Sie sind auf Seite 1von 39

FINAL REPORT OF

Automatic Segmentation of Red Blood Cells from


Microscopic Blood Smear Images using Image
Processing Techniques
A Graduate Project Report submitted to Manipal Academy Of Higher Education

in partial fulfilment of the requirement for the award of the degree of

BACHELOR OF TECHNOLOGY

in

E LECTRONICS AND C OMMUNICATION


E NGINEERING
Submitted by

SUBHRANEIL DAS ABHIROOP DATTA


R EG N O . : 150907434 R EG N O . : 150907120

under the guidance of

K. T. NAVYA
Assistant Professor (Senior Scale)
M ANIPAL I NSTITUTE O F T ECHNOLOGY
ABSTRACT

Human blood is characteristic of various operations that occur within the human body and hence
is a very effective parameter to detect, diagnose and hence rectify ailments of the human body.
Complete blood count is a method to clinically obtain a statistical measure of blood and its related
parameters i.e. red blood cells (RBCs), white blood cells (WBCs), platelets, haemoglobin concen-
tration to name a few. This helps to determine the physical state of the subject. A blood film or
peripheral blood smear is a thin layer of blood smeared on a microscope slide and stained using
various staining methods. A blood smear generally consists of three main cells, red blood cells,
white blood cells and platelets. The main scope of this project is to segment and count red blood
cells from other types of cells to help diagnose RBC related disorders. This project aims to create
an automatic method of segmenting and counting of RBCs from such smear images using image
processing techniques based on MATLAB. However existing methods mostly consist of manually
counting the RBCs from the smear images. Although there are certain software based approaches
to tackle the problem but most of them are not robust for all staining methods. Thus the need of
the hour is to create an automated algorithm which will work for different staining types thereby
alleviating both the aforementioned drawbacks. Robustness is needed as the proposed method
should work for all staining methods i.e. it has to have a certain process of standardization of the
smear images to normalise the colour, contrast and tone of the image. In the proposed method,
the images are first pre-processed, i.e. the images were standardised to a uniform colour profile
which made the components of complete blood count more prominent using contrast enhance-
ment, adaptive histogram equalisation followed by Reinhard stain normalization [9]. Then the
noise from the images were removed by subjecting the processed images through a Wiener filter.
After pre-processing, the image was subjected to green channel extraction, HSI colourspace con-
version and Otsu’s thresholding[18] to differentiate foreground and background. Thereafter using
subsequent steps for segmentation, i.e. active contour algorithm, a count of total RBCs present in
the concerned microscopic blood smear image were obtained. This count was then matched with
a ground truth to obtain the accuracy of the proposed method. All algorithms and processes are
implemented on MATLAB using the Image Processing Toolbox.

i
Contents

1 INTRODUCTION 1

1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

1.2 Organization of the Report . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.3 Project Schedule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2 LITERATURE REVIEW 6

3 METHODOLOGY 9

3.1 Pre-processing of Microscopic Images . . . . . . . . . . . . . . . . . . . . . . . 10

3.2 Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

3.2.1 Removal of WBCs and platelets . . . . . . . . . . . . . . . . . . . . . . 12

3.2.2 Segmentation Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . 13

4 RESULTS 16

4.1 Formation of Ground Truth . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

4.2 Comparison of Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

4.3 Comparison of Colour Correction Methods . . . . . . . . . . . . . . . . . . . . 25

4.4 Comparison of Contour and Morphology Based Approach . . . . . . . . . . . 26

4.5 Comparison of Segmentation Algorithms . . . . . . . . . . . . . . . . . . . . . 26

4.6 Accuracy of Segmentation Algorithms . . . . . . . . . . . . . . . . . . . . . . . 27

4.7 Drawbacks of the Proposed Method . . . . . . . . . . . . . . . . . . . . . . . . 28

ii
5 CONCLUSION AND FUTURE SCOPE OF WORK 30

5.1 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

5.2 Future Plan of Action . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

iii
List of Figures

1 (a) General Blood Smear Image [26] (b) RBC with Central Pallor [28] . . . . . . . 1

2 Microscopic Blood Smear Images of varying lens objective : (a) Wright-Geimsa


Stain (40x image), Image courtesy: Isfahan MISP Online Database [29] (b) Leish-
man Stain (100x image), Image courtesy: KMC Haematology Dept. . . . . . . . . 2

3 Methodology Block Diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

4 Noise Removal Block Diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

5 Methodology Flowchart . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

6 Contrast Enhancement Results : (a) Original 100x image (b) Enhanced image. . . 16

7 GrayWorld Colour Correction Results : (a) Original Image (b) Image after Gray-
world Colour Correction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

8 Gamma Correction Results : (a) Original Image (b) Image after Gamma Correction. 17

9 Reinhard Stain Normalization Results : (a) Original Image (100x lens objective)
(b) Target Image (100x lens objective) (c) Resultant Image. . . . . . . . . . . . . . 17

10 Reinhard Stain normalization Results : (a) Original Image (100x lens objective)
(b) Target Image used (40x lens objective) (c) Resultant Image. . . . . . . . . . . . 18

11 Reinhard Stain Normalization Results : (a) Original image with poor illumination
(40x lens objective) (b) Target Image used (40x lens objective) (c) Resultant Image. 18

12 Noise Removal : (a) Reinhard Normalised Image (b) Sharpened Image using Un-
sharp Masking (c) Image after Noise Removal. . . . . . . . . . . . . . . . . . . . 18

13 WBCs and platelets removal using Active Contour Method : (a) Original Image
(b) S-channel of original image (c) Original binarized image (d) Applied mask for
contour detection (e) Detected WBC (f) Resultant subtracted image. . . . . . . . . 19

iv
14 WBCs and platelets removal using Morphology : (a) S-channel of original image
(b) Green Channel Extraction (c) Extracted WBC Image (d) Subtracted image (e)
Image smoothened by morphological dilation (f) Resultant image after morpho-
logical erosion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

15 WBCs and platelets removal using Morphology;image contains 2 WBCs : (a) S-


channel of original image (b) Green Channel Extraction (c) Extracted WBC Image
(d) Subtracted image (e) Image smoothened by morphological dilation (f) Resul-
tant image after morphological erosion. . . . . . . . . . . . . . . . . . . . . . . . 20

16 RBC Segmentation using Convex Hull Technique : (a) Resultant cleaned image
using erosion (b)Union Convex Hull representation (c) Convex Hull of Image Ob-
jects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

17 RBC Segmentation using Watershed Algorithm : (a)Watershed Transform (b) Su-


perimposition of Watershed Ridgelines (c) Resultant Segmented Image . . . . . . 21

18 RBC Segmentation using Active Contour Algorithm : (a) Initial Contour Mask (b)
Subsequent Contour Mask (c) Resultant Segmented Image . . . . . . . . . . . . . 22

19 RBC Segmentation using Active Contour Algorithm : (a) Initial Contour Mask (b)
Subsequent Contour Mask (c) Resultant Segmented Image . . . . . . . . . . . . . 22

20 Detected RBCs along with their shape enclosed within binary mask . . . . . . . . 23

21 Ground Truth Images : (a) Image 36 (b) Image 40 . . . . . . . . . . . . . . . . . . 24

List of Tables

1 Project Schedule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2 Segmentation Accuracy Chart . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

v
1 INTRODUCTION

Blood sampling and diagnosis of diseases based on the blood count and characteristics is an im-
portant aspect of medicinal science and engineering. Hence to begin with it is necessary to know
some preliminary information regarding human blood and its components along with a generalised
outlook of blood smear, its formation and the microscopic image acquisition of the same.
Over the years since Leeuwenhoek invented the microscope, various techniques have been de-
veloped to obtain blood count in a given subject. Initially manual techniques followed by semi-
automated and automated processes for obtaining the complete blood count has been established.
The blood smear thus taken is analysed and hence subsequent conclusion helps in knowing the
current state of health of the patient as well as diagnose a specific ailment. Human blood is an es-

Figure 1: (a) General Blood Smear Image [26] (b) RBC with Central Pallor [28]

sential source for developing diagnosis patterns and antigens for numerous diseases. Thus knowing
about human blood and its components is a prerequisite.
Blood is a specialised body fluid. It has four main components [12]:

• Plasma, produced in our bone marrow is the liquid component of blood is a mixture of
water, sugar, fat, protein, and salts. The main job of the plasma is to transport blood cells
throughout your body along with nutrients, waste products, antibodies, clotting proteins,
chemical messengers such as hormones, and proteins that help maintain the body’s fluid
balance.

1
• Erythrocytes or red blood cells as shown in Figure 1.(b) are microscopic cells formed in the
bone marrow abundant in haemoglobin (which helps carry oxygen from the lungs to the rest
of the body and then returns carbon dioxide from the body to the lungs so it can be exhaled)
and lack nuclei. The shape of a red blood cell is a biconcave disk with a flattened centre.
Unlike many other cells, red blood cells can easily change shape, helping them fit through
the various blood vessels in your body.

• Leukocytes or white blood cells which exist in the spleen, liver, and lymph glands. They
produce antibodies which are proteins that specifically target bacteria, viruses, and other
foreign materials.. The most common type of white blood cell is the neutrophil, which is
the ”immediate response” cell. The other major type of white blood cell is a lymphocyte.
Lymphocytes help regulate the function of other immune cells and directly attack various
infected cells and tumours.

• Thrombocytes or platelets are small cell fragments without nuclei that work with blood clot-
ting chemicals at the site of wounds and thus help in blood coagulation.

Figure 2: Microscopic Blood Smear Images of varying lens objective : (a) Wright-Geimsa Stain
(40x image), Image courtesy: Isfahan MISP Online Database [29] (b) Leishman Stain (100x im-
age), Image courtesy: KMC Haematology Dept.

The various functions of blood include transporting oxygen and nutrients to the lungs and tissues,
forming blood clots to prevent excess blood loss, carrying cells and antibodies that fight infection,
bringing waste products to the kidneys and liver, which filter and clean the blood, regulating body
temperature.
Figure 1.(a) shows a typical blood film or peripheral blood smear It can be described as a thin layer
of blood smeared on a microscope slide and stained using Romanowsky staining method to exam-
ine the blood cells microscopically. Staining plays a very important role in analysis of blood smear
images as the colour profile, contrast and feature details in an microscopic blood smear image is

2
determined initially by this aspect. Popular staining methods under Romanowsky staining include
Wright-Geimsa stain, Leishman stain shown in figure 2.
A blood smear generally consists of the three main cells, red blood cells, white blood cells and
platelets. The main scope of this project is analysing and segmenting red blood cells from other
types of cells i.e. white blood cells, platelets in order to help diagnose RBC related disorders.
The essence of the project lies in the fact that previous segmentation algorithms often poses a
problem when multiple overlapping red blood cells are present in the blood smear image. Hence a
more analytic approach is taken for this aspect and is aimed at solving the accuracy of the estimate
that is obtained irrespective of the contents of the blood smear image. Accuracy in this scenario
refers to a performance parameter which will provide a measure of efficiency of the constructed
algorithm with respect to a predetermined ground truth. Considering the manual techniques fol-
lowed previously for analysing blood smear images, it was very tedious and not reliable. Thus
This project is aimed at developing an autonomous and robust system for image pre-processing
followed by segmentation of red blood cells in a given blood smear image thereby establishing an
estimate of the same using image processing techniques in MATLAB.

1.1 Motivation

Now a clear understanding of the reasons behind the framing of this project is needed in order to
understand the finer elements of the technical procedures followed. Also grasping the concept of
blood count and its existing methodologies is important to further progress in our quest to obtain a
robust and efficient algorithm for the same.
As mentioned earlier the existing and past methods used for obtaining total blood count have
been lacking in accuracy, speed and efficiency. Generally a complete blood count (CBC) [11] is
performed in clinics and hospitals to obtain an overall report of the statistical perspective of blood
and related components in the subject’s body. However the usage of automate blood analysers
has been reported. Typically, analysis begins when a well mixed whole blood sample is placed
on a rack in the analyser. The instrument utilises flow cells, photometers and apertures in order
to analyse different elements in the blood. The cell counting component counts the numbers and
types of different cells within the blood Although complete blood count is typically not a definitive
diagnostic test. Depending on the reason this test is taken, results outside the normal range may
or may not require follow-up. The results of a CBC might need to be compared alongside results
of other blood tests or additional tests may be necessary. However to make the process digitised,
optimal and mobile at the same time the aspect of software development on the same is considered,
i.e. image processing techniques specifically designed for microscopic blood smear images.

3
Current image processing techniques for blood smear images, generally have not been reported as
robust, i.e. most of the pre-processing and segmentation techniques exist for a specific set of smear
images with a particular colour profile or similar contrasts. Also depending on the staining used
in the smear, overlapping cells in the image and the lab setup utilised, it is necessary to develop a
comparatively robust system. So the need for a robust and automated technique for a wider variety
of smear images arises.
The primary objectives are as follows :

• To develop a robust technique for pre-processing of a peripheral blood smear image.

• Identifying and segmenting the red blood cells from white blood cells and platelets in the
smear image using image processing techniques.

1.2 Organization of the Report

This report aims at providing the reader a clear understanding of the entire project in a sequential
manner. The organization of the report is as follows.
Chapter 1 gives us a brief introduction into the world of haematology and its connected compo-
nents. It gives an insight into the current methods used for human blood analysis and the need for
a more robust and effective algorithm for the same.
Chapter 2 provides us a brief literature review of past work done in various aspects of the project.
The tabular format clearly depicts the work done versus the drawbacks involved in that paper, thus
giving the reader a fair idea of the methodology applied in this project.
Chapter 3 delves deeper into the fine technical aspects of the project involving the various stages
in pre-processing and segmentation. The stages include colour and contrast correction, colour nor-
malization, background and foreground differentiation, thresholding and binarization and active
contour algorithm.
Chapter 4 gives us a brief overview of the definitive results obtained from the application of the
aforementioned processes on the concerned microscopic smear images. Also it presents an opti-
mal comparison of the all the algorithms applied and justifies the one which fits best as per the
requirement. It also explains the process of obtaining the ground truth. It also gives the definitive
explanation and quantitative comparison of all the segmentation algorithms in terms of accuracy
with respect to the sample space used for each algorithm. In addition to that it cites the drawbacks
of the current method.
Chapter 5 provides the reader with a bottomline of the progress achieved and the final conclusion
of the proposed method and the ways in which the results can be further processed and improved.

4
1.3 Project Schedule

A brief schedule of the project has been shown in Table 1.

Table 1: Project Schedule

January - Mid-February 2019 Acquisition and pre-processing of mi-


croscopic smear images :

• Effective Noise Removal


(Wiener filtering/Median filter-
ing techniques

• Contrast enhancement and illu-


mination correction

Pre-processing (continued) and re-


Mid-February - March 2019
moval of WBCs & platelets
Segmentation and Counting of RBCs
April 2019
in the microscopic smear image

May 2019
• Conclude aforementioned seg-
mentation and counting of RBCs

• Related documentation

5
2 LITERATURE REVIEW

This section attempts at giving a brief overview of various past developments in the field of au-
tomated blood counting and its related parameters such as colour normalization, pre-processing,
RBC segmentation, background and foreground differentiation using various segmentation meth-
ods based on underlying principles of image processing.

• J. M. Sharif et. al.[1] elaborates a robust methodology for RBC segmentation using mask-
ing and morphological operation functions to eliminate unwanted objects. The paper has
the combination of pixel based, region based and morphological segmentation. Marker con-
trolled watershed algorithm can avoid over-segmentation issue and it has being used to han-
dle overlapping cell.It is hoped that a better mix Methods can be developed from a variety of
other methods

• Siti Madihah Mazalan et. al.[2] proposes a method to count the RBCs in microscopic images
using CHT (Circular Hough Transform) automatically. Method is cost effective and provides
an alternative way to recognising, analysing and counting the circular cells; identify and
count separately the overlapping RBC. Results are unable to get 100% accuracy due to the
out of range nature of the radius during cropping process.

• Kareem et. al.[3] provides a robust method for pre-processing of blood smear images irre-
spective of the colour profile and contrast of the image. The time and complexity required
for setting up different acquisition and enhancement facilities for the pre-processing of the
images is reduced.The method does not require any binarisation as the method works directly
on the grayscale image. The method is limited to the MATLAB environment and hence lacks
platform mobility.

• M.Habibzadeh et. al.[4] gives an insight into a generalised and robust algorithm for pre-
processing and segmentation.The main breakthroughs are: (a) robust noise removal and mit-
igating its side effect problem (blurring) ; (b) Separation of blood smear particles into two
sub-images using a mask-based algorithm. Experimental results indicate that the current
analysis of blood cells is accurate, easy to implement with a remarkable accuracy, which is
asserted by comparing the automatically extracted blood particles with traditional segmen-
tation techniques.

• Razali Tomaria et. al.[5] describes a framework for automatically classify the RBC into
overlap, normal and abnormal cluster is proposed. The system consist of combination of

6
three main blocks which are segmentation and processing block, feature extraction block
and classification block. Otsu segmentation method applied on green colour channel image
with a series of post processing is found to effectively prune out solid RBC shape from the
background. ANN framework was tested to classify the RBC in blood cell images, and the
results are encouraging with an average of 83% accuracy, 82% average precision and 76%
average recall.

• Sherif Abbas et. al.[6] presents a method for Red Blood Corpuscles (RBCs) counting has
been developed using RBCs light microscopic images and MATLAB . A RBC mask was
used to count the number of RBCs in the blood smear image. The discusses method is
useful for recognising, analysing, identifying and counting separately the overlapping RBCs.
However it lacks platform compatibility and mobility owing to development in MATLAB.

• Derek Magee et. al.[10] holds up two novel colour normalization approaches against the
previously utilised method of linear normalization in lαβ colourspace. It showed reasonable
consistency in the stained areas, but poor consistency in the background.

• Erik Reinhard et. al.[9] focuses on the fact that a colourspace with decorrelated axes is a
useful tool for manipulating colour images. Imposing mean and standard deviation onto
the data points is a simple operation. Usage of statistical parameters (mean and SD) ensure
mathematical consistency.

• Miguel et. al.[15] exhibits evaluation of a method to segment and separate WBCs and subse-
quently segmenting RBCs using watershed algorithm. Reasonable consistency in uniformly
stained areas, but fails to segment WBCs/RBCs properly when staining and illumination are
not uniform.

• D. Baswaraj et. al.[13] demonstrates the different methods using active contours employed
to obtain required features from an image. Active contour methods pull the contours to the
exact boundaries and are capable in their powerful abilities in object segmentation. However,
local minimum problems still exist within these models.

• Kass et. al. [16] demonstrates development of practical methods to segment and hence
separate image objects using active contour or otherwise known as snakes. It Provides the
background theory and practicality behind the working of the active contours. Thus this is
the rudimentary work done on contour algorithms and hence is very important to understand
the working of the algorithm as stated in the proposed methodology.

• Jinping et. al. [15] evaluates a method to segment and separate overlapping RBCs in a blood
smear image using morphological operations and convexity closure. The method presents

7
reliable consistency in majority of series, parallel, series-parallel cases of cell overlapping.
However if the number of overlapping cells is more than 6 and one cell nuclei is overlapping
with another cell nuclei, the method proposed lacks result accuracy.

Generally a series of salient observations were noted from the previous section. As far as the
pre-processing phase is concerned, Reinhard’s method of colour normalization suits the problem
statement of standardization of images in the best way possible provided a proper standard image
is available to be used as the target image. In this situation controlled gamma correction is advis-
able. The other methods including grayworld colour correction and linear contrast enhancement
performs inconsistently and hence are not preferred. Further drawbacks and a detailed discussion
is given in subsection 5.1 & 5.2. Generally a series of methods have been implemented in these
papers to obtain the total RBC count. The methods range from the basic circular Hough Transform
to the use of active contours or snakes. However for the purpose of segmentation and counting,
active contouring models have proven to be the most efficient. This can be verified as per the
accuracy results discussed in latter section.

8
3 METHODOLOGY

This section aims at giving the reader a clear scenario of the sequence of events that are to be
followed to obtain the desired results. A methodology as depicted in Figure 3 has been devised
which the readers can refer at all times to keep track of the algorithms and steps followed in the
process.
The project can be summarised by the block diagram shown in Figure 3.

Figure 3: Methodology Block Diagram

• Image Acquisition
The required blood smear image is acquired from KMC blood database with each smear
image being obtained on a microscopic setup with a 100x lens objective. Apart from this
smear images from Isfahan MISP online database[29] have been acquired.

• Pre-processing
The very first step in processing the given blood smear image is to make sure that the image is
ready to be segmented. For this reason various stages involving noise removal, illumination
correction and colour normalization is implemented at this stage.

9
• Segmentation
As the name suggests this is the primary objective of this project wherein the pre-processed
image will be further subjected to contour detection and background subtraction techniques
to segregate and identify the red blood cells present in the blood smear image following
which an automatic system is to be developed for the entire process to be accommodated in
a single bundle thus simplifying the issues of manual labour and accuracy.

3.1 Pre-processing of Microscopic Images

The various steps in pre-processing can be further elaborated as :

• Linear Contrast Enhancement using adaptive histogram equalisation


The acquired images gave faded features and hence it is very difficult to proceed with any
operation as such. Hence the images are first subjected to conversion to LAB colourspace
where a consistent normalization of the L parameter (Luminosity) is performed followed by
grayscale conversion and subsequent adaptive histogram equalisation. It is important to note
that the maximum value of the L parameter is taken as 100 units.

• Gray world colour correction method


Gray world assumption [7] is a colour correction function based on the principal that on av-
erage the world is grey. The average pixel value of an Unsigned 8-bit integer image is 127.5,
so by calculating the real average pixel value, the scaling value is computed. This scaling
value is used to scale the entire image linearly so that the average of the image is 127.5 In
practice the average of each individual channel is used to calculate a separate scaling value
for each channel. This way the illumination on the different channel is eliminated. Here a
higher scaling value has been used.

• Gamma correction algorithm


Gamma correction [8] is a non-linear method to scale the colour tone of an image, i.e. am-
plify the shadows or the bright regions of the image as per the requirement. It is used to
encode and encode luminance or tristimulus values in video or still image systems. Gamma
correction is defined as :
Vout = AVinγ

10
Powers larger than 1 make the shadows darker, while powers smaller than 1 make dark re-
gions lighter. A gamma value < 1 is called an encoding gamma, and the process is called
gamma compression (compressive power-law nonlinearity application); conversely a gamma
value > 1 is called a decoding gamma and is called gamma expansion (expansive power-law
nonlinearity application).

• Reinhard’s stain normalization algorithm


As mentioned earlier due to the varied colour profile and contrast of the microscopic smear
images, there was a need to standardise the images to obtain a consistent colour profile. This
algorithm thus maps the colour distribution of an image to that of a well stained target image
thereby solving the problem of inconsistency [10]. Technically it matches the mean and
standard deviation of each colour channel in the two images in that colourspace [9]. The
statistical approach to colour mapping can be shown as :

loriginal − ˜loriginal ˆ
lmapped = ltarget + ˜ltarget
ˆloriginal

αoriginal − α̃original
αmapped = α̂target + α̃target
α̂original
βoriginal − β̃original
βmapped = β̂target + β̃target
β̂original
where the superscripted variables ˜l and ˆl represent the mean and standard deviation of lu-
minosity parameter respectively in the lαβ colourspace. Similarly the subscripted variables
loriginal , ltarget and lmapped indicate whether the parameter concerned is either of the original,
target or the mapped image of the lαβ colourspace parameters. An image with acceptable
colour profile, contrast and illumination was chosen as the target image. The source image
was subjected to Reinhard staining normalization.

• Noise Removal
The blood smear image post colour normalization is subjected to noise removal and deblur-
ring using Wiener filter. The filter processes a two dimensional image i.e. grayscale image in
this case and eliminates granular (salt and pepper ; Gaussian) noise from the image. It is also
useful in removing motion blur from the image thereby making the image ready for further
processing. The generalized block diagram clearly shows how the original signal (micro-
scopic image region pixels in this case) in which noise is present by default is subjected
to Wiener filtering to receive a smoothened and granulation free image ready for further
processing.

11
Figure 4: Noise Removal Block Diagram

3.2 Segmentation

Prior to segmentation, the images after being standardised need to be subjected to various processes
to make it suitable for segmentation and counting. The processes include:

• Removal of WBCs and platelets

• Highlighting remaining red blood cells in the microscopic image

This section further elaborates the previously stated processes and emphasises on the underlying
principles of image processing.

3.2.1 Removal of WBCs and platelets

The initial concept required for removal of all non necessary components of the microscopic im-
age involves two main principles of image processing, namely morphology and usage of HSI
colourspace. As the name suggests, HSI stands for hue, saturation and intensity. The reason for
the usage of HSI colour space is that under Rowmanosky staining, WBCs become clearly visible
in the saturation model (S-model) compared to other components. The S model is then subjected
to thresholding and binarization, such that the concerned WBCs is clearly the white foreground
and the remaining portion of image is the black background.
Simultaneously the image obtained after Reinhard staining method is then subjected to green chan-
nel extraction, thresholding and subsequent binarization. The second image obtained from green
channel extraction and binarization of the original image consists of RBCs, WBCs, and platelets
as the white foreground with the remaining image as black background. The image with the ex-
tracted WBCs is subjected to morphological dilation to enlarge the detected WBCs to include their

12
outer cytoplasm. This dilated WBCs image is then subtracted from the green channel binarized
image thus resulting in a image devoid of WBCs and platelets. This image is complemented and
morphologically eroded for smoothening and clearing feature boundaries of the RBCs.

3.2.2 Segmentation Algorithms

A considerable amount of thought was given to the algorithms which were to be used to segmenting
the RBCs. Factors such as the shape of the extracted cell, resolution of overlapping cells, over-
segmentation were carefully taken under consideration while deciding the algorithm. The four
basic algorithms that were taken under development are as follows.

• Convex Hull Technique


Convex hull can be defined as the largest polygon that can contain or rather enclose all the
objects present in an image which in this case are the red blood cells. As stated in [17] a
convex hull is constructed as per the image objects and clusters it contains and then masking
with the original binary image results in detection of singular blobs in the image. However
some blobs which are formed due to overlapping cells are detected as singular and thus
present the problem of under-counting of cells. This deviation is very large in this method
owing to the lack of specificity in the method, i.e. its inability to separate clumps of objects.
Thus using convex hull compromises the counting accuracy as it clumps overlapping cells
into a singular large polygon. After applying convex hull ’regionprops’ is used to detect the
centroids of the detected objects. Using the above data, the detected cells are encircled for
the purpose of clarity and counting.

• Watershed Algorithm
The images are now subjected to watershed transform. However the image needs to be pre-
processed rigorously before being subject to watershed algorithm to avoid over-segmentation.
Referenced publications[1] and Mathworks documentation[20][21] recommend the use of
marker controlled watershed algorithm. The processing done on the binary image helps
in identifying the catchment areas of the image i.e. the local maximas. On making these
points clear, the algorithm recognizes the regions of maxima and is able to draw watershed
ridge lines to distinguish the overlapped cells. The method can be summarised as follows.
The negative distance transform of the complement of the image is obtained following the
catchment areas are highlighted. Then the image is subjected to watershed algorithm and
the ridgelines thus formed distinguish the overlapping cells as can be seen from the super-
imposed ridgeline mask on the original image in order to get a clear understanding of the

13
scenario. Thus thin connecting lines at the overlap regions now serve as ideal points for es-
tablishing watershed ridgelines. However some of the images obtained clearly exhibit cases
of over-segmentation is some regions of the image thereby posing inconsistencies in the final
results.

• Active Contour Detection Algorithm


Kass et. al.[16] states ”A snake is an energy-minimizing spline guided by external constraint
forces and influenced by image forces that pull it toward features such as lines and edges.
Snakes are active contour models, i.e. they lock onto nearby edges, localizing them accu-
rately”. In much more simpler terms an active contour can be defined as a closed curve
defined in terms of gradient, internal energy and parametric quantities which has a tendency
to fit on image objects based on image boundaries and features thereby providing us with a
clear scenario of object detection. The method is heavily used in blob detection and dynamic
motion tracing. The equations associated with an active contour are generally the energy and
equilibrium equation respectively. The basic premise of active contours is energy minimis-
ing models i.e. the snake tends to minimize its energy thereby shrinking onto the boundaries
of the image objects. The entire process of fitting onto the boundaries occur in several iter-
ations to get the best fit possible. Now, considering the snake as an continuous parametric
variable, we can define its position in the image as

v(t) = (x(t), y(t)),

where v(t) is the active contour and x(t),y(t) are the continuous contour coordinates. Thus
the energy equation[16] associated with active contours are as follows :
Z 1 Z 1

Esnake = Esnake (v(t))dt = Eint (v(t)) + Eimage (v(t)) + Econ (v(t))dt
0 0

where Eint is the internal energy of the defined snake, Eimage stands for forces existing in
the image, and Econ gives rise to the external constraint forces to help fit the contour better.
Minimizing this energy i.e. causing it to fit in the image such that all forces so defined are in
equilibrium.

From the above discussion notable points were observed. It is very clear that as per the nature of the
blood smear images used, segmenting cells as per their position in the image is difficult to achieve
using convex hull approach as the convex polygon that encloses objects will be unable to separate
cell clusters and marks it as a singular cell thereby leading to inaccuracy. Watershed algorithm on
the other hand poses the threat of over-segmentation of RBCs as the number of watershed regions
as detected by the algorithm is not accurate for the images being used. Hence the method which

14
gives the closest count with respect to the number of RBCs per image is active contour method.
Thus the proposed method which has been followed to obtain all results pertaining to the scope of
this project can be sequentially represented by the following flowchart.

Figure 5: Methodology Flowchart

This project can be summarised as a continuous process of image acquisition, pre-processing,


image refining (WBC and platelet subtraction) followed by segmentation and counting of RBCs.

15
4 RESULTS

The algorithms on being implemented on a set of images gave a series of results based on their
resolution, contrast standardization and processing method used as stated in this section. Thus it
gives a clearer depiction of the results obtained for each operation involved along with the charac-
teristics and drawbacks related to that process in particular.

Figure 6: Contrast Enhancement Results : (a) Original 100x image (b) Enhanced image.

Figure 7: GrayWorld Colour Correction Results : (a) Original Image (b) Image after Grayworld
Colour Correction.

In Figure 6(a), the original faded microscopic image of a blood smear is shown which is almost
devoid of sharpness, contrast and colour. Figure 6(b) shows the corrected image wherein the linear
colour normalization using Lαβ colourspace followed by adaptive histogram equalisation is per-
formed on the source image.

16
Figure 8: Gamma Correction Results : (a) Original Image (b) Image after Gamma Correction.

Figure 9: Reinhard Stain Normalization Results : (a) Original Image (100x lens objective) (b)
Target Image (100x lens objective) (c) Resultant Image.

The image is also subjected to Grayworld colour correction algorithm as shown in Figure 7 in order
to correct uniformity of colour, contrast and primarily for illumination correction. However results
obtained show consistency in image regions with proper contrast and deviation for other regions
thereby proving to be inefficient for the required purpose. Figures 7(a) & 7(b) show the original
images and colour corrected images respectively as obtained by Grayworld’s algorithm.
Gamma correction primarily used for the purposes of colour correction is implemented on these
images to analyse its performance. Figure 8 clearly depicts that with a controlled gamma exposure
i.e. proper value of γ parameter as mentioned in subsection 3.2.1, a proper colour tone can be
achieved. Figures 8(a) represents the original image and similarly 8(b) represents gamma correc-
tion wherein the value of γ is specified by trial and error method.
The images are then subjected to the statistical means of colour matching in order to achieve a
fixed level of uniform tone and contrast which can then be applied and fixated to any image which
is fed as an input to the algorithm. The mean and standard deviation parameters are relied on

17
Figure 10: Reinhard Stain normalization Results : (a) Original Image (100x lens objective) (b)
Target Image used (40x lens objective) (c) Resultant Image.

Figure 11: Reinhard Stain Normalization Results : (a) Original image with poor illumination (40x
lens objective) (b) Target Image used (40x lens objective) (c) Resultant Image.

Figure 12: Noise Removal : (a) Reinhard Normalised Image (b) Sharpened Image using Unsharp
Masking (c) Image after Noise Removal.

18
Figure 13: WBCs and platelets removal using Active Contour Method : (a) Original Image (b)
S-channel of original image (c) Original binarized image (d) Applied mask for contour detection
(e) Detected WBC (f) Resultant subtracted image.

Figure 14: WBCs and platelets removal using Morphology : (a) S-channel of original image (b)
Green Channel Extraction (c) Extracted WBC Image (d) Subtracted image (e) Image smoothened
by morphological dilation (f) Resultant image after morphological erosion.

19
Figure 15: WBCs and platelets removal using Morphology;image contains 2 WBCs : (a) S-channel
of original image (b) Green Channel Extraction (c) Extracted WBC Image (d) Subtracted image
(e) Image smoothened by morphological dilation (f) Resultant image after morphological erosion.

extensively to provide a stable platform for mapping of these values from the target image onto
the source image. The target image ideally should possess uniformity in regard to the factors of
illumination, colour tone, contrast, warmth and sharpness. Here the source image is subjected to
Reinhard stain normalization with reference to the target image wherein both the images are of
100x magnification. Figures 9 to 11 show the concerned results for this step including images of
both images of 100x and 40x lens objective with varying levels of contrast, tone and illumination.

This colour normalised image is now ready to be subjected to further operations for foreground
and background differentiation. However the image being augmented with granular and salt and
pepper noise by default needs to pass through a noise filtering unit. But prior to that the object
boundaries in the image need to distinct with respect to the background especially at the edges
of the RBCs. Thus the Reinhard normalised image is then subjected to unsharp masking[19] for
sharpening of object boundaries. The name value pair of ’Radius’ = 2 and ’Amount’ = 1 is used.
The resultant image is then passed through a Wiener filter with a filtering window of [3,3] as spec-
ified in the MATLAB script. The resultant images are smoothened and exhibit a much less extent
of jagged edges in the features of the image. Figures 12(a), (b) & (c) clearly depict the image
sharpening and noise removal process.

20
Figure 16: RBC Segmentation using Convex Hull Technique : (a) Resultant cleaned image using
erosion (b)Union Convex Hull representation (c) Convex Hull of Image Objects

Figure 17: RBC Segmentation using Watershed Algorithm : (a)Watershed Transform (b) Super-
imposition of Watershed Ridgelines (c) Resultant Segmented Image

Active contour method was experimented as one of the methods for WBC removal. In this method
the code was designed so as to fit a square shaped mask around the proposed WBC region. Then
the mask boundary (contour) would start converging towards the WBC border and hence the shape
of the WBC is extracted. However sometimes, neighbouring RBCs are also taken into the contour
and the exact WBC is not present in the result because of the vicinity of the RBCs to the WBC.
The concerned results can be seen in Figure 13.

The previous predicament of contour method which was causing neighbouring RBCs to be taken
into the proposed contour region resulting in unwanted subtraction of those required cells was
solved to a certain extent by the use of morphological operators in MATLAB. The WBC after be-
ing extracted using Otsu’s thresholding method is subjected to morphological opening via a disk
shaped morphological element of radius range of 80 to 90, which in turn detect the extracted WBC
along with its cytoplasm and then a similar sequence of events is followed i.e. platelets removal,
subtraction and smoothening to obtain the final image. The appropriate results of the same with

21
Figure 18: RBC Segmentation using Active Contour Algorithm : (a) Initial Contour Mask (b)
Subsequent Contour Mask (c) Resultant Segmented Image

Figure 19: RBC Segmentation using Active Contour Algorithm : (a) Initial Contour Mask (b)
Subsequent Contour Mask (c) Resultant Segmented Image

the occurrence of a single as well as two WBC in the smear images can be seen in Figure 14 & 15
respectively.
Images after being cleared of WBCs and platelets are then subjected to segmentation algorithms.
Although three different approaches were tested only one was finally implemented for processing
the entire database and the accuracy of the same was documented.
Figure 16(b) & 16(c) shows the process of convex hull operation for RBC segmentation on Image
40. The results clearly exhibit the same problem as stated earlier, i.e. clumping of overlapping or
clustered RBCs. This causes the final count to be much lesser than the actual thereby leading to
loss in accuracy. It can now be safely claimed that the major problem with this method is under-
counting of image objects owing to clumping of image objects.
The images were then subjected to the next segmentation method, i.e. the watershed algorithm.
However using watershed requires a fair bit of pre-processing in the resultant binary image after
WBC and platelet subtraction. This pre-processing aims at clearing the image further and limiting
the regions of local maxima present in the image only to within the cell boundaries. Doing so
causes the distance transform to clearly depict the catchment areas, i.e. the local maxima regions.
This causes the juxtaposition of watershed ridgelines to be more accurate. As can be seen in figure
17 (a),(b) & (c), oversegmentation is clearly visible in the final image (superimposed on the orig-

22
Figure 20: Detected RBCs along with their shape enclosed within binary mask

inal binary image). Thus contrary to convex hull technique this method presents the problem of
overcounting.
The priority being proper detection of RBCs in the image, the focus shifted to the active con-
tour algorithm. As explained earlier in section 3.2.3, the method conforms to our requirements of
correctly counting the RBCs present in a single microscopic blood smear image. However as op-
posed to the general notion of active contours where the contour usually has a converging gradient
force, the current project requirement was a diverging or expanding contour. The divergence of
the contour was achieved in MATLAB by exercising a certain degree of ”contraction bias”. This
parameter ranges from -1 to 1 and a negative value indicates expanding contour whereas a positive
value indicates a shrinking contour. Figures 18 & 19 clearly depict the entire operation of active
contouring and centers detection of detected cells. An overlay mask for image 40 was created
from the resultant binary images of contour detected RBCs as can be seen in figure 20, in order to
present the detected cells clearly.

4.1 Formation of Ground Truth

The binary image thus obtained is now devoid of platelets and WBCs. This image can now be
subjected to segmentation algorithms for obtaining the count of RBCs. However as in the case of
any algorithm a reference or ground truth is required for measuring accuracy of the used algorithm.

23
Here we make use of ImageJ Tool to segment the microscopic blood smear images to get a clear
count of the RBCs present in the image. The process of segmenting each and every image manu-
ally is tedious but yields an accurate count to serve as the ground truth for calculating algorithm
accuracy. The process followed is very straightforward and can be summarised as:

Figure 21: Ground Truth Images : (a) Image 36 (b) Image 40

• Loading of image from database

• Contrast enhancement (Pixel saturation = 0.01%)

• Demarcating clustered or ambiguous cell boundaries using freehand line tool to obtain clean
cell boundaries

• Removal of WBCs, platelets and border cells using free hand select tool followed by filling
of those regions

• Splitting of image into three individual channels out of which green channel is used

• Application of threshold is done so as to get a clear binary image of only RBCs

• Analysis of image objects is done to obtain a outlined image of the same along with the RBC
count

The count thus obtained is later used for estimating accuracy of the segmentation algorithm used.
Sample ground truth images have been shown below.

24
4.2 Comparison of Algorithms

This section presents a clear comparison of all the methods as previously mentioned under method-
ology chapter, based on the results obtained. This along the accuracy results as stated in the accu-
racy chart further helped in determining the final method which is best suited for the purpose of
this project.

4.3 Comparison of Colour Correction Methods

A comprehensive implementation of the aforementioned methods gave us the following observa-


tions.

• The Linear method of contrast enhancement though limited to correcting over illuminated
images, provides satisfactory results were obtained for most smear images.

• Grayworld colour normalization method is not uniform and fails in consistency owing to
scaling factor imbalances

• Gamma correction performs decently on most of the images provided that a proper gamma
parameter is chosen to differentiate foreground and background in the image. However the
gamma correction method used gives undesirable results if the gamma parameter is even
slightly varied from the preferred value.

• Reinhard stain normalization on the contrary exhibits consistent performance while correct-
ing images based on staining of a target image. However in this method the resolution of
the target image has to be more than the source image at all instances. Since the database
being used has uniform resolution 100x images, the pixel values in the code are fixed and
need not be manipulated in any manner for the entire database and for a suitable target im-
age. However for different databases the pixel values of the source and target images need to
be manually changed. However given that such technical flexibility is allowed, this method
proves to be the most feasible in terms of performance and consistency.

From the above, it is evident that advisable gamma correction followed by Reinhard’s method and
is suited for the required colour normalization process.

25
4.4 Comparison of Contour and Morphology Based Approach

This project has obtained a binarized image without WBC and platelets so far based on the two
approaches, i.e. active contouring and morphology respectively. As evident of the results the
following observations were noted :

• The active contour method demands the usage of a properly aligned mask with respect to
the location of WBCs in order to extract it completely along with the cell cytoplasm re-
gions. Hence achieving a dynamic masking algorithm is complex and would work finer if a
diverging contour is applied instead of the converging contour as applied currently.

• The Morphology is better in terms of performance ad consistency as compared to the con-


verging contour algorithm. However it too shares the similar issue of clipping of RBCs in
vicinity of the concerned WBC. Although the extent of clipping is much less than the previ-
ous method, a much more sophisticated algorithm is preferable for proper segmentation of
the RBCs. Hence a late update was made in the code for this approach wherein the follow-
ing improvements were made. In addition to the previously used rudimentary morphological
operations of erosion and dilation, operations involving binary image morphing[22] and hole
filling were implemented and specifically adjusted for robust operation. The structuring ele-
ment size was modified in terms of pixels to get perfect subtraction of WBCs from the image.
In a latter update when the diverging contour model was materialised, it still was not used in
place of this owing to the level of robustness and stability which had already been achieved
using this approach.

4.5 Comparison of Segmentation Algorithms

Using a total of three segmentation algorithms namely convex hull, watershed algorithm and active
contour model a series of microscopic blood smear images were processed to get trial results before
the final code was determined to process the entire database consisting of 150 images. From these
trial runs the following observations were noted :

• Convex Hull Technique although widely used for segmentation purposes in the field of char-
acter segmentation and recognition, is unsuitable for scenario at hand. Given the fact seg-
menting cells of any kind requires accuracy in terms of cluster separation and proper centroid
detection, generating convex hull for image objects degrades the preexisting binary image by

26
clumping individual cells if they are very finely separated from a nearby cluster as is evident
in figure 16.

• Watershed algorithm as observed from the trial runs itself, was resulting in over-segmentation
of the binary image owing to presence of ambiguous local maximas in the binary image even
after implementing a series of pre-processing steps which was specific to this algorithm. This
results in over-counting of total RBCs leading to inaccuracy as depicted in figure 17.

• Active Contour Model provides us with the solution to both the problems as mentioned in
the aforementioned algorithms, i.e. the counts obtained by applying the proposed method
are in the vicinity of the actual count as specified by the ground truth. Erosion limit of the
binary image is done to strictly restrict the white elements in the binary image to be only
those pertaining to the central portion of all RBCs present in the binary image.

4.6 Accuracy of Segmentation Algorithms

The segmentation algorithms used has already been compared in terms of their performance in a
qualitative sense earlier in this article. Now using the count obtained from the ground truth of the
images and the count obtained via our proposed algorithm we determine the accuracy of all three
methods to showcase their efficiency and robustness.
Accuracy in the scope of this article for a single microscopic blood smear image can be defined as
the percentage of detected count divided by the actual count of RBCs present in that image. The
accuracy (let it denoted as a(t) for a smear image t) of the algorithm used for this image can be
mathematically represented as :
d(t)
a(t) = ∗ 100%
c(t)
where d(t) represents detected count, c(t) represents actual count.
The following table depicts the accuracy for three segmentation algorithms namely for convex
hull, watershed technique and active contour method which was tested against a set of 5, 5 and 143
images (entire database) respectively.

It is evident from the above table that the count accuracy was the least in the case of
watershed algorithm followed by slightly better accuracy as shown by convex hull technique. It is
important to note that the accuracy field for watershed algorithm does not provide any numerical
data owing to the fact that due to heavy over-segmentation the number of cells detected by the
algorithm is huge as compared to the actual count. As per the formula for accuracy as stated in
section 6, the ratio becomes more than 1 which is not a feasible figure in terms of efficiency and

27
Table 2: Segmentation Accuracy Chart

Technique Used Sample Images Average Accuracy with respect to sample space
Convex Hull 5 63.054%
Watershed Algorithm 5 Heavy Over-segmentation causes negligible accuracy
Active Contour Model 150 73.22%

robustness. However both these methods show low accuracy as compared to the active contour
algorithm. The accuracy for active contour model being the closest to the actual count can be
safely stated as the better choice for a segmentation algorithm within the aforementioned three
algorithms. Using active contour algorithm yields an accuracy greater than 80% for 50 images.
However overall accuracy as shown by active contour is not completely robust because though it
resolves a significant amount of overlapped cells present in an image, some overlapped cells and
clusters still remain unresolved. Further discussion on improvement of the proposed method has
been done in the following section.

4.7 Drawbacks of the Proposed Method

The proposed method of segmentation using using Reinhard stain normalization and active contour
model as the main components of the pre-processing and segmentation stages respectively, yielded
satisfactory results. Although most of the red blood cells get detected by this method, a significant
number of cells still remain undetected thereby decreasing the overall accuracy. The following
major drawbacks of this method are enumerated as follows.

• Although the current method gives us majority of cell centers in a given microscopic image
the fact that cell Segmentation of any kind requires a high level of accuracy in terms of
detection and Segmentation, the traditional approach in MATLAB without the use of any
machine learning techniques for classification lags behind by a huge margin in terms of
efficiency as a CNN (convolutional neural network ) or a shape classifier makes the job of
WBC and platelet removal a lot easier. And post Segmentation of RBCs the shape of the
RBCs can be used as an input to a trained SVM (Support Vector Machine) to classify RBC
types. Unfortunately lacking the knowledge to wield these tools, the exact shape of every
RBC in an image could not be clearly demarcated.

• The degree of robustness achieved with this model tends to limit itself to an average value
of as seen in section 6 (accuracy section). The reason for this gap in achieving a much more

28
satisfactory level of accuracy is due to the fact that the proposed method uses erosion limit to
get the centers of individual RBCs including overlapped cells. However quite a few images
presented the problem in 1separating overlapped and clustered RBCs. Thus the average
count accuracy is significantly affected.

• Several images of the database had to be omitted from being processed as they exhibited
a heavy degree of clustering (for example, more than half of one RBC covered by another
RBC) and owing to the poor contrast and focus of the objective lens the cells could not be
detected even with proper refining of the code. In addition to these semi clustered images
were present in the database which consisted of both clean cells as well as densely clustered
regions and hence could not completely ignored. This overall accuracy decreased due to the
individual accuracy of these images.

29
5 CONCLUSION AND FUTURE SCOPE OF WORK

This section summarises the entire methodolgy and the consequences of the implementation of
the proposed method. The final outcome of the work done with respect to the level of robustness
achieved and the scope for further improvement have been discussed.

5.1 Conclusion

The methods implemented for achieving the primary objectives of this project provides a clear
scenario of the project and the accuracy of results that has been achieved so far. Pre-processing
of the images proved to be a cumbersome task and after several trial runs, Reinhard’s method
proved to be the most effective method of colour normalization. Similarly post WBC and platelet
removal, active contour model provided a RBC count closest to the actual values as obtained from
the ground truth.

5.2 Future Plan of Action

The next step would be to overcome these previously mentioned problems by introducing a robust
algorithm especially for segmenting overlapping and densely clustered RBCs in the resultant image
to get a much more accurate count. The proposed plan of action would be to go through a series of
segmentation processes using convolutional neural networks, and artificial intelligence toolboxes
to create a shape classifier for identifying overlapped cells and cell clusters accurately. Also using
feature extraction in CNN enables the developer to robustly detect the edges of the detected RBCs
without losing any major features. This shape can then be used to create a training set to train a
model to classify types of RBCs, i.e. healthy RBCs, sickle cells, microcytes, macrocytes and burr
cells in order to detect specific diseases and anomalies.

30
References

[1] J. M. Sharif, M. F. Miswan, M. A. Ngadi, Md Sah Hj Salam, ”Red blood cell segmentation
using masking and watershed algorithm: A preliminary study”, 2012 International Conference
on Biomedical Engineering (ICoBE), Penang, Malaysia,27-28 February 2012, Publisher : IEEE.

[2] Siti Madihah Mazalan, Nasrul Humaimi Mahmood, Mohd Azhar Abdul Razak, “Automated
Red Blood Cells Counting in Peripheral Blood Smear Image Using Circular Hough Transform”,
2013 IEEE First International Conference on Artificial Intelligence, Modelling & Simulation,
Publisher : IEEE.

[3] Kareem, S. Morling, R. C. Kale I., “A Novel Method to count the red cells in thin blood films”,
2011 IEEE International Symposium of Circuits and Systems (ISCAS), Publisher : IEEE.

[4] M.Habibzadeh, A. Krzyżak, T.Fevens, A.Sadr, “Counting of RBCs and WBCs in noisy normal
blood smear microscopic images”, Medical Imaging 2011, Proc. of SPIE Vol. 7963, 79633-1.

[5] Razali Tomaria, Wan Nurshazwani, Wan Zakariaa, Muhammad Mahadi Abdul Jamila, Faridah
Mohd Norb, Nik Farhan Nik Fuadb, ”Computer Aided System for Red Blood Cell Classification
in Blood Smear Image”, International Conference on Robot PRIDE 2013-2014 - Medical and
Rehabilitation Robotics and Instrumentation, ConfPRIDE 2013-2014, Publisher : Elsevier

[6] Sherif Abbas, ”Microscopic Images Dataset for Automation of RBCs Ccounting”, Middle East
Technical University 2015, Publisher: Elsevier

[7] Kwok, N. M., Wang D., Jia X., Chen S. Y., Fang G., Ha Q. P., ”Gray World based Color Cor-
rection and Intensity Preservation for Image Enhancement”, 2011 4th International Congress on
Image and Signal Processing, Publisher : IEEE

[8] Wilfried Kubinger, Markus Vincze, Minu Ayromlou, “The Role of Gamma Correction in
Colour Image Processing”, 9th European Signal Processing Conference (EUSIPCO 1998), Pub-
lisher : IEEE

[9] Erik Reinhard, Michael Ashikhmin, Bruce Gooch, and Peter Shirley, University of Utah,
“Colour transfer between Images”, IEEE 2001

[10] Derek Magee, Darren Treanor, Doreen Crellin, Mike Shires, Katherine Smith, Kevin Mohee,
and Philip Quirke, ”Colour Normalisation in Digital Histopathology Images”, Elsevier 2009.

31
[11] Complete Blood Count (CBC)
Source : Wikipedia

[12] Human Blood and Its Components


URL : https://www.hematology.org/Patients/Basics/

[13] D. Baswaraj, Dr. A. Govardhan, Dr. P. Premchand, “Active Contours and Image Segmen-
tation: The Current State of the Art” , Global Journal of Computer Science and Technology
Graphics & Vision Volume 12 Issue 11 Version 1.0 Year 2012

[14] Jinping Fan, Yonglin Zhang, Ruichun Wang, Shiguo Li, “A Separating Algorithm for Over-
lapping Cell Images”, Department of Electronic Communication Technology, Shenzhen Insti-
tute of Information Technology, Shenzhen, China; Institute of Optoelectronic Engineering, Ji-
nan University, Guangzhou, China

[15] Miguel Fabián Romero-Rondón, Laura Melissa Sanabria-Rosas, Lola Xiomara Bautista-
Rozo, Alfonso Mendoza-Castellanos, “Algorithm for detection of overlapped red blood cells
in microscopic images of blood smears”, Universidad Nacional De Colombia, 2016

[16] Michael Kass, Andrew Witkin, and Demetri Terzopoulos, ”Snakes : Active Contour Models”,
International Journal Of Computer Vision, 1987

[17] Tien-Chien Chang, Shu-Yuan Chen, ”Character Segmentation Using Convex Hull Tech-
niques”, Yuan-Ze University, Taiwan, 1999

[18] Otsu’s Thresholding Method


http://www.labbookpages.co.uk/software/imgProc/otsuThreshold.html

[19] Unsharp Masking in MATLAB Image Processing


https://in.mathworks.com/help/images/ref/imsharpen.html

[20] Watershed Transform in MATLAB Image Processing


https://in.mathworks.com/help/images/ref/watershed.html

[21] Marker Controlled Watershed Algorithm in MATLAB Image Processing


https://in.mathworks.com/help/images/marker-controlled-watershed-segmentation.
html;jsessionid=e0202274f87693439cb599f21e09

32
[22] Morphological Operations on binary images
https://in.mathworks.com/help/images/ref/bwmorph.html

[23] Active Contour Implementation in MATLAB


https://in.mathworks.com/help/images/ref/activecontour.html

[24] Generation of convex hull on a binary image


https://in.mathworks.com/help/images/ref/bwconvhull.html

[25] Thresholding Active Contours, Surgical Planning Laboratory,Boston MA, USA


http://www.spl.harvard.edu/publications/item/view/1579

[26] Labelled Microscopic Human Blood Smear Image


Image Courtesy : Slideshare
https://www.slideshare.net/TeaNonikashvili/1-human-blood-smear

[27] Preparation of a microscopic slide for analysis of human blood


Image Courtesy : Nuffield Foundation
http://www.nuffieldfoundation.org/practical-biology/closer-look-blood

[28] Red Blood Cell with Central Pallor


Image Courtesy : Haematology Outlines Textbook
http://www.hematologyoutlines.com

[29] Isfahan MISP Datasets, Blood Cell Archives


http://www.biosigdata.com

33

Das könnte Ihnen auch gefallen