Figure
1.3). The conventional method is time-consuming to complete and the counting task is
laborious [5].
4
Image processing is very useful and widely used nowadays as alternative for
improving the way of RBC counting and analysis. It improves the effectiveness of the
analysis in term of accuracy, time consuming and cost. Such a procedure has been used
in medical diagnosis for processing blood cell image especially for segmenting RBC
from its complex environment in the blood. Segmentation of cell image is the key of
medical image processing that directly impact the analysis and accuracy of the work
[13]. From the analysis and counting, it will lead the doctor to make a decision of a
patient health status. Various software provide tool for image processing such as
MATLAB [7], Microsoft Visual Studio and Open CV [1], Lab VIEW [2], telemetry
system [3] and in this project, National Instrument (NI) Vision Builder Automated for
Inspection (AI). Most of the techniques can be found in the software, the importance is
the understanding of how to apply it to obtain the desired outcome.
Figure 1.3: Image of Blood Smear
5
1.2 Problem Statement
Current method to count and analysis RBC is using a laboratory device known as
hematocytometer. It consists of a thick glass microscope slide with a rectangular
indentation creating a chamber of certain dimensions. This chamber is etched with a grid
of perpendicular lines. It is possible to count the chamber of cells in a specific volume of
fluid and calculate the concentration of cells in the fluid [20]. To count blood cells, lab
technologist must view hemocytometer through a microscope and count blood cells
using hand tally counter. However, the overlapped blood cells cannot be counted by
using hemocytometer and be neglected. This method is laborious, tedious and time
consuming to complete it. Knowledge, skill and experience will help to identify and
count the cells.
RBCs are not suspended alone in the blood volume thus the issue of how to
distinguish and segment RBCs from other components of blood such as WBCs, platelet
and plasma is always being raised. Besides, those cells sometimes are found overlapping
and clumping to each other. The complicated and high degree of overlapping condition
is very difficult to solve. There is also possibility of having irregular and normal shape
of cells that mix and stick together.
Development of automated system for analysis and counting RBCs in the blood
sample is very useful. It will improve the existing method and help the pathologist and
doctors to diagnose the patients accurately.
1.3 Aim and Objectives
The aim of this project is to count the total number of red blood cell in blood
samples since medical diagnosis concern about that pretty much. The accuracy of the
result for the RBC number in a sample image and fast performance of operation will also
be the aims of this project. From the total count of RBCs, it can help the doctors
determine the disease of the patient. To support the aims, there are objectives that will be
formulated and focused in this project.
6
The objectives of this project are:
1. To develop a method for automatically segment out RBC region in various
environment condition.
2. To construct a procedure for counting the RBC number in overlapping and non-
overlapping condition.
3. To access the performance of the develop method in real condition.
1.4 Scope of the Project
This project will be focusing in the boundary that related to the stated objectives. It
will involve the image processing techniques and come out with the algorithm for
counting the RBC numbers.
7
CHAPTER 2
LITERATURE REVIEW
This chapter will discuss the previous researches that have been done within the same
area of the topic. Literature review is important to tell whether the research is relevant
and feasible to be studied. It also can be a guidance and boundary so that the research is
reasonable and can produce some useful findings. The review is done on the techniques
used for processing image and counting the RBC.
2.1 Image Processing in RBC Segmentation and Counting
Image processing is a successive tool to improve the capability of human to
identify and extract the important information or data from images or video by using a
machine language [7]. Image can be defined as a matrix of light intensity level that can
be manipulated or operated by using computer algorithm [10]. Image processing
provides many techniques to increase the effectiveness of decision making, shorten the
consumed time and reduce the cost [4, 18].
8
The use of image processing for RBC analysis has been produced in previous
studies. Many techniques and algorithm have been proposed to achieve the good
findings. Recent studies on this area mainly focus in segmenting and classification of the
cells for counting purpose.
Software based for developing automatic system to analysis RBC can reduce cost
compared to expensive hardware that being used nowadays. There are several examples
of software that commonly used to perform image processing such as MATLAB, Open
CV, and Lab View. This project uses a part of embedded system of NI Lab View which
is Vision Builder. Image processing techniques mostly were utilized to perform pre-
processing, segmentation of the image, classification or identification and counting the
cells.
