Research Article Segmentation of White Blood Cell from...

Research ArticleSegmentation of White Blood Cell from Acute LymphoblasticLeukemia Images Using Dual-Threshold Method

Yan Li12 Rui Zhu1 Lei Mi1 Yihui Cao12 and Di Yao3

1State Key Laboratory of Transient Optics and Photonics Xirsquoan Institute of Optics and Precision Mechanics ofChinese Academy of Sciences Xirsquoan 710119 China2University of Chinese Academy of Sciences 52 Sanlihe Road Beijing 100864 China3Shenzhen Vivolight Medical Device and Technology Co Ltd Shenzhen 518000 China

Correspondence should be addressed to Yan Li liyanoptcn

Received 30 December 2015 Revised 7 April 2016 Accepted 21 April 2016

Academic Editor Jayaram K Udupa

Copyright copy 2016 Yan Li et al This is an open access article distributed under the Creative Commons Attribution License whichpermits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

We propose a dual-threshold method based on a strategic combination of RGB and HSV color space for white blood cell (WBC)segmentation The proposed method consists of three main parts preprocessing threshold segmentation and postprocessing Inthe preprocessing part we get two images for further processing one contrast-stretched gray image and one H component imagefrom transformed HSV color space In the threshold segmentation part a dual-threshold method is proposed for improving theconventional single-threshold approaches and a golden section search method is used for determining the optimal thresholds Forthe postprocessing part mathematical morphology and median filtering are utilized to denoise and remove incomplete WBCsThe proposed method was tested in segmenting the lymphoblasts on a public Acute Lymphoblastic Leukemia (ALL) image datasetThe results show that the performance of the proposed method is better than single-threshold approach independently performedin RGB and HSV color space and the overall single WBC segmentation accuracy reaches 9785 showing a good prospect insubsequent lymphoblast classification and ALL diagnosis

1 Introduction

Automatic white blood cell segmentation which plays animportant role in automatic blood cell morphology analysisremains a challenging issue because of the morphologicaldiversity of WBCs and the complex background of bloodmicroscopic images In this paper our focus is the segmen-tation of lymphoblastmdashone kind abnormal white blood cellfrom Acute Lymphoblastic Leukemia images for ALL classi-fication

Acute Lymphocytic Leukemia also known as Acute Lym-phoblastic Leukemia is a serious hematic disease character-ized by the overproduction and continuous multiplication ofmalignant and immature WBCs (referred to as lymphoblastsor blasts) It is fatal if left untreated due to blastsrsquo rapid spreadinto the bloodstream and other vital organs Fortunatelyearly diagnosis of the disease is helpful and beneficial to therecovery of patients especially in the case of children [1]

According to the FAB (French-American-British) clas-sification standard built on blast cellrsquos shape and patternvariability ALL can be classified into three types L1 L2 andL3 (Figures 1(b)ndash1(d)) [2] Accurate classification can offerdoctors useful information for therapeutic plan selection andfollow-up work The microscope inspection of blood smearsby counting different kinds of WBCs is one of the mostfrequently performed blood tests by hematologists [3] Fordecades the operation is performed by experienced opera-tors However this is a time-consuming and tedious work Inaddition diagnosis results always tend to be subjective andimprecise and also are difficult to be reproduced afterwards

As a substitute automatic identification technology ofWBCs known for low-cost homogenous accuracy has gainedmore and more focus in domain of hematic related diseasediagnosis In general the system consists of four parts WBCsegmentation feature extraction classification and countingSince segmentation results directly influence the accuracy of

Hindawi Publishing CorporationComputational and Mathematical Methods in MedicineVolume 2016 Article ID 9514707 12 pageshttpdxdoiorg10115520169514707

2 Computational and Mathematical Methods in Medicine

(a) (b) (c) (d)

Figure 1 Morphological variability associated with the blast cells according to the FAB classification (a) Healthy lymphocytes cell fromnon-ALL patients (bndashd) Lymphoblasts from ALL patients where (b) (c) and (d) are L1 L2 and L3 respectively

classification and counting it has become a very hot topic inclinical diagnosis [4]

Some work on microscopic WBC image segmentationis available in the literature [3ndash25] Methods can be clas-sified into three types threshold-based methods patternrecognition-based methods and deformable model-basedmethods Threshold-based methods include Otsursquos method[4ndash6] the region growing method [7] watershed method[8ndash10] and their combination [11] Cseke [4] presented afast segmentation schemewith automatic thresholding wherethresholds are selected with a simple recursive methodderived from maximizing the interclass variance betweendark gray and bright regions based on the method proposedby OtsuThemethod works well for nucleus and backgroundsegmentation However it cannot separate cytoplasm fromthe red blood cells Wu et al [6] developed an iterativeOtsursquos threshold approach based on circular histogram forthe leukocyte segmentation using HampS components of HSImodel Experimental results show that the method workssuccessfully in the segmentation of WBC nucleus but losesthe cytoplasm informationDorini et al [11] divided theWBCsegmentation process into two steps In the first step theyextracted the cell nucleus using the watershed transformby Image Forest Transform (IFT) Then they segmentedthe WBC cytoplasm using basic operations such as thresh-olding and morphological opening via the size distributioninformation of the RBC but the method tends to produceoversegmentation in presence of noise

As leukocytes in microscopic images can be treated asobjects pattern recognitionmethods are also used to performthe segmentation which can be categorized as supervised orunsupervised [12] Supervised methods classify the objectsusing learning-based approaches such as Support VectorMachine (SVM) [13] and Artificial Neural Network (ANN)while unsupervised methods also known as clustering meth-ods mainly including 119896-means clustering [14ndash16] fuzzy 119862-means [17] and expectation-maximization extract the objectsfrom the data itself In [13] Guo et al proposed multispectralimaging techniques with a spectral calibration method toacquire device-independent images and then applied SVMdirectly to the spectrum of each pixel to segment thewhole microscopic image into four types of regions nucleus

cytoplasm erythrocytes and background Segmentationresults are satisfactory but the implementation speed needsto be boosted

Deformable model-based methods can be classified intoparametric models and geometric models based on contourrepresentation Besides level set method [9] the active con-tour model also known as a snake is the most common [1819] In [18] Ko et al introduced a new WBC image segmen-tation method using stepwise merging rules based on mean-shift clustering and boundary removal rules with a gradientvector flow (GVF) snake Removal rules are used to removethe boundary and noise edges while a GVF snake is forced todeform to the cytoplasm boundary edges Due to a weak dif-ference between the cytoplasm and the background or con-tact with RBCs some experimental results were slightly over-segmented Other segmentation methods for WBCs asidefrom the above three categories aremorphological operations[3 11 20] hybrid methods [9 10 19 21] and so on Hybridmethods combine two or more methods mentioned aboveto achieve better results such as combination of a Lab colorspace based segmentationmethod and a gray level threshold-ing method [21] and combination of an Otsu method and anactive contour method [19]

Generally the ultimate goal of WBC segmentation is toextract whole WBC from a complicated background andsegment every WBC into morphological components suchas nucleus and cytoplasm Among the methods mentionedabove RGB color space based threshold approaches aremost widely used due to their high efficiency and reliabilityHowever cytoplasm has a big variance of color and its coloris quite similar to image backgroundThus in many previousworks gray image based threshold methods are only utilizedto segment nucleus (Figures 2(a) and 2(d)) As to cyto-plasm extraction (Figure 2(c)) other auxiliary segmentationschemes are needed [19] What is more aside from single-threshold-based segmentation methods nucleus and cyto-plasm are respectively segmented with different methods inmany other segmentation schemes too [3 8 9 11 18] To facil-itate cytoplasm segmentation via threshold-based methodssome researchers have turned their attention to segmenting intransformed color space such as HSV [22 23] HSI [24] andLab [21] As an example Eldahshan et al [22] proposed to

Computational and Mathematical Methods in Medicine 3

(a) (b) (c) (d) (e)

Figure 2 Background of dual-threshold method (a d) Segmentation results of single-threshold method based on RGB color space or Schannel image from HSV color space (only nucleus is extracted) (b) Segmentation result of uniform image using single-threshold methodbased on H channel image of HSV color space (nucleus and cytoplasm are all extracted) (c) Expected segmentation result of cytoplasm(e) Segmentation result of image having illumination variations by single-threshold method based on H channel image of HSV color space(much noise exists in segmentation result)

Input

Grayscaling

H component Segmented WBC

Backgroundextraction

RBC removalIntersection

Binarization

Morphological erosion

Medianfiltering

MCR extraction

Contrast stretching

Thresh1Thresh2I

G

HS

Postprocessing

RGB

Threshold segmentationPreprocessing

G998400

H998400

Figure 3 Flowchart of the proposed dual-threshold segmentation scheme

segment WBC from its background using hue channel ofHSV color space based single-threshold method The resultsshow that the proposed framework works well for uniformimages (Figure 2(b)) but inconsistent for the images havingillumination variations (Figure 2(e))

In this paper we propose a general method to segmentWBC (both nucleus and cytoplasm are included) from imagebackground regardless of illumination variations by effec-tively combining the RGB color space based single-thresholdmethod and HSV color space based single-threshold methodto construct a new technique which is named dual-thresholdmethod to overcome the weakness of the component meth-ods In this way cytoplasm can further be segmented by sub-tracting nucleus which can be obtained through various intu-itivemethods such as Otsursquos method [6] and 119896-means cluster-ing [14] from the above result easilyThough it is not includedin this paper To determine the value of the two thresholdsgolden section search method is selected At last except

qualitative assessment we also make a quantitative compari-son of ourmethodwith two single-threshold-basedmethodsrespectively implemented in RGB and HSV color space [23]in terms of a widely used evaluation metric DSC in imagesegmentation field

The rest of this paper is organized as follows In Section 2we elaborate the completemethodology of the dual-thresholdalgorithm In Section 3 we describe an evaluation of thismethod in terms of its accuracy and robustness In Section 4concluding remark is given

2 Materials and Methods

21 Overview of the Approach The proposed dual-thresholdmethod consists of three phases preprocessing thresholdsegmentation and postprocessing Figure 3 presents anoverview of the proposed approach In the preprocessing


(a) (b) (c)

Figure 4 Global contrast stretching (a) Original image (b) Gray image (c) Contrast-stretched image

phase we get two images for further processing one contrast-stretched gray image and one H component image fromtransformed HSV color space The threshold segmentationphase consists of three main steps image background extrac-tion red blood cell separation and the optimal thresholdselection In the background extraction step RGB color spacebased single-threshold method is employed while in redblood cell separation step H channel image based single-threshold method is utilized For the optimal thresholdselection part golden section search method is used Finallyin the postprocessing phase mathematical morphology andmedian filtering are utilized to denoise and remove incom-plete WBCs The whole process can be summarized as thefollowing five steps The details of each step are given in thefollowing subsections

Step 1 Given an input image I get its gray-scaling image G aswell as H component image H from transformed HSV colorspace

Step 2 Do contrast stretching operation to G and then usethreshold Thresh1 to extract image background which isshown in black in G1015840 (Figure 3)

Step 3 Through threshold Thresh2 separate the red bloodcells from H (resultant image is H1015840)

Step 4 Get a gray image S by intersecting G1015840 with H1015840

Step 5 Binarization morphological erosion and medianfiltering are performed on S followed bymaximumconnectedregion (MCR) extraction to get a binary image fill smallholes connect narrow gaps and remove small dots as well asincomplete WBCs in the image

22 Preprocessing In this step two roughly processed imagesare obtained for further processing At first we convert thecolor image into a gray one (Figure 4(b)) At the same timetransform the source RGB color space into HSV color spaceand then extract H channel image denoted by H in Figure 3from HSV color space From Figure 4 we can see thatthe contrast between foreground and background pixels inthe grayscale is typically not sufficient to classify the pixelsprecisely To increase the contrast of the image the global

contrast stretching (GCS) techniquewhich can spread out therange of scene illumination is employed GCS is performed bysliding a window (called the KERNEL) across the image andadjusting the center element using the following formula [25]

119868119901(119909 119910) =

255 ∙ [119868119900(119909 119910) minusmin]

(max minusmin) (1)

where (119909 119910) is the coordinate of the image pixel 119868119901(119909 119910) is

the output color level for the pixel (119909 119910) from the contraststretching process 119868

119900(119909 119910) is the input color level for the pixel

(119909 119910) min and max are the minimum and maximum colorlevel value in the input image

After GCS (Figure 4(c)) we can see that the contrastthroughout the image is equalized and it has been easier tosee the image details in the regions that are originally veryobscure More importantly contrast between foreground andbackground pixels is greatly enhanced which is supposed tofacilitate image background extraction via single-thresholdmethod which will be clarified in the next part

23 Threshold Segmentation

231 Image Background Extraction From Figure 2(e) it canbe seen that noise in final segmentation result mainly comesfrom two aspects background and red blood cells Thefollowing two steps background extraction and red blood cellseparation are all somewhat based on this fact In Figure 5histogram of the contrast-stretched gray image presentsa triple-modal respectively representing white blood cellnucleus red blood cell (cytoplasm) and background AfterGCS background now has a certain contrast with othercomponents in the image so it is feasible to extract it throughsingle-threshold (marked as Thresh1 in Figure 5) methodHowever as it has been stated in Introduction it is still diffi-cult to separate the whole white blood cell (both nucleus andcytoplasm included) from the image in that cytoplasm has asimilar gray intensity with red blood cells So the first stageof our method is to extract the image background shown inblack in Figure 5(c)

232 Red Blood Cell Separation In Section 231 imagebackground has been extracted If we can continue finding


3000

2000

1000

0

0 100 200

(a)

Thresh1

3000

2000

1000

0

0 100 200

(b) (c)

Figure 5 Image background extraction through gray image based single-threshold method (a) Original gray image and its histogram (b)Contrast-stretched image and its histogram (c) Original color image and resultant binary image after background extraction (the extractedbackground is shown in black in (c))

an effective method to remove red blood cells in the imageWBCcan be segmented easily by subtracting background andred blood cells from the original image As a result our goalin this step is to separate the red blood cells from the image

The purpose of a color space is to facilitate the speci-fication of colors in some standard [26] A color space istypically represented by a three- or four-dimension matrix inmathematics such as RGB HSV Lab and CMYK The RGBcolor space is themost common color space used in electronicdevices In this color space each color can be obtained by theaddition of three primary colors red green and blue Gener-ally the original stained blood smear image is represented bythe RGB color space in the RGBmodel The HSV color spacehas three components hue (H) saturation (S) and value(V) Hue represents color In this model it is an angle from 0to 360 degrees Saturation indicates the range of gray in thecolor space It ranges from 0 to 100 Sometimes the valueis calculated from 0 to 1 When the value is ldquo0rdquo the color isgray and when the value is ldquo1rdquo the color is a primary color Afaded color is due to a lower saturation level whichmeans thecolor contains more gray Value is the brightness of the colorand varies with color saturation It ranges from 0 to 100When the value is ldquo0rdquo the color space will be totally blackWith the increase in the value the color space brightness isup and shows various colors The HSV color space is quitesimilar to the way in which humans perceive colorThe colorsused in this space can be clearly defined byhumanperceptionwhich is not always the case with RGB These characteristics

make the HSV color space more suitable for image segmen-tation and analysis than the RGBmodel In Lab color space Ldefines lightness a denotes redgreen value and b representsthe yellowblue value [21] It is known as device independentmeaning the Lab color space can communicate differentcolors across different devices The above color spaces can beconverted into each other according to related formulas

In Figure 6 a comparison of blood cell image respec-tively in RGB HSV and Lab color spaces and each channelimage is given From the figure it can be observed thatin most single-channel images (such as G- S- L- and b-channel) cytoplasm has a similar gray intensity with thebackground and red blood cells making it difficult to beseparated from red blood cells and background throughOtsursquosmethod exceptH channel imagewhere cytoplasmhas asimilar gray intensity with nucleus and a certain contrast withred blood cells Therefore to avoid cytoplasm being removedtogether with red blood cells (which is the case in G-channelS-channel and some other channels) H channel image isselected for red blood cell removal in this step In Figure 7(a)H channel image and its histogram are givenWe can see thatin the image red blood cells are the brightest part So we canremove them through a suitable thresholdmarked asThresh2in the figure Figure 7(b) shows the resultant image after redblood cell separation

233 Threshold Selection The proposed algorithm has twoparameters Thresh1 and Thresh2 which are respectively


R channelRGB G channel B channel

H channel

L channel

HSV

Lab

S channel

a channel

V channel

b channel

Figure 6 Images in RGB HSV and Lab color spaces and their component images (R G B H S V and L a b)

used in image background extraction and red blood cellremoval steps of the white blood cell segmentation From thehistograms of contrast-stretched gray image (Figure 5(b)) andH channel image (Figure 7(a)) we can see that the rangesof Thresh1 and Thresh2 can be roughly estimated through apriori knowledge But in concrete algorithm implementationrandomly selected thresholds even when they are in the rea-sonable scope just cannot guarantee any kind of optimality Sowe formulate threshold selection as an optimization problemwith two variables (Thresh1 and Thresh2) and tend to use anappropriate method to find the optimal solution

To formulate a proper objective function 119891 let us intro-duce Dice Similarity Coefficient (DSC) first which is usuallyused to evaluate segmentation effect quantitatively In thenext section we will use it for quantitative comparison of theproposed method and two other segmentation schemes It isdefined as

DSC (119860 119861) = 2 lowast (119860 cap 119861)(119860 + 119861) (2)

where119860 is the area of the target region of ground truth imageacted by the manually segmented images in this paper 119861 isthe area of the target region of the result of an automaticallysegmented image DSC varies between 0 and 1 The higher it

is the better segmentation accuracy it indicates According tothis fact the objective function 119891 we formulate is

DSC = 119891 (Thresh1Thresh2) (3)

In this way the optimization problem here can bedescribed as follows make DSC be as large as possible byselecting appropriate Thresh1 and Thresh2 Since approxi-mate ranges ofThresh1 andThresh2 can be determined whatwe should do next is to find the optimal value from all thepossible choices

The golden section search method is a technique forfinding the extremum (minimum or maximum) of a strictlyunimodal function by successively narrowing the range ofvalues inside which the extremum is known to exist Thetechnique derives its name from the fact that the algorithmmaintains the function values for triples of points whose dis-tances form a golden ratio known as 0618 It has been provedthat compared to bisection method this value can enable usto obtain an optimal reduction factor for the search intervalandminimal number of function calls when searching for themaximum point

Assume 119891(119909) is a unimodal function in search region[119886 119887] the maximum point is 119909 and the assumed algorithmprecision is epsilon Let 1199091 1199092 be two points in region [119886 119887]and 119886 lt 1199091 lt 1199092 lt 119887 To use the golden section search


Thresh2

1000

500

0

0 05 1

(a) (b)

Figure 7 Red blood cell removal through H channel image based single-threshold method (a) H channel image and its histogram (b)Original image and the H channel image after red blood cell removal

algorithm to determine 119909 the following six steps are included(algorithm flowchart is given in Figure 8)

Step 1 Let [119886 119887] be initial search interval and let algorithmprecision be epsilon

Step 2 Let 1199091 = 119886+0382lowast(119887minus119886) let 1199092 = 119886+0618lowast(119887minus119886)and compute 119891(1199091) 119891(1199092)

Step 3 If 119891(1199091) gt 119891(1199092) 119887 = 1199092 jump to Step 4 if 119891(1199091) lt119891(1199092) 119886 = 1199091 jump to Step 5 if 119891(1199091) = 119891(1199092) 119886 = 1199091119887 = 1199092 jump to Step 6

Step 4 If |119887 minus 119886| lt epsilon 119909 = (119886 + 119887)2 stop Otherwiselet 1199092 = 1199091 119891(1199092) = 119891(1199091) and 1199091 = 119886 + 0382 lowast (119887 minus 119886)Compute 119891(1199091) jump to Step 3


Step 6 If |119887 minus 119886| lt epsilon 119909 = (119886 + 119887)2 stop Otherwisejump to Step 2

In the next section we will use this method to find theoptimal solutions of Thresh1 andThresh2

24 Postprocessing After threshold segmentation whiteblood cells can be roughly segmented (S in Figure 3) At thetime it is still a gray image with some noise so in this step

binarization is used to convert this resultant image to binaryimage at first Then we use morphological erosion medianfiltering (size 15 times 15) and maximum connected region(MCR) extraction operations to fill small holes connectnarrow gaps and remove small dots as well as the incompleteWBCs in the image At last convert the binary image intoRGB color image but this is not a must

3 Results and Discussion

31 Dataset The proposed method was tested on 130 ALLimages taken from ALL IDB [27] a public and free availabledataset provided by Department of Information Technologyspecifically designed for the evaluation and comparison ofalgorithms for segmentation and image classification Theimages of the dataset have all been captured with an opticallaboratory microscope coupled with a Canon Power Shot G5camera For each image in the dataset the classification ofALL lymphoblast is provided by expert oncologists ALL IDBincludes two subsets ALL IDB1 and ALL IDB2 The formeris composed of 108 images containing about 39000 blood ele-ments taken with different magnifications of the microscoperanging from 300 to 500 So the images in the dataset maydiffer in background color (Figure 9)The latter is a collectionof cropped area of interest of normal or blast cells fromALL IDB1

ALL IDB2 which contains single WBC in each image isused for testing the performance of our proposal It has 260imagesThe first half is fromALL patients (lymphoblasts) and


Yes

No

Yes

No

Stop

YesYes

No

Yes

NoNo

Start

[a b]epsilon gt 0

x1 = a + 0382 lowast (b minus a) f(x1)x2 = a + 0618 lowast (b minus a) f(x2)

a = x1b = x2

x2 = x

x1 = a + 0382 lowast (b minus a)

f(x1)

1

f(x2) = f(x1)

a = x1

x1 = x


f(x2)

2

f(x1) = f(x2)

b = x2

x = (a + b)2

f(x1) == f(x2)

f(x1) gt f(x2)

|b minus a| lt epsilon |b minus a| lt epsilon |b minus a| lt epsilon

Figure 8 Flowchart of golden section search method

(a) (b) (c)

Figure 9 Sample images differ in background color caused by different magnifications being used in image taken

the last half is from non-ALL patients (normal WBCs) Thefinal task of us is to classify the lymphoblasts into three classesL1 L2 and L3 (Figure 1) for targeted treatment and follow-up so only the first 130 lymphoblast images of ALL IDB2 areconcerned in this study But experimental results show thatthe proposedmethod alsoworkswell for the other 130 imagesTwo samples from ALL IDB1 and ALL IDB2 are shown inFigure 10 Subimage which contains only one WBC can beextracted frommulti-WBCs image through a proper method[4 6 20]

32 Experimental Results We evaluate the performance ofour proposed algorithm both visually and quantitatively in

this part To this end we find the optimal value of Thresh1and Thresh2 first through the iterative golden section searchmethod described in Section 2 Two doctors are invited tomanually segment all test images for the purpose of gener-ating ground truth for evaluation

321 Threshold Selection In the context of this paper ourgoal is to find two optimum thresholds to make value of theobjective function119891(Thresh1Thresh2) (3) be as big as possi-ble Since there are two variables in this function the goldensection searchmethodneeds to be used several times to deter-mine their respective optimum value Each time one of themis fixed to find the other variablersquos optimum value until both


(a) (b)

Figure 10 Acute Lymphoblastic Leukemia images from the dataset (a) ALL IDB1 (b) ALL IDB2 Arrow indicates that (b) is cropped from(a)

of them are unchangedThe idiographic step of thismethod isgiven in the following

Step 1 Assume that the optimum values of Thresh1 andThresh2 are respectively 1199051 1199052 To make the procedureoperate initial value of 1199052 is set at 230

Step 2 Let Thresh2 = 1199052 119886 = 08 119887 = 10 epsilon = 001119909 = Thresh1 and 119891(119909) = mean(DSC(130)) and computethe optimum value of Thresh1 1199051 through the golden sectionsearch method

Step 3 Let Thresh1 = 1199051 119886 = 150 119887 = 255 epsilon = 5 119909 =Thresh2 and119891(119909) = mean(DSC(130)) and compute the opti-mum value of Thresh2 1199052 through the golden section searchmethod If 1199052 = 230 repeat Step 2 and Step 3 Otherwise stop

Note

(1) The numbers above 230 08 10 001 150 255 and 5are selected through a priori knowledge

(2) mean(DSC(130)) is mean of all 130 test imagesrsquoDSC value obtained after segmentation on conditionThresh1 = 119909 Thresh2 = 1199052 (in Step 2) or Thresh1 =1199051 Thresh2 = 119909 (in Step 3)

Through the above steps optimal values of Thresh1 andThresh2 1199051 1199052 can be determined

322 Qualitative Evaluation Through the above step theoptimal values Thresh1 = 09 Thresh2 = 220 are obtainedSegmentation results of the 130 test images are comparedwith those of manual segmentation as well as two single-threshold methods based respectively on RGB [18] andHSV [15] color space The segmentation result is consideredaccurate when the autodetected boundary closely matchesthe manually traced boundary In Figure 11 we give five testimagesrsquo segmentation resultsThreemethodswere carried outon the same morphological structure element and median

Table 1 Mean and standard deviation of all test imagesrsquo DSC value

Method 1 Method 2 Method 3Mean 09542 07737 09795Standard deviation 00403 02163 00144

filter sizes Results present thatMethod 1 shows a goodperfor-mance in nucleus segmentation but the cytoplasm cannot besegmented in some cases (the 1st 2nd 4th and 5th cells) Asto Method 2 it segments the WBC perfectly in some cases(the 1st and 2nd cells) but does not in others (the 3rd 4thand 5th) On the contrary our proposed Method 3 performswell in all the cases

It has been shown that the proposed method achieves ahigh accuracy in segmenting single lymphoblast frommicro-scopic images in Figure 11 Figure 12 shows that the methodalso performs well in segmenting normal WBCs in the latterhalf part of ALL IDB2 and when multiple WBCs exist in oneimage (in this case MCR in postprocessing step should beremoved)

323 Quantitative Evaluation One kind quantitative depic-tion of segmentation performance is given in Figure 13where DSC values of the 130 test images are calculated aftersegmentation with the proposed method and two single-threshold-based methods It can be seen that DSC values gotthrough our method are higher and stable than the othertwo methods in most cases To those images in latter partof all the test images Method 2 has a good performance butthat cannot be extended to whole dataset As has been saidin dataset part images in ALL IDB are taken with differentmagnifications so some of them differ in background color(Figure 9) Thus H channel image based single-thresholdmethod does not perform well all the time Mean andstandard deviation of the 130DSCs are given inTable 1 By cal-culation we can learn that our method has 266 and 266higher accuracy over Method 1 and Method 2 respectivelyin terms of DSC mean value metric Meanwhile its standard


(a) (b) (c) (d) (e)

Figure 11 Comparison of the proposed method and the manual method as well as single-threshold methods (a) Original image (b) Manualsegmentation results (ground truth) (c)M1 RGB color space based single-thresholdmethod (d)M2HSV color space based single-thresholdmethod (e) M3 our proposed method

deviation is 6438 and 9336 lower than them suggestingour method is also more robust than the other two

4 Conclusion

In this paper we have proposed a dual-threshold methodfor segmenting white blood cells from Acute LymphoblasticLeukemia images The method effectively combines RGBand HSV color space based single-threshold methods to

exploit their complementary strengths It consists of threemain parts preprocessing threshold segmentation andpostprocessing Background and red blood cells of the imageare extracted via two different thresholds in segmentationprocess The experimental results suggested that an overallsegmentation accuracy of DSC asymp 098 can be achieved Asthe first step of automatic white blood cell differential systemit shows a good prospect in further WBC feature extractionALL classification and diagnosis


(a) (b) (c) (d)

Figure 12 Performance of the proposed method on normal WBCsrsquo and multi-WBCsrsquo segmentation (a c) Original images (b d)Segmentation results

1

09

08

07

06

05

04

DSC

DSC comparison

0 5 10 15 20 25 30

Test image number (times5)

Method 1Method 2

Method 3

Figure 13 Each test imagersquos DSC value obtained using different segmentation methods Method 1 RGB color space based single-thresholdmethod Method 2 HSV color space based single-threshold method Method 3 our proposed method

Competing Interests

The authors declare that there is no conflict of interestsregarding the publication of this paper

Acknowledgments

The authors would like to thank Professor Fabio Scotti forproviding the microscopic Acute Lymphoblastic Leukemia


cell images This work was supported by the National KeyScience and Technology Support Program (2015BA101B06)and Science and Technology Coordination and InnovationPlan (2013KTCQ03-09)

References

[1] V Piuri and F Scotti ldquoMorphological classification of bloodleucocytes by microscope imagesrdquo in Proceedings of the IEEEInternational Conference on Computational Intelligence forMea-surement Systems and Applications (CIMSA rsquo04) pp 103ndash108Boston Mass USA July 2004

[2] F Scotti ldquoAutomatic morphological analysis for acute leukemiaidentification in peripheral blood microscope imagesrdquo in Pro-ceedings of the IEEE International Conference on ComputationalIntelligence for Measurement Systems and Applications (CIMSArsquo05) pp 96ndash101 Messian Italy July 2005

[3] C Fatichah M L Tangel M R Widyanto F Dong and KHirota ldquoInterest-based ordering for fuzzymorphology onwhiteblood cell image segmentationrdquo Journal of Advanced Computa-tional Intelligence and Intelligent Informatics vol 16 no 1 pp76ndash86 2012

[4] I Cseke ldquoA fast segmentation scheme for white blood cellimagesrdquo in Proceedings of the 11th IAPR International Conferenceon Pattern Recognition C Image Speech and Signal Analysis vol3 pp 530ndash533 IEEE 1992

[5] N Otsu ldquoA threshold selection method from gray-level his-togramsrdquo Automatica vol 11 no 285ndash296 pp 23ndash27 1975

[6] JHWu P P Zeng Y Zhou andCOlivier ldquoAnovel color imagesegmentation method and its application to white blood cellimage analysisrdquo in Proceedings of the 8th International Confer-ence on Signal Processing (ICSP rsquo06) vol 2 IEEE Beijing ChinaNovember 2006

[7] J Duan and L Yu ldquoA WBC segmentation methord based onHSI color spacerdquo in Proceedings of the 4th IEEE InternationalConference on Broadband Network and Multimedia Technology(IC-BNMT rsquo11) pp 629ndash632 Shenzhen China 2011

