research article automatic detection and quantification of...

Research ArticleAutomatic Detection and Quantification of WBCs and RBCsUsing Iterative Structured Circle Detection Algorithm

Yazan M Alomari1 Siti Norul Huda Sheikh Abdullah1

Raja Zaharatul Azma2 and Khairuddin Omar1

1 Pattern Recognition Research Group Center for Artificial Intelligence Technology Faculty of Information Science and TechnologyUniversiti Kebangsaan Malaysia 43600 Bangi Malaysia

2 Department of Pathology UKMMedical Center Universiti Kebangsaan Malaysia Cheras 56000 Kuala Lumpur Malaysia

Correspondence should be addressed to Yazan M Alomari yazanitgmailcom

Received 30 December 2013 Revised 21 February 2014 Accepted 8 March 2014 Published 2 April 2014

Academic Editor Shengyong Chen

Copyright copy 2014 Yazan M Alomari et al This is an open access article distributed under the Creative Commons AttributionLicense which permits unrestricted use distribution and reproduction in any medium provided the original work is properlycited

Segmentation and counting of blood cells are considered as an important step that helps to extract features to diagnose somespecific diseases like malaria or leukemia The manual counting of white blood cells (WBCs) and red blood cells (RBCs) inmicroscopic images is an extremely tedious time consuming and inaccurate process Automatic analysis will allow hematologistexperts to perform faster and more accurately The proposed method uses an iterative structured circle detection algorithm for thesegmentation and counting of WBCs and RBCs The separation of WBCs from RBCs was achieved by thresholding and specificpreprocessing steps were developed for each cell type Countingwas performed for each image using the proposedmethod based onmodified circle detection which automatically counted the cells Several modifications were made to the basic (RCD) algorithm tosolve the initialization problem detecting irregular circles (cells) selecting the optimal circle from the candidate circles determiningthe number of iterations in a fully dynamic way to enhance algorithm detection and running time The validation method used todetermine segmentation accuracy was a quantitative analysis that included Precision Recall and F-measurement testsThe averageaccuracy of the proposed method was 953 for RBCs and 984 for WBCs

1 Introduction

The analysis of microscopy images is extremely importantin both the medical and the computer science fields Manyresearch problems are related to the analysis of microscopyimages such as complete blood count (CBC) tests [1] andthe analysis of blood smears which is considered the firststep in detecting and diagnosing malaria leukemia andanemia Additionally during a complete physical exam aseries of tests are performed One of these tests is the CBCwhich is used to evaluate the composition and concentrationof all cellular blood components The CBC determines redblood cell (RBC) counts white blood cell (WBC) countsplatelet counts hemoglobin (HB) measurements and meanred blood cell volumes [2]

CBC tests and the analysis of blood smear images helpto evaluate diagnose andmonitor various health conditions

such as anemia leukemia infections and allergic conditions[3] For blood disorders such as anemia which is based onHB level the production and destruction of red blood cellsare evaluated In red blood cell disorder such as anemiaother red cell indices such as (mean cell volume) MCV meancell hemoglobin (MCH) mean corpuscular hemoglobinconcentration (MCHC) RBC and red blood cell distributionwidth (RDW or RCDW) are evaluated to narrow down onthe causes of anemia If the red cell indices are suggestedof iron deficiency anemia (IDA) further tests to confirmthe IDA will be done In normal blood red blood cell(RBC) counts range from 42 to 59 million cells per squarecentimeter High RBC counts can be indicative of seriousmedical conditions such as heart lung or kidney diseasePrimary or secondary polycythemia in polycythemia HBis also raised a bone marrow disorder also causes highRBC counts [2] Normal WBC counts range from 4500 to

Hindawi Publishing CorporationComputational and Mathematical Methods in MedicineVolume 2014 Article ID 979302 17 pageshttpdxdoiorg1011552014979302

2 Computational and Mathematical Methods in Medicine

10000 WBCs per microliter of blood [4] High WBC counts(above 30000 cells per microliter) indicate an infectionsystemic illness inflammation allergy leukemia or burn-induced tissue injury If leukemia is suspected analysis ofblood smear is done to look for morphology of the leukemiccells and followed by bone marrow examinations [2ndash4] Forplatelets which are small blood cell fragments that assist inblood clotting normal counts range from 150000 to 450000platelets per microliter In patients with low platelet countsuch as in patients with dengue infection their platelet countis monitored closely and the value is within critical levelthe patient might need platelet transfusion [2] Generallyany abnormal blood smear reading indicates an infection ordisease