The images basically are obtained from the samples of blood that being captured
using light microscope which the process involve the preparation of blood smear in the
lab [1, 7]. Blood smear is a process to put the blood specimen on the slide that being
observed under the microscope. The image then will be filtered to reduce or minimize
the noise such as using average or median filter. The conversion of colour from RGB
into other channels helped in solving illumination problem [7] and as a preparation for
the next stage of image processing.
According to Ventalakshmi in 2013 [3] histogram thresholding and
morphological operator is used to segment the image as well to extract or differentiate
RBC from other cells (WBC and platelets) and its background. For RBC counting
process, Circle Hough Transform (CHT) technique is used to find a circle characterized
by a center point and a known radius (Figure 2.1). CHT is modified version from classic
Hough Transform that detect parametric curve in images. Another paper is by Wei et. al
in 2009 [16] that adopting extended HT for cover different sizes of co-center circles and
multiple radii called Multi-Scale Circular Hough Transform (MSCHT).
9
Figure 2.1 : Circle Hough Transform accumulation array [23]
Study by Sharif et. al in 2012 [7] proposed mathematical morphological operator
to segment the image based on elimination of WBC. The step involved erosion, dilation,
opening, closing and reconstruction. To separate the overlapped cells, marker controlled
watershed algorithm is used where it based on the distance of overlapping cells (Figure
2.2). However, it still has weaknesses where it cannot handle the area that contains
important information.
Figure 2.2: Separating overlapped objects using distance transform and watershed
[redrawn using MATLAB]
10
In 2013, Tulsani et. al [6] also implemented the similar method of morphological
watershed transform algorithm for segmentation of the cells (Figure 2.3). The paper also
proposed a technique to eliminate WBC and platelet by using color conversion in
YCbCr format. The Cb component has been chosen as it showed the clear appearance of
WBC and platelets.
An improvement has been made before that by Huang in 2010 [13] that
combined the watershed algorithm with mathematical morphology using corrosion and
expansion algorithm.
Figure 2.3: Morphological reconstruction to remove unwanted object (a) Mask
(b) Marker (c) Result [redrawn from [6]]
A paper from Nguyen in 2010 [12] proposed algorithm for splitting the clumped
cells. It used Otsu’s algorithm for getting threshold in order to separate background and
foreground and for detection of cell central points, it applied Euclidean distance
transform. The advantages of this method are it can reduce the disadvantage of
concavity analysis methods and model-based approach, detect other distorted structures
and independent cell size and the merge process is much simpler compared to watershed
segmentation merge process. Concavity analysis is measuring split lines in overlapped
cells but has the problem that cannot cope with multiple overlapping cells. However,
11
there was still weakness regarding quality of the image such as blur and non-focused
image.
In study by Tomari et.al in 2014 [1], the image was processed through three
main stages which is segmentation, feature extraction and classification. Segmentation
process was done by using adaptive global thresholding. The extracted features such as
moment values, area and perimeter are used to identify overlapping and non-overlapping
region. For classification, artificial neural network (ANN) classifier is used to train and
test the data.
Multilayer perceptron artificial neural network is a method used to classify RBC
as proposed by Jambhekar in 2011 [10]. In the paper too, segmentation process was
performed using histogram thresholding. By removing the overlapped cells, it
successively counted the number of RBC and WBC. The result also showed the
complete health RBC and incomplete non-circular (sickled cells) of RBC. The accuracy
was 81% and increased when training was increased.
Another paper by Poomcokrak in 2008 [18] also implemented ANN to classify
and count RBC. The function was performed by adjusting the weight between the
elements and the mean squared error (MSE) was calculated. The value obtained is used
to separate sickle RBC from normal RBC and the number of normal RBC is counted.
The result shows 74% accuracy of counting RBC.
Natsution in 2008 [17] reported a comparison of two methods for counting RBC.
It used backprojection ANN and connected component labelling (CCL). The result
showed the accuracy was higher (86.97%) by using ANN back projection than using
CCL (87.74%).
Other method used for segmentation and counting process by Adagale in 2013
[4] was Pulse Coupled Neural Network (PCNN) combined with template matching
algorithm since PCNN alone cannot cope with overlapped cells. PCNN consider
overlapped cells as a single cell. Template matching uses a template of RBC to be
matched to the object for separating small cell in shape and size. However the accuracy
decreased whenever the RBCs are overlapped totally because the area of one cell was
considered as a template and the algorithm works only in 100x microscope scale.
12
Contour tracing approach has been used to segment scanning electron
microscope (SEM) images as reported by Vromen in 2009 [15]. The method viewed
contour detection and negotiating perceived problem areas one at a time but it still has
lack when facing overlapping cells. It applied Bayesian tracking framework. The
contours were categorized into three fully correct, minimally obscured and incorrect.
The result showed 95.7% were either unobscured or minimally obscured with 0.6% false
negative and 4.3% false positive.
A study by Ajay in 2014 [2] used Lab View software with Vision Assistant to
apply morphological operation for diagnosis blood cells in real time. It implemented
pattern matching method to identify the abnormal RBC for anemia analysis. It also
developed a man-machine interface so that the procedure could be in real time and user
friendly.
Using Lab View software too, Goyal in 2006 [21] proposed an artificial
intelligent method for diagnosis anemia disease using the concept of fuzzy logic. It has
two parts; find the group of disease using fuzzy interferences and another one using
knowledge base.
From the previous researches, the process that apparently important was
segmentation or classification of the RBC from the blood image. To execute the process,
most of the studies applied the similar method as stated above. They also improved the
classic method for getting the better result or to complement their objective.
In this project, there are two methods will be applied to count RBC number.
First, segmentation and object detection based on particle area and then mathematical
function to count. This method is implemented using NI Vision Builder AI. Second,
finding and counting connected objects in binary image using MATLAB.
2.2 Summary
RBC counting and analysis using image processing is not a new thing in medical
diagnosis. It always goes through improvement as a new research is carried out. The
researchers offered many different methods as long as it can be giving a better accuracy
13
and promising results. However there are still weaknesses and constraints due to the
image itself such as color similarity, weak edge boundary, overlapping condition, image
quality, contrast, brightness, illumination and noise. Thus, more studies must be done to
handle those matters to produce strong analysis approach for medical diagnosis purpose.
This project is hoped can build a better solution and help to improve the current methods
so that it can be more capable, robust, and effective whenever any sample of blood cell
is analyzed.
14
CHAPTER 3
METHODOLOGY
This chapter will discuss the method or technique used to achieve the research aim. The
discussion will be based on the image processing steps used for analysis the RBC image
which has been shown in the project work flow.
3.1 Research Work Flow
The research work flow is showing steps on how to achieve the objective. The
flow is including image acquisition, image pre-processing, segmentation and
classification of cells, RBC counting and validation. The explanation of each step is
included to give a clear understanding of the methods.
15
Figure 3.1: Work flow
3.1.1 Image acquisition
The blood images are captured by using light microscope which the blood
sample or specimen, specifically called blood smear, is viewed under the microscope.
Blood smear is the process of preparing the blood specimen onto the slide to be observed
under microscope. This is done in laboratory. To display the images, the process
involves the digitization of image and interpolation from optical image with 40X
objective (Figure 3.2). Those processes create noise and distortion to the images. Hence,
the images are then need to be pre-processed in the next stage.
Image acquisition
Pre-processing
Segmentation and classification
RBC counting and validation
16
Figure 3.2: Process of image acquisition from light microscope
3.1.2 Image pre-processing
Pre-processing of the images is a stage for improving quality of the images. It is
using digital image processing software that available to process several steps including
color conversion and thresholding, enhancement and filtering, contrast improvement and
brightness correction, and several morphological operation. In this project, National
Instrument (NI) Vision Builder for Automated Inspection (AI) and Vision Assistant and
MATLAB are used as image processing tool. The flow of pre-processing image is
illustrated in Figure 3.3.
17
Figure 3.3 Pre-processing flow diagram
3.1.2.1 Threshold
The original image is in RGB colour. RGB color image has 24 bits (8 bits red, 8
bits green and 8 bits blue) and the problem of this image is backward compatibility with
older hardware that may not be able to display 16 million colors in one time. Before pre-
processing stage, the image has to be prepared in binary or monochrome representation
for convenience and fast processing. Gray level (monochrome) has 8 bits per pixel and
range from 0 (black) to 255 (white) in between gray while binary image has 1 bit per
pixel which is 0 for black and 1 for white (Figure 3.4).
18
Figure 3.4: Binary image [Marques,MATLAB]
Threshold is actually an initial and a simplest non-contextual technique to
segment the image. It can distinguish the object or foreground from the background. It is
based on intensity of the image where it determines which pixels belong to object and
which pixels should be labeled as background.
The process is comparing each of pixel intensity with a reference value or
threshold and replacing it with a value that means white or black depending on the
outcome of the comparison. Mathematically, the process of thresholding an input image
f(x,y) and producing a binarized version g(x,y) can be expressed as equation (1):
𝑔(𝑥, 𝑦) = { 1 𝑖𝑓 𝑓(𝑥, 𝑦) > 𝑇0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
where T is threshold. When the same value of T is adopted for the entire image, it is
called global thresholding and when the choice of value for T at a point of coordinates
(1)
19
(x,y) depends on statistical properties of pixel values in a neighborhood around (x,y), it
is called local or adaptive thresholding.
In global thresholding, a single value of T is used as a threshold for the entire
image. To choose the value of T, it can be done manually or automatically and there are
many strategies for selecting threshold values. They rely on assumed statistical models
but unfortunately that models neglect some factors such as border and continuity,
shadow and non-uniform reflectance. That’s why for some cases, manual way produce
better result that statistical approach.
Most popular automated approach is by Nobuyuki Otsu [24], using function
graytresh that choose the threshold to minimize the intra class variance (within class) of
the black and white pixels. Otsu’s thresholding perform iterating through all the possible
threshold values and calculating a measure of spread for the pixel levels each side of the
threshold. It wants to find the threshold value where the sum of foreground and back
ground spreads is at minimum. There are equations for calculating between class
variance and within class variance to find the threshold, as follow equation:
Within class variance:
𝜎2𝑊 = 𝜔0𝜎2
0 + 𝜔1𝜎21
Between class variance:
𝜎2𝐵 = 𝜔0(𝜇0 − 𝜇𝑟)2 + 𝜔1(𝜇1 − 𝜇𝑟)2 = 𝜔0𝜔1(𝜇1 − 𝜇0)2
Where 𝜇 = 𝜔0𝜇0 + 𝜔1𝜇1
In this project, the most prominent colour to qualitatively and quantitatively
differentiate between RBC and its background is Green channel. Green channel is used
because it has the best contrast between RBC and background [1]. The threshold value is
ranged from value 0 until 255 and among this values, the most significant threshold that
matches for the whole sample is in range of 145 to 148 (Figure 3.5).
(2)
(3)
20
Figure 3.5: Threshold value setting
A testing has been done to select the correct channel to be used. By using img2,
the actual count for RBC number that manually derived is 82. From Table 3.1, it can be
seen that measurement in channel Red gave the lowest number from the actual count. As
comparison, channel Green and Blue produced the close number to the actual. If Red
channel is selected, the threshold value needs to be higher so that the image can be
differentiated quite clear from the background.
Table 3.1: Comparison of each RGB channel for thresholding purpose
Channel Threshold value Measured count
Red 178 67
Green 148 81
Blue 148 80
21
(a) Green channel
(b) Red channel
(c) Blue channel
Figure 3.6: Comparison of threshold images in different RGB channels
22
Figure 3.6 show the comparison images in each RGB channel. The set threshold
is 148 for all channels. However, when channel Red is set at 148, the result shows the
blank image, only WBCs are appeared. It is meaning the RBC images are not
differentiated from the background. That is the reason channel Red is not selected. This
gives the idea that red channel can be used in order to eliminate WBC but need to be
applied with marker and masking technique.
Thus, channel Green and Blue are suitable to be used. However channel Green
is most prominent for differentiating the foreground and background because the
measurement for all images showed that Green channel tresholding gave better results
compared to Blue channel.
3.1.2.2 Morphological operations
After thresholding process, the next step is to edit or ‘make up’ the image by
using several morphological operations. Morphological functions can improve the
information in a binary image by removing unwanted information such as noise
particles, particles touching the border of images, particle touching each other or particle
with uneven borders. Morphological operations rely only on the relative ordering of
pixels values not the numerical values. Therefore it especially suited to the processing
on binary images and also greyscale images. That is why the conversion of the RGB
image into binary is essential in the image processing.
In this case, the function of NI Vision Assistant is involved. In MATLAB, the
algorithm of morphological operation is quite similar. Since the focus is on the RBC as
the objects, the incomplete shaped RBC or such as attached to the border line will be
ignored. “Remove border” means eliminating the particles that touch to the image
border.
23
The process followed by “fill holes” which is to fill the holes in the RBC images.
The morphological operation for that is closing, which is similar in some ways to
dilation is as expressed in equation (4):
𝐴 ∙ 𝐵 = (𝐴 ⊕ 𝐵) ⊖ 𝐵
It tends to enlarge the boundaries of foreground regions in an image and shrink
background colour holes in such regions but it is less destructive of the original
boundary shape. Holes are filled with the pixel value of 1. The holes are appeared due to
the hemoglobin in the RBC. Thresholding process will leave some of the inner RBC area
become hollow. It needs to be filled up because in the classification stage, area in the
RBC image will be considered.
Besides, the small objects also can be noise that has to be cleared. They come
from platelet area or the imperfection of background identification after thresholding
process. “Remove small objects” is to diminish such noise by erosion many times
(iterations). The morphological operation for that process is opening, which the basic
effect is somewhat like erosion as expressed in equation (5):
𝐴 ∘ 𝐵 = (𝐴 ⊖ 𝐵)⨁𝐵
Erosion is to shrink or thin the binary image. In other words is reducing or minus the
image pixel. This can be applied to eliminate unwanted objects.
“Equalize” is increasing the intensity dynamic by distributing a given grayscale
interval over the full grayscale. It redistributes pixel intensities on order to provide a
linear cumulated histogram.
Lastly, the image has to be in state where objects are in black and background is
in white for be fed into the next stage. Thus, “Reverse” is to reverse the pixel value and
produce the photometric negative of the original image. The outcome of the pre-
processing images will be shown in the Chapter 4.
(4)
(5)
24
3.1.3 Segmentation and classification
Image segmentation and classification is the step for partitioning the images into
its main components. There are two basic properties for most segmentation algorithm
that can be extracted from pixel values, which are discontinuity and similarity or
combination of both. Many problems contribute challenges in segmentation such as non-
uniform, lighting, shadows, overlapping among objects, and poor contrast between
objects and background. The classification or detection stage is involving thresholding
(as in segmentation), particle analysis and binary morphology to locate area features that
match specific criteria that has been decided. Area features must have uniform pixel
intensities or colours that vary significantly from the background pixels.
Particle analysis consists of a series of processing operations and analysis
functions that produce some information about the particle in an image. A particle is a
contiguous region of non-zero pixels. Previously, the image has been extracted by
thresholding process into background and foreground states. Zero valued pixels are in
the background state and non-zero valued pixels are in the foreground. Particle analysis
is performed to detect connected regions or grouping of pixels in an image and then
make selected measurements of those regions. Hence, with this information, the features
can be detected in various part shape or orientation.
In a blood image, there is condition where cells might be overlapped or clumped
more than two cells. To overcome that matter, this project classifies the RBC features
into single (non-overlap) and multi-overlapped cells, which overlapped condition is
could be into double cells, triple cells and more than three cells. The classification and
detection of objects with those features will be done based on the area of the particle.
Particle is defined here as RBC image and area is the number of pixels within the image.
Hence, the area in the RBC is determined in the certain range for single and others.
As shown in Table 3.2, RBC is ranged based on the particle area in pixels. The
pixels value below than 5000 is neglected since it might be non-RBC component such as
platelet or any noise that missed to be extracted out in the early process. Maximum
pixels value is 300000 and upper than that might be the overlapping cells more than four
or could be any larger component other than RBC.
57
REFERENCES
[1] Tomari, R., Wan Zakaria, W. N., Abdul Jamil, M. M., Mohd Nor, F. and Nik
Fuad, N. F., Computer Aided System for Red Blood Cell Classification in Blood
Smear Image, Procedia Computer Science, vol 42,pp. 206-213, 2014
[2] Ajay P. Dhawale and S. R. Hirekhan, Real- Time Image Processing for
Biological Applications Through Morphological Operations using LabVIEW ,
International Journal of Engineering research & Technology (IJERT), vol. 3
issue 5, May 2014.
[3] Venkatalakshmi, B. and Thilagavathi, K. Automatic Red Blood Cell Counting
using Hough Transform, IEEE Conference on Information & Communication
Technologies (ICT), 2013.
[4] Adagale, S.S. and Pawar, S.S. Image Segmentation using PCNN and Template
Matching for Blood Cell Counting, IEEE International Conference on
Computational Intelligence and Computing Research, pp. 1-5, 2013.
[5] Mahmood, N.H. Poon C.L., Mazalan, S.M. and Abdul Razak, M. A. Blood
Cells Extraction using Color Based Segmentation Technique, International
Journal of Life Science Biotechnology and Pharma Research, Vol.2 No. 2, pp.
233-240, April 2013.
58
[6] Tusani, H., Saxena, S. and Yadav, N. Segmentation using Morphological
Watershed Transformation for Counting Blood Cells. International Journal of
Computer & Information Technology. Vol (2). Issue III. Apr-May 2013.
ISSN:2278-7720. Pp.28-36.
[7] Sharif, J. M., Miswan, M. F., Ngadi, M. A., Salam, M. S. and Abdul Jamil, M.
M., Red Blood Cell Segmentation Using Masking and Watershed Algorithm: A
Preliminary Study, International Conference of Biomedical Engineering (IcoBE),
Penang Malaysia, 27-28 February 2012.
[8] Cai, R., Wu, Q., Zhang, R., Fan, L. and Ruan, C. Red Blood Cell Segmentation
using Active Appearance Model, IEEE 11th
International Conference on Signal
Processing, pp. 1641-1644, Vol. 3, 2012.
[9] Berge, H., Taylor, D., Krishnan, S. and Douglas, T.S. Improved Red Blood
Cells Counting in Thin Blood Smear, IEEE- ISBI, pp.204-207, 2011.
[10] Jambhekar, N. J. Red Blood Cells Classification using Image Processing,
Science Research Reporter, 1(3): p151-154, November 2011.
[11] Wang, R., MacCane, B. and Fang, B. RBC Image Segmentation based on
Shape Reconstruction and Multi-scale Surface Fitting, 3rd
International
Symposium of Information Science and Engineering, pp. 586-589, 2010.
[12] Nguyen, N-T., Duong A-D. and Vu, H-Q. A New Method for Splitting Clumped
Cells in Red Blood Images, 2nd
International Conference on Knowledge and
System Engineering, Ho Chi Minh Vietnam, 2010.
[13] Huang J., An Improved Algorithm of Overlapping Cell Division, Intelligent
Computing and Integrated System, International Conference of Digital Object
Identifier, pp. 687-691, 2010.
59
[14] Fox, S.I. Human Physiology, Mc Graw Hill International Ed., 11th
Ed., New
York, 2009.
[15] Vromen, J. and McCane, B. Red Blood Cell Segmentation from SEM Images,
24th
International Conference Image and Vision Computing, New Zealand, 2009.
[16] Wei, X., S.H. Ong, C. Kang, Joo H. L., Jiang, L., D. Racoceanu and K. Foong,
Cell Clumping Qualification and Automatic Area Classification in Peripheral
Blood Smear Image, IEEE , 2009.
[17] Natsution A.M.T, and Suryaningtyas, E.K. Comparison of Red Blood Cells
Counting using two Algorithms: Connected Component Labelling and
Backprojection of Artificial Neural Network, IEEE, Singapore, pp. 1-4, 2008.
[18] Poomcokrak, J. and Neatpisarnvanit, C. Red Blood Cells Extraction and
Counting, The 3rd
International Symposium on Biomedical Engineering, 2008.
[19] Wang, R. and MacCane, B. Red Blood Cell Classification Through Depth Map
and Surface Feature, International Symposium on Computer Science and
Computational Technology, pp. 339-342, 2008.
[20] D.R. Caprette, “Introduction Laboratory : using a counting chamber,” Bios211,
2007.
[21] D. Goyal., Thesis on Diagnosis Disease System using LabView, Thapar Institute
of Engineering and Technology (Deemed University) Patiala, June 2006.
[22] Waters, J. C. Chapter 5- An Introduction to Electronic Image Acquisition during
Light Microscopic Observation of Biological Specimens. Cell Biology. 2006.
Elsevier Science (USA). pp. 49-56.
[23] H. Rhody, Lecture 10:Hough Circle Transform,Chester F. Carlson Center for
Imaging Science, Rochester Institute of Technology, 2005.
60
[24] Webster, J.G. and Cazzanti, S.C. Bioinstrumentation-Hematology, John Wiley
& Sons, Inc., 2004, ISBN978-0-471-26327-2, pp.170-188.
[25] Otsu, N., A Threshold Selection Method from Gray-Level Histogram. IEEE
Transaction on System, Man and Cybernetics, Vol. 9, No. 1, 1979.pp. 62-66.
[26] RBC image. Retrieved from http://vikaspedia.in/health/diseases/genetic-
disorders/sickle-cell-disease on Jun 2015.