[8] K Jiang Q-M Liao and Y Xiong ldquoA novel white blood cellsegmentation scheme based on feature space clusteringrdquo SoftComputing vol 10 no 1 pp 12ndash19 2006

[9] L B Dorini R Minetto and N J Leite ldquoSemiautomatic whiteblood cell segmentation based on multiscale analysisrdquo IEEEJournal of Biomedical and Health Informatics vol 17 no 1 pp250ndash256 2013

[10] S Arslan E Ozyurek and C Gunduz-Demir ldquoA color andshape based algorithm for segmentation of white blood cells inperipheral blood and bone marrow imagesrdquo Cytometry A vol85 no 6 pp 480ndash490 2014

[11] L B Dorini R Minetto and N J Leite ldquoWhite blood cellsegmentation using morphological operators and scale-spaceanalysisrdquo in Proceedings of the 20th Brazilian Symposium onComputer Graphics and Image Processing (SIBGRAPI rsquo07) pp294ndash301 Minas Gerais Brazil October 2007

[12] M Saraswat and K V Arya ldquoAutomated microscopic imageanalysis for leukocytes identification a surveyrdquoMicron vol 65pp 20ndash33 2014

[13] N N Guo L B Zeng andQ SWu ldquoAmethod based onmulti-spectral imaging technique for white blood cell segmentationrdquoComputers in Biology and Medicine vol 37 no 1 pp 70ndash762007

[14] C Zhang X Xiao X Li et al ldquoWhite blood cell segmentationby color-space-based k-means clusteringrdquo Sensors vol 14 no 9pp 16128ndash16147 2014

[15] S Mohapatra and D Patra ldquoAutomated leukemia detectionusing hausdorff dimension in blood microscopic imagesrdquo inProceedings of the International Conference on IEEE Roboticsand Communication Technologies (INTERACT rsquo10) pp 64ndash68Chennai India December 2010

[16] N M Salem ldquoSegmentation of white blood cells from micro-scopic images using K-means clusteringrdquo in Proceedings of the31st National Radio Science Conference (NRSC rsquo14) pp 371ndash376Cairo Egypt April 2014

[17] P K Mondal U K Prodhan M S Al Mamun et al ldquoSegmen-tation of white blood cells using fuzzy C means segmentationalgorithmrdquo IOSR Jornal of Computer Engineering vol 1 no 16pp 1ndash5 2014

[18] B C Ko J-W Gim and J-Y Nam ldquoAutomatic white blood cellsegmentation using stepwise merging rules and gradient vectorflow snakerdquoMicron vol 42 no 7 pp 695ndash705 2011

[19] M Hamghalam M Motameni and A E Kelishomi ldquoLeuko-cyte segmentation in giemsa-stained image of peripheral bloodsmears based on active contourrdquo in Proceedings of the Interna-tional Conference on Signal Processing Systems (ICSPS rsquo09) pp103ndash106 IEEE Singapore May 2009

[20] CDi Rubeto ADempster S Khan andB Jarra ldquoSegmentationof blood images using morphological operatorsrdquo in Proceedingsof the 15th International Conference on Pattern Recognition vol3 pp 397ndash400 IEEE Barcelona Spain September 2000

[21] F Scotti ldquoRobust segmentation and measurements techniquesof white cells in bloodmicroscope imagesrdquo in Proceedings of theIEEE Instrumentation and Measurement Technology Conference(IMTC rsquo06) pp 43ndash48 IEEE Sorrento Italy 2006

[22] K A Eldahshan M I Youssef E H Masameer and MA Mustafa ldquoSegmentation framework on digital microscopeimages for acute lymphoblastic leukemia diagnosis based onHSV Color Spacerdquo International Journal of Computer Applica-tions vol 90 no 7 pp 48ndash51 2014

[23] K A ElDahshan M I Youssef E H Masameer and M AHassan ldquoComparison of segmentation framework on digitalmicroscope images for acute lymphoblastic leukemia diagnosisusing RGB and HSV color spacesrdquo Journal of BiomedicalEngineering and Medical Imaging vol 2 no 2 pp 26ndash34 2015

[24] V Singhal and P Singh ldquoCorrelation based feature selection fordiagnosis of acute lymphoblastic leukemiardquo in Proceedings ofthe 3rd ACM International Symposium onWomen in Computingand Informatics (WCI rsquo15) pp 5ndash9 Kochi India August 2015

[25] S I Sahidan M Y Mashor A S W Wahab et al ldquoLocal andglobal contrast stretching for color contrast enhancement onziehl-neelsen tissue section slide imagesrdquo in 4th Kuala LumpurInternational Conference on Biomedical Engineering 2008BIOMED 2008 25ndash28 June 2008 Kuala LumpurMalaysia vol 21of IFMBE Proceedings pp 583ndash586 Springer Berlin Germany2008

[26] R C Gonzalez and R E Woods ldquoColor image processingrdquo inDigital Image Processing pp 416ndash478 Prentice Hall LondonUK 2002

[27] R D Labati V Piuri and F Scotti ldquoAll-IDB the acute lym-phoblastic leukemia image database for image processingrdquo inProceedings of the 18th IEEE International Conference on ImageProcessing (ICIP rsquo11) pp 2045ndash2048 IEEE Brussels BelgiumSeptember 2011

Submit your manuscripts athttpwwwhindawicom

Stem CellsInternational

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014


MEDIATORSINFLAMMATION

of


Behavioural Neurology

EndocrinologyInternational Journal of



Disease Markers


BioMed Research International

OncologyJournal of



Oxidative Medicine and Cellular Longevity


PPAR Research

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Immunology ResearchHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Journal of

ObesityJournal of



Computational and Mathematical Methods in Medicine

OphthalmologyJournal of


Diabetes ResearchJournal of



Research and TreatmentAIDS


Gastroenterology Research and Practice


Parkinsonrsquos Disease

Evidence-Based Complementary and Alternative Medicine

Volume 2014Hindawi Publishing Corporationhttpwwwhindawicom


(a) (b) (c) (d)

Figure 1 Morphological variability associated with the blast cells according to the FAB classification (a) Healthy lymphocytes cell fromnon-ALL patients (bndashd) Lymphoblasts from ALL patients where (b) (c) and (d) are L1 L2 and L3 respectively

classification and counting it has become a very hot topic inclinical diagnosis [4]