Malaria and Babesia are parasites that infect RBCs theanalysis of thin blood smear remains the gold standard fordiagnosis in such disease [5] Microscopy images are alsostill used for early diagnosis analysis and count of someblooddisorder such as sickle cell anemia and leukemia beforeconfirming it with other laboratory tests However manualor visual quantification of parasitemia in thin blood filmsand WBCs in leukemia is an extremely tedious subjectiveand time-consuming task with a high probability of count-ing error [6 7] An accurate segmentation and countingmechanism that gathers information about the distribution ofmicroscopic particles may help diagnose abnormalities dur-ing clinical analysis Our objective in this paper is to developand validate an algorithm that segments and automaticallycounts red and white blood cells in microscopy imagesThe ground truth of the images was determined by expertsFor the evaluation quantitative analyses were performedon the segmentation results based on the ground truthand the 119865-measure method was used to confirm accuracyIn the following sections of this paper we will summarizerelated work on the segmentation and counting of RBCs andWBCs (Section 2) present themethodology used (Section 3)discuss the results and experiments (Section 4) and reviewthe conclusions

2 Related Work

Many researchers have investigated blood cell segmentationand counting Some researchers [5 8ndash10] usedmorphologicaloperations and thresholding to do the segmentation andcounting Berge et al [5] presented an approach basedon a morphological method and iterative threshold tech-niques Segmentation was performed on red blood cellswhich included clumped cells and boundary curvatureswere used to construct a Delaunay triangulation They usedreal microscopy images prepared in the laboratory andthe ground truth was determined by a laboratory expertA 28 difference was calculated between the manual andautomatic counting of red blood cellsTheirmethod tolerateda degree of overlapping but in cases with a high degreeof overlapping cells the cells were unable to be detectedAdditionally the iterative threshold method was unable todetect faint red cells Damahe et al [8] used the S andV imagecomponents of a HSV color model with Zackrsquos thresholdingtechnique for cell segmentationThresholding combinedwith

a sequential edge-linking algorithm was used to increasesegmentation accuracy The experimental dataset and bloodcell images were collected from the Dhruv Pathology Laband the CDC-DPDx respectively and the ground truth wasdetermined by experts The dataset size was limited SeveralRBCs that were detected possessed holes Additional prepro-cessing steps should have been implemented to increase theaccuracy Panchbhai and Vishal [9] proposed an automatedanalysis that counted red blood cells and detected malariaparasites in thin blood smear samples The green color layerwas processed to count all of the RBCs segmentation wasperformed on the infected RBCs using Otsu thresholding Ahistogramwas used to determine the optimal threshold CDCdatasets were used for the experimental part and the groundtruth was determined by a pathology expert who comparedtheir results with the manual counts However becausethe detection algorithm used morphology and thresholdingtheir method was unable to detect clumped and overlappingcells Khan et al [10] proposed a method to count WBCsRBCs and platelets Several preprocessing steps were per-formed before converting the image to binary Segmentationand cell counting were performed based on the optimalthreshold value which was determined from a histogramThey achieved 95 accuracy with their proposed methodcompared to manual counting and a hematology analyzerHowever this method is unable to detect overlapping cellsWhen using iterative thresholds the probability of losinguseful information from the image is high this decreases theaccuracy of segmentation

Nguyen et al [11] used distance transform to solve theoverlapping cells problem they proposed a method thatconcentrated on clumped cells First they assigned centralpoints based on a distance transform The optimal centerpoints were selected by checking the degree of boundarycovering the center point and the average size of a cell wasestimated by the extraction of a single cellThen an algorithmwas developed that used a single cell mask to split the cellsTheir dataset ground truth was labeled by experts and theaccuracy of the proposed method using 119865-measurementswas 935 and 82 Clumped cells were tolerated but thecells had to be regularly shaped and focused at a highmagnification Not all blood cells have regular cell shapesand this is especially true if the blood cells are diseasedTherefore cell detection methods should be able to detectirregular cells Additionally because of noise the performanceof their method was not good

Rhodes and Bai [12] presented a circle detection methodusing specific properties of Gabor wavelet filter to detect theimage features such as circularity It is able to extract radiuswraps around the origin and the plane wave radiates fromthe center of the filter They test their proposed method onsynthetic images and real microscopic images which allow acertain degree of overlapping cellsThe results were 913 and87 of the cells detected in the two microscopic test images

Since blood cells are approximately circular shape circledetection algorithms can still handle the challenge of bloodcells detection Hough Transform is considered as one of themost known algorithms for line and circle detection It wasdeveloped by Richard Duda and Peter Hart in 1972 [13] they

Computational and Mathematical Methods in Medicine 3

called it as a generalized Hough Transform after the related1962 patent of Paul Hough [13] Hough Transforms mapsevery edge pixel into parameter space and use conventionalHT to detect lines circles or any other parametric shapeIn this paper we concentrate on circle detection algorithmsThere are many techniques used for circle detection ManyHT-based algorithms that detect circles were developed usingdifferent methods Yip et al [14] reduced the accumulatorarray to enhance computation time and improve memoryconsumption Other methods use pixel gradient information[15 16] to reduce computation time and the accumulatorarray Ho and Chen used the geometric properties of circlesto improve performance [17] Xu et al [18 19] developedthe randomized Hough Transform (RHT) which randomlyselects three noncollinear edge pixels (119886 119887 119903) maps theminto parameter space and requires less computation time andmemory storage compared with Standard HT methods Asimple voting strategy in the accumulator is used to collectevidence and determine the existence of a circle Chung andHuang [13] presented a method that can substitute variousshape detection algorithmsThismethod enhanced the speedof the original algorithmThey applied their method in RHTand randomized circle detection (RCD) algorithms to detectlines circles and ellipses Their method presented goodresults when compared with original methods

Chiu et al [20] presented a fast randomized HoughTransform method for detecting circles to improve RHTwhich is less efficient in complex images due to its probabilityusage problem They pick one random edge pixel from theimage and it is considered as a seed point Then a checkingmethod was developed to observe if this selected seed pointis lying on a true circle or otherwise The checking criterionis based on finding two other points whose distances arethe same from the selected seed point by using a 119882 times 119882

window centered by the selected seed point This methodenhanced the probability to find relevant points on a truecircle They have proven that using one random selectedpointrsquos probability is sufficient in comparison to three randomselected points [13ndash19]

Chen and Chung [21] presented a circular algorithmcalled RCD They claimed that RCD outperformed othermost efficient Hough Transform based algorithms [13ndash19]Regardless of accumulator usage as in [13ndash19] RCD works byrandomly selecting four edge pixels from the whole imageThen these pixels are examined if they are noncollinear andit will proceed to form a candidate circle RCD determinesthat circle is a possible circle based on distance criteria Afterfinding its center and radius it checks the number of pixelslying on the boundary of this possible circle This checkingcriterion is performed by calculating the distance between alledge pixels in the image and the boundary of this possiblecircle If this distance is lesser than a fixed threshold valuethen it will be considered as a boundary of this possiblecircle Finally another fixed threshold value has also beenused to decide a true circle or otherwise based on numberof edge pixels lying on a possible circlersquos boundary Otherrelated issues on RCD are less efficient when dealing withhuge image size consisting of a high number of edge pixelsRCD also requires four selected edge pixels randomly from

the whole image which causes low probability forming a truecircle Furthermore RCD has a drawback in terms of its fixednumber of iterations in which it is highly correlated to theimage texture In addition RCD acquires many parametersand threshold values to be predefined and it ignores irregularcircles

Since the Hough Transform presented a good perfor-mance in different fields many researchers [22ndash25] usedHough transformmethod for detecting circle when perform-ing RBCrsquos and WBCrsquos calculation task in the microscopicimages Mahmood et al [22] applied Hough transformmethod for counting the RBCrsquos and WBCrsquos for the micro-scopic images In the first step they converted the sourceimage to 119871 lowast 119886 lowast 119887 color space model and performedcolor segmentation process of red and white cells based onfeature lightness over the 119871 channel ranging between 130ndash150and 80ndash100 respectively Then the second step begins withapplying some morphological operators which are followedby applyingCanny operator for edge detection process Lastlythey performed cell detection and counting using CircularHough Transform The dataset was composed of 108 imagesfrom the ldquoAcute Lymphoblastic Leukemia Image Databasefor Image Processingrdquo database that was established andmaintained by Labati et al [26] Depending on the processingand type of cells being analyzed they achieved an accuracyranging from 64 to 87 The Hough Transform consumedmemory storage and required a long computation time todetermine a large range of accepted radii Additionally theability to tolerate a high degree of overlapping and irregularcells was limited therefore accuracy was not high

Mahmood and Mansor [23] examined 10 image samplesof normal blood cells image transformed to the HSV colorspace and then Saturation or ldquoSrdquo channel was selectedto proceed with image analysis Morphological operