2017 2nd IEEE International Conference On Recent Trends in Electronics Information & Communication Technology (RTEICT), May 19-20,

2017, India

Computer Aided Leukemia Detection using

Digital Image Processing Techniques
Shaikh Mohammed Bilal N Sachin Deshpande
Department of Computer Engineering, Department Of Computer Engineering,
University Of Mumbai, University Of Mumbai,
Bilals.sb@gmail.com Sachin.deshpande@vit.edu.in

Abstract—The machine-controlled identification of malignant and classification of Acute leukemia subtype after
cells from microscopic diagnostic assay pictures of blood determining the type of Acute leukemia.
samples helps in assuaging the diagnostic problems of leukemia Acute leukemia is of two types namely, Acute
and provides higher results if the biologically explainable and Lymphoblastic Leukemia (ALL) and Acute Myeloblastic
clinically important feature primarily based approaches are
Leukemia (AML). Further, acute leukemia is divided into
used for the identification of malignancy and the severity of the
disease. Identification method might have error rates up to the subsequent subtypes:
40%; Due to this drawback it is necessary to develop a second 1. ALL subtypes are L1, L2, and L3.
opinion tool which helps the doctors and pathologists in a 2. AML subtypes are M0, M1, M2, M3, M4, M5, M6, and
classified detection of the leukemia disease. The analysis shows M7.
a discourse method for leukemia and its subtypes detection
from microscopic blood cell images which can be done by B. Exiting Detection Systems
applying image processing techniques followed by a detailed Because of the irregular staining and overpopulated cell
classification analysis on the dataset images. smears the segmentation of cytoplasm and nucleus is
difficult. SVM along with Discrete Fourier Transform and
Keywords—leukemia, detection, segmentation, feature segmentation was used in some works, which in turn
extraction, classification, subtypes. provided an automated learning mechanism to differentiate
the blood smear image specifications (i.e cytoplasm cells,
I. INTRODUCTION
nucleus and platelets). This method is more acceptable and
Leukemia is a type of cancer that originates in the bone efficient in contrast with the “Tresholding and the
marrow of the subject. It occurs as a result of large Wwatershed Algorithms”, as in [1]. Moreover, a similar
production of immature leucocytes which override the version of this method named as “simulated visual
normal blood cells (WBC, RBC, and platelets). The body attention via learning by on-line sampling” is projected in
gets exposed to multiple other diseases due to the decreased [2]. Due to the issue of overlapping cells in blood images
WBC efficiency to fight against the same. within the some algorithms were introduced. They usually worked at
diagnosing of cancer, additionally to think about the signs splitting and merging cells at the edges, as shown in [3].
and symptoms of the subjected patient, it's needed to Detailed classification was deployed by Mohapatra et. al.
identify the malignant or blast cells. To get the amount of using various classifiers to classify the healthy and
the blood cells in every unit volume a blood count is cancerous cells, which is stated in [4]. Lim et. al. projected
calculated for different types of cells in blood(RBC, WBC, on a method thereby showcasing “thresholding,
and platelets). For the determination of the number of morphological operations, and watershed mechanism” as
malfunctioning leucocytes, a bone marrow cellular analysis shown in [5].
is carried out after looking for abnormalities in the blood
count. A diagnostician examines the microscopic cell III. LITERATURE SURVEY
images in correlation to a research to determine the As shown in Table 1, Aswin. R.B and J. Abdul et al. in
abnormalities in the leucocytes in order to observe the “Detection and Classification of Cancer from Microscopic
occurrence of blood cancer and further its type and subtype. Biopsy Images Using Clinically Significant and Biologically
Classification decides the prescription for the subjected
Interpretable Features” 2014(IEEE) used a hybrid GA ANN
patient of blood cancer, hence it is very important. Although
classifier after applying image pre-processing (dull razor &
Flow Cytometry is used by various pathologists as a
dependable mechanism for blood cancer diagnosis, still in filtering), segmentation (color threshold seg) and feature
many medical institutes an hospitals it is not feasible, mainly extraction (grey level co occurrence matrix-GLCM)
the public sector medical hospitals, as described in [11]. techniques on images of skin for melanoma detection with
During this work, the type and subtype of Acute Leukemia an accuracy of 88% [7].
can be identified by working upon the information provided Rajesh Kumar in his paper “Detection and
by the microscopic blood cell images of the dataset of Classification of Cancer from Microscopic Biopsy Images
different human subjects or patients. Using Clinically Significant and Biologically Interpretable
Features” 2015 showed an approach to detect and classify
II. PROBLEM DEFINITION cancer from blood cell images specific pathological features,
A. Statement Of Problem which is shown in [6]. For segmentation of images color ݇-
means based method is used. The various hybrid features
To analyze blood cells digital pictures from blood smear
which are extracted from the segmented images include
samples, then perform a discourse approach for the detection
shape and morphological features, GLCM texture features,
Tamura features, Law’s Texture Energy based features,
978-1-5090-3704-9/17/$31.00 © 2017 IEEE
344
 2017 2nd IEEE International Conference On Recent Trends in Electronics Information & Communication Technology (RTEICT), May 19-20, 2017, India
TABLE 1 1) Differentiating the Cytoplasm (background) and
Sr. Paper Author Advantages Drawbacks
No.
1. “Computer Aided Detection of Aswin R. B. Hybrid Classification Cell overlapping not addressed,
Skin Cancer” [7][16] and Abdul J. Approach, Can be used for detecting melanoma
Effective feature only,
extraction [7]. Accuracy is comparatively less
2. “Detection and Classification of Rajesh K.Texture, shape and It was observed that the proposed
Cancer from Microscopic Biopsy morphology, HOG, method is performing better for
Images Using Clinically Significant wavelet color, connective tissues type sample and
and Biologically Interpretable Tamura’s feature, was not tested for leukemia blood
Features” [6]16] and LTE used smear samples.
innovatively with an
accuracy of
92 % [6].
3. “Detection Of Skin Cancer Using Greeshma Cell segmentation, Cell overlapping not addressed,
Hybrid of SVM-ID3 Algorithm” Rajan1, G Third Harmonic Texture features not addressed.
[13] Shivaraj Generated
Microscopy, SVM,
ID3, Hybrid of ID3-
SVM Algorithm,
Nucleus-to-
Cytoplasm (NC) ratio
implemented in a
simple effective
manner [13].
4. “Computer Aided Diagnostic Ammara M. Detailed comparison Limited morphologic
Support System for Skin Cancer: and classification of Information.
A Review of Techniques and various models used Blood cancer not involved.
Algorithms” [14] [17] in diagnosis of
cancer such as
ultrasound and optic
coherence
tomography [14].
Nucleus( foreground) :
histogram of oriented gradients, wavelet features, and color
features. For classification purposes, ݇-nearest neighbor
based method was proposed to be used with an average
accuracy of 92.19%.
IV. PROPOSED SYSTEM
A. Issues Focused
We propose cell classification model using various
cellular morphological features. For this purpose, it is
necessary to Carry out pixel level pre processing and the
separation of cytoplasm and nucleus. This approach
considers segmentation as an important task for the Figure 1. L A B color space by CIE for the dataset image.
differentiation of blood cells. Segmentation proves to be
accurate and has effective results on such medical diagnostic Fig. 1 shows the three channels for a same blood cell
machine based approach. Hence as an option we use a cell dataset image We set to use the ‘CIE l_a_b_color format’ as
segmentation procedure that considers colors, textures, and in [18]. we have a tendency to took this call as it marks the
other features even in overpopulated cell smears of blood at changes in color groups. Moreover it provides a color
the pixel level image evaluation thereby extracting the differntiation feature based on senses and has high results of
cytoplasm and nucleus in the same. The cytoplasm evalualation as shown in [12]. As the dataset’s blood cell
background cells and nucleus differentiation makes this images are in RGB color space, it has to converted to CIE L
work enhance its results in contrast with other works. a b space using mathematical rules.