Some work on microscopic WBC image segmentationis available in the literature [3ndash25] Methods can be clas-sified into three types threshold-based methods patternrecognition-based methods and deformable model-basedmethods Threshold-based methods include Otsursquos method[4ndash6] the region growing method [7] watershed method[8ndash10] and their combination [11] Cseke [4] presented afast segmentation schemewith automatic thresholding wherethresholds are selected with a simple recursive methodderived from maximizing the interclass variance betweendark gray and bright regions based on the method proposedby OtsuThemethod works well for nucleus and backgroundsegmentation However it cannot separate cytoplasm fromthe red blood cells Wu et al [6] developed an iterativeOtsursquos threshold approach based on circular histogram forthe leukocyte segmentation using HampS components of HSImodel Experimental results show that the method workssuccessfully in the segmentation of WBC nucleus but losesthe cytoplasm informationDorini et al [11] divided theWBCsegmentation process into two steps In the first step theyextracted the cell nucleus using the watershed transformby Image Forest Transform (IFT) Then they segmentedthe WBC cytoplasm using basic operations such as thresh-olding and morphological opening via the size distributioninformation of the RBC but the method tends to produceoversegmentation in presence of noise

As leukocytes in microscopic images can be treated asobjects pattern recognitionmethods are also used to performthe segmentation which can be categorized as supervised orunsupervised [12] Supervised methods classify the objectsusing learning-based approaches such as Support VectorMachine (SVM) [13] and Artificial Neural Network (ANN)while unsupervised methods also known as clustering meth-ods mainly including 119896-means clustering [14ndash16] fuzzy 119862-means [17] and expectation-maximization extract the objectsfrom the data itself In [13] Guo et al proposed multispectralimaging techniques with a spectral calibration method toacquire device-independent images and then applied SVMdirectly to the spectrum of each pixel to segment thewhole microscopic image into four types of regions nucleus

cytoplasm erythrocytes and background Segmentationresults are satisfactory but the implementation speed needsto be boosted

Deformable model-based methods can be classified intoparametric models and geometric models based on contourrepresentation Besides level set method [9] the active con-tour model also known as a snake is the most common [1819] In [18] Ko et al introduced a new WBC image segmen-tation method using stepwise merging rules based on mean-shift clustering and boundary removal rules with a gradientvector flow (GVF) snake Removal rules are used to removethe boundary and noise edges while a GVF snake is forced todeform to the cytoplasm boundary edges Due to a weak dif-ference between the cytoplasm and the background or con-tact with RBCs some experimental results were slightly over-segmented Other segmentation methods for WBCs asidefrom the above three categories aremorphological operations[3 11 20] hybrid methods [9 10 19 21] and so on Hybridmethods combine two or more methods mentioned aboveto achieve better results such as combination of a Lab colorspace based segmentationmethod and a gray level threshold-ing method [21] and combination of an Otsu method and anactive contour method [19]

Generally the ultimate goal of WBC segmentation is toextract whole WBC from a complicated background andsegment every WBC into morphological components suchas nucleus and cytoplasm Among the methods mentionedabove RGB color space based threshold approaches aremost widely used due to their high efficiency and reliabilityHowever cytoplasm has a big variance of color and its coloris quite similar to image backgroundThus in many previousworks gray image based threshold methods are only utilizedto segment nucleus (Figures 2(a) and 2(d)) As to cyto-plasm extraction (Figure 2(c)) other auxiliary segmentationschemes are needed [19] What is more aside from single-threshold-based segmentation methods nucleus and cyto-plasm are respectively segmented with different methods inmany other segmentation schemes too [3 8 9 11 18] To facil-itate cytoplasm segmentation via threshold-based methodssome researchers have turned their attention to segmenting intransformed color space such as HSV [22 23] HSI [24] andLab [21] As an example Eldahshan et al [22] proposed to


(a) (b) (c) (d) (e)


Input

Grayscaling




Binarization


Medianfiltering

MCR extraction

Contrast stretching

Thresh1Thresh2I

G

HS

Postprocessing

RGB


G998400

H998400









(a) (b) (c)










119868119901(119909 119910) =

255 ∙ [119868119900(119909 119910) minusmin]

(max minusmin) (1)










3000

2000

1000

0

0 100 200

(a)

Thresh1

3000

2000

1000

0

0 100 200

(b) (c)









H channel

L channel

HSV

Lab

S channel

a channel

V channel

b channel




DSC (119860 119861) = 2 lowast (119860 cap 119861)(119860 + 119861) (2)








Thresh2

1000

500

0

0 05 1

(a) (b)
















Yes

No

Yes

No

Stop

YesYes

No

Yes

NoNo

Start

[a b]epsilon gt 0


a = x1b = x2

x2 = x


f(x1)

1

f(x2) = f(x1)

a = x1

x1 = x


f(x2)

2

f(x1) = f(x2)

b = x2

x = (a + b)2

f(x1) == f(x2)

f(x1) gt f(x2)



(a) (b) (c)







(a) (b)






Note











(a) (b) (c) (d) (e)



4 Conclusion




(a) (b) (c) (d)


1

09

08

07

06

05

04

DSC

DSC comparison

0 5 10 15 20 25 30


Method 1Method 2

Method 3


Competing Interests


Acknowledgments




References

































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















(a) (b) (c) (d) (e)


Input

Grayscaling




Binarization


Medianfiltering

MCR extraction

Contrast stretching

Thresh1Thresh2I

G

HS

Postprocessing

RGB


G998400

H998400









(a) (b) (c)










119868119901(119909 119910) =

255 ∙ [119868119900(119909 119910) minusmin]

(max minusmin) (1)










3000

2000

1000

0

0 100 200

(a)

Thresh1

3000

2000

1000

0

0 100 200

(b) (c)









H channel

L channel

HSV

Lab

S channel

a channel

V channel

b channel




DSC (119860 119861) = 2 lowast (119860 cap 119861)(119860 + 119861) (2)








Thresh2

1000

500

0

0 05 1

(a) (b)
















Yes

No

Yes

No

Stop

YesYes

No

Yes

NoNo

Start

[a b]epsilon gt 0


a = x1b = x2

x2 = x


f(x1)

1

f(x2) = f(x1)

a = x1

x1 = x


f(x2)

2

f(x1) = f(x2)

b = x2

x = (a + b)2

f(x1) == f(x2)

f(x1) gt f(x2)



(a) (b) (c)







(a) (b)






Note











(a) (b) (c) (d) (e)



4 Conclusion




(a) (b) (c) (d)


1

09

08

07

06

05

04

DSC

DSC comparison

0 5 10 15 20 25 30


Method 1Method 2

Method 3


Competing Interests


Acknowledgments




References

































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















(a) (b) (c)










119868119901(119909 119910) =

255 ∙ [119868119900(119909 119910) minusmin]

(max minusmin) (1)










3000

2000

1000

0

0 100 200

(a)

Thresh1

3000

2000

1000

0

0 100 200

(b) (c)









H channel

L channel

HSV

Lab

S channel

a channel

V channel

b channel




DSC (119860 119861) = 2 lowast (119860 cap 119861)(119860 + 119861) (2)








Thresh2

1000

500

0

0 05 1

(a) (b)
















Yes

No

Yes

No

Stop

YesYes

No

Yes

NoNo

Start

[a b]epsilon gt 0


a = x1b = x2

x2 = x


f(x1)

1

f(x2) = f(x1)

a = x1

x1 = x


f(x2)

2

f(x1) = f(x2)

b = x2

x = (a + b)2

f(x1) == f(x2)

f(x1) gt f(x2)



(a) (b) (c)







(a) (b)






Note











(a) (b) (c) (d) (e)



4 Conclusion




(a) (b) (c) (d)


1

09

08

07

06

05

04

DSC

DSC comparison

0 5 10 15 20 25 30


Method 1Method 2

Method 3


Competing Interests


Acknowledgments




References

































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















3000

2000

1000

0

0 100 200

(a)

Thresh1

3000

2000

1000

0

0 100 200

(b) (c)









H channel

L channel

HSV

Lab

S channel

a channel

V channel

b channel




DSC (119860 119861) = 2 lowast (119860 cap 119861)(119860 + 119861) (2)








Thresh2

1000

500

0

0 05 1

(a) (b)
















Yes

No

Yes

No

Stop

YesYes

No

Yes

NoNo

Start

[a b]epsilon gt 0


a = x1b = x2

x2 = x


f(x1)

1

f(x2) = f(x1)

a = x1

x1 = x


f(x2)

2

f(x1) = f(x2)

b = x2

x = (a + b)2

f(x1) == f(x2)

f(x1) gt f(x2)



(a) (b) (c)







(a) (b)






Note











(a) (b) (c) (d) (e)



4 Conclusion




(a) (b) (c) (d)


1

09

08

07

06

05

04

DSC

DSC comparison

0 5 10 15 20 25 30


Method 1Method 2

Method 3


Competing Interests


Acknowledgments




References

































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of


















H channel

L channel

HSV

Lab

S channel

a channel

V channel

b channel




DSC (119860 119861) = 2 lowast (119860 cap 119861)(119860 + 119861) (2)








Thresh2

1000

500

0

0 05 1

(a) (b)
















Yes

No

Yes

No

Stop

YesYes

No

Yes

NoNo

Start

[a b]epsilon gt 0


a = x1b = x2

x2 = x


f(x1)

1

f(x2) = f(x1)

a = x1

x1 = x


f(x2)

2

f(x1) = f(x2)

b = x2

x = (a + b)2

f(x1) == f(x2)

f(x1) gt f(x2)



(a) (b) (c)







(a) (b)






Note











(a) (b) (c) (d) (e)



4 Conclusion




(a) (b) (c) (d)


1

09

08

07

06

05

04

DSC

DSC comparison

0 5 10 15 20 25 30


Method 1Method 2

Method 3


Competing Interests


Acknowledgments




References

































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















Thresh2

1000

500

0

0 05 1

(a) (b)
















Yes

No

Yes

No

Stop

YesYes

No

Yes

NoNo

Start

[a b]epsilon gt 0


a = x1b = x2

x2 = x


f(x1)

1

f(x2) = f(x1)

a = x1

x1 = x


f(x2)

2

f(x1) = f(x2)

b = x2

x = (a + b)2

f(x1) == f(x2)

f(x1) gt f(x2)



(a) (b) (c)







(a) (b)






Note











(a) (b) (c) (d) (e)



4 Conclusion




(a) (b) (c) (d)


1

09

08

07

06

05

04

DSC

DSC comparison

0 5 10 15 20 25 30


Method 1Method 2

Method 3


Competing Interests


Acknowledgments




References

































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















Yes

No

Yes

No

Stop

YesYes

No

Yes

NoNo

Start

[a b]epsilon gt 0


a = x1b = x2

x2 = x


f(x1)

1

f(x2) = f(x1)

a = x1

x1 = x


f(x2)

2

f(x1) = f(x2)

b = x2

x = (a + b)2

f(x1) == f(x2)

f(x1) gt f(x2)



(a) (b) (c)







(a) (b)






Note











(a) (b) (c) (d) (e)



4 Conclusion




(a) (b) (c) (d)


1

09

08

07

06

05

04

DSC

DSC comparison

0 5 10 15 20 25 30


Method 1Method 2

Method 3


Competing Interests


Acknowledgments




References

































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















(a) (b)






Note











(a) (b) (c) (d) (e)



4 Conclusion




(a) (b) (c) (d)


1

09

08

07

06

05

04

DSC

DSC comparison

0 5 10 15 20 25 30


Method 1Method 2

Method 3


Competing Interests


Acknowledgments




References

































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















(a) (b) (c) (d) (e)



4 Conclusion




(a) (b) (c) (d)


1

09

08

07

06

05

04

DSC

DSC comparison

0 5 10 15 20 25 30


Method 1Method 2

Method 3


Competing Interests


Acknowledgments




References

































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















(a) (b) (c) (d)


1

09

08

07

06

05

04

DSC

DSC comparison

0 5 10 15 20 25 30


Method 1Method 2

Method 3


Competing Interests


Acknowledgments




References

































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of


















References

































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of





















of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of
















Research Article Segmentation of White Blood Cell from...

Documents

Transcript of Research Article Segmentation of White Blood Cell from...