B. Proposed Methodology
The proposed mechanism is as follows:

345
 2017 2nd IEEE International Conference On Recent Trends in Electronics Information & Communication Technology (RTEICT), May 19-20, 2017, India

3) Wold’s decomposition

Figure 4. Wold’s Decomposition Procedure.

We need to determine the texture specifications of the

dataset image of the blood cells. For this, ‘Wold’s
Figure 2. Forming groups by the similarity features from channel intensity
Decomposition’ can be used. It consists of fields for texture
Channel L is called the luminosity channel which focuses on evaluation of the image, namely a) Harmonic Field, b)
the brightness or luminosity constrain of the image with Stochastic Field and c) Evanescent Field, as stated in [9].
specific distinctive reflections as shown in Fig 2. Color The texture function is the total of evanescent and harmonic
description of cells is done by channel B, which marks the elements, generalized. As in Fig 4, It can be seen that, in
presence of blue and purple shades. A clustering approach order to initialize the harmonic field, it has to be followed by
can be used keeping in picture the channel L and B values a sine transformation by DFT or FFT for the specific
for different cells for producing two groups having values of channel of all orthogonal components. Hough transform
k as 2 and 3 respectively as shown in fig 2. along with DFT can be used to activate the evanescent field,
generating four evanescent lines. The three fields correspond
2) Finding overlapping cells in highly populated smears: to three texture specifications as regularity, direction
constrains, and random variation constrain.

4) The Classification Process

The following flowchart represents the classification
paradigm of the system.

Dataset

Attribute selection

Class labeling

Cross validation and

classification
Figure 3. Cell Separation process

Fig 3 shows a cell separation procedure. In some works Performance

the overlapped region was identified initially and the evaluation
separation was done by getting the edges from the region of
interest and then finding the midpoint surrounded by the Results
curved portions as the portion of operation as shown in [3].
Then, we carry out linear interpolation or Discrete Fourier
Transform to create a graph of the discontinuities to Figure 5. Classification model
determine overlapping cells in the smear image followed by
morphological operations in MATLAB.

346
 2017 2nd IEEE International Conference On Recent Trends in Electronics Information & Communication Technology (RTEICT), May 19-20, 2017, India

Fig. 5 shows an easy way for developing an acute

leukemia diagnosing algorithm supported by a mixture of
coupled classifiers using SVM (Support Vector Machine)
along with ANN (Artificial Neural Network). SVM will be
functioning with the majority voting approach to determine
the type and subtype of Acute Leukemia along with ANN
since the dataset needs to be trained effectively to reduce the
chances of error. In each step we perform classification to
distinguish between class of acute leukemia (ALL and
AML), subclasses of ALL (L1 and L2), subclasses of AML
(M2, M3, M5). For ALL subclass, we can perform binary as
well as multiclass classification to analyze the features of
subtype. By Fusing different Classifiers (ANN and SVM)
and the use of demonstrative features of nucleus and
Figure 6. Classification procedure cytoplasm cells it is possible to reduce the number of false
positives and false negatives in both classifiers.
Feature extraction of the cells contained in the
cytoplasm as well as the nucleus leads to the classification V. OBJECTIVE
of blood cells in terms of malignant cells and healthy cells, Increase efficiency & accuracy (more than 90%)
furthermore determining its subtypes, as shown in [6]. In thereby decreasing time complexity of each task in the
this section, we evaluate the different features such as algorithm i.e. using a simple fast algorithm such as DFT or
geometry, color, texture, size and statistics from the portions FFT along with above discussed procedures in MATLAB
of interest popped up after image segmentation process, and for image processing, SVM (Support Vector Machine) and
we use them to spot the leukemia type followed by its ANN (Artificial Neural Network) for classification using
subtype. The following table 2 shows some of the WEKA instead of using algorithms that increase complexity
distinguishing features the analysis of whose can be done as shown in [8] and [10] respectively. Classified diagnosis
using different training sets, selection of attributes and leading to sublevel detection of the acute leukemia disease
classification programs supported in WEKA [8]. by applying image processing methods using MATLAB
followed by the classification and subtype classification
Table 2
using SVM along with ANN.
Sr Feature Type Information VI. FIELD WORK
.
The TATA Memorial Hospitals (TMH), Mumbai was
no
visited by me as part of the project detailing and research
1 Area Geometry Total no. of pixels in
and analysis of the topic. It was reported by the senior
the Region Of Interest.
pathologist Dr. Mani Subramanian of the Hematology
2 Circularity Geometry Perimeter2/4πarea
department that for the blood cells count and determination
3 Eccentricity Geometry Distance(centre, focus) of abnormalities they used flow cytometry method
4 Elongation Geometry Width/length of ROI. implemented by Beckman Coulter imported by- Beckman
5 Mean Statistics Avg. no. of pixels value Coulter, Inc. The COULTER VCS uses unit volume,
in ROI. conductivity and scatter light parameters for the
6 Mode Statistics Highest Frequency determination of WBC and its diagnosis, which is shown in
value of pixels in ROI. the manual at [11], and elaborated as follows:
7 Standard Statistics Standard deviation of A. Volume Analysis:
Deviation pixels value in ROI. The unit volume of the individual blood cells is
8 Sum Statistics Total of pixels value in evaluated by the Beckman Coulter using very low frequency
ROI current. This method has been used since ages to evaluate
9 Contrast Color Intensity variation the WBC count in a blood sample, which is further
between a pixel and its processed for the diagnosis of leukemia.
neighborhood pixel B. Conductivity Analysis:
over the image High frequency current passes easily through cell walls
10 Correlation Color Similarity constrain of the blood. As they are good conductors, the evaluation of
between two neighbor insulation at the walls of the cell as well as the specifications
pixels. of the nucleus components along with the chemical uptake
11 Entropy Texture Smoothness determined and traces inside the cell is evaluated.
by gray levels. C. Light Scatter Analysis:
12 Homogeneity Texture Correlation between the Scattering of laser light throughout the population of
elements in co cells contained in a blood sample of a patient also known as
occurrence matrix to its Fulwyler’s method for cell analysis, is deployed in Beckman
diagonal elements. Coulter so as to determine the blast cell count, the
containment proportion of the cells and the overpopulated
regions.

347
 2017 2nd IEEE International Conference On Recent Trends in Electronics Information & Communication Technology (RTEICT), May 19-20, 2017, India

VII. DATASET [11] COULTER HmX Hematology Analyzer, PN 4237523CB (March

2011) Beckman Coulter, Inc.250 S. Kraemer Blvd.Brea, CA 9282.
As urged by the doctors of the Tata Memorial [12] Paschos G. Perceptually uniform color spaces for color texture. IEEE
Hospital, Mumbai; the request to accumulate the dataset was Transactions on Image Processing. 2011; 10(6): 932–937.
sent to Fabio Scotti, University of Milano, Italy. The [13] Greeshma Rajan et al, International Journal of Computer Science and
database ALL IDB was gathered that consisted of Mobile Computing, Vol.3 Issue.5, May- 2014, pg. 143-153.
ALL_IDB1 and ALL_IDB2 datasets consisting of pictures [14] Ammara Masood, “Computer Aided Diagnostic Support System for
Skin Cancer:A Review of Techniques and Algorithms” Hindawi
that were captured by optical laboratory magnifier as well as Publishing Corporation, International Journal of Biomedical Imaging
a Canon Power Shot G5 camera. The resolution of the JPEG Volume 2013, Article ID 323268.
format 24 bit pictures is 2592 x 194, which is stated in [15]. [15] F. Scotti “ALL_IDB: the acute lymphoblastic leukemia image
ALL_IDB1 consists of 108 pictures (healthy and database for image processing”, 2011 IEEE, Brussels, Belgium, pp.
malignant) having 39000 blood components in which the 2045-2048, ISBN: 978-1-4577-1302-6.
lymphocytes are labeled by skilled oncologists. [16] www.scilit.net
ALL_IDB2 dataset is for testing and training the [17] www.semanticscholar.org / http://dx.doi.org/10.1155/2013/323268
classification system. It's a set of separated regions to be [18] C. Reta. " Leukocytes Segmenta ion Using Markov Random Fields"
, Advances in Experimental Medicine and Biology, 2011.
considered pertaining to healthy and leukemia cells from
ALL_IDB1 dataset.
ACKNOWLEDGMENT
I wish to offer my sincere gratitude to our principal Dr.
S.A. Patekar for providing me an opportunity to be a part of
the Vidyalankar Institute of Technology, Mumbai under the
postgraduate program and moreover to allow me work on
this final year Project on the respective domain.
I sincerely thank my Guide Prof. Sachin Deshpande for
the guidance and encouragement in carrying out this work.
I also wish to thank the Expert Panel members who
carried out the reviews on this seminar topic for myself and
other students and rectified the mistakes in the same for the
betterment of our work.
I would like to offer gratitude to Dr. Subramanian and
the other pathologists of the Hematology Department, TMH,
Mumbai for their support and information gathering.
REFERENCES
[1] Pan C, Lu H, Cao F. “Segmentation of blood and bone marrow cell
images via learning by sampling.” Emerging Intelligent Computing
Technology and Applications, Springer Berlin Heidelberg. 2009;
336–345.
[2] Pan C, Park DS, Yoon S, Yang JC. “Leukocyte image segmentation
using simulated visual attention.” Expert Systems with Applications.
2012; 39(8): 7479–7494.
[3] Wang W, Hao S. “Cell cluster image segmentation on form analysis”.
IEEE Third International Conference on Natural Computation. 2007;
4: 833–836.
[4] Mohapatra S, Patra D. “An ensemble classifier system for early
diagnosis of acute lymphoblastic leukemia in blood microscopic
images.” Neural Computing and Applications. 2014; 24(7–8): 1887–
1904.
[5] Lim HN, Mashor MY, Hassan R. “White blood cell segmentation for
acute leukemia bone marrow images.” IEEE International Conference
on Biomedical Engineering (ICoBE) 2012; 357–361.
[6] Aswin.R.B and J. Abdul ”Detection and Classification of Cancer from
Microscopic Biopsy Images Using Clinically Significant and
Biologically Interpretable Features.” 2014(IEEE).
[7] Rajesh Kumar,” Detection and Classification of Cancer from
Microscopic Biopsy Images Using Clinically Significant and
Biologically Interpretable Features” Hindawi Publishing Corporation
Journal of Medical Engineering ,Volume 2015, Article ID 457906,
2015.
[8] Weka—Hall M, Frank E, Holmes G, Pfahringer B, Reutemann P,
Witten IH. “The WEKA data mining software: an update”; ACM
SIGKDD Explorations Newsletter. 2009; 11(1): 10–18.
[9] Liu F, Picard RW. “A spectral 2D Wold decomposition algorithm for
homogeneous random fields”. IEEE Proceedings of the International
Conference on Acoustics, Speech and Signal Processing. 1999; 6:
[10] 3501–3504.Pankaj Agrawal, S.K.Shriwast ava and S.S.Limaye,
“MATLAB Implementation of Image Segment at ion Algorithms”,
IEEE Pacific, pp.68–73.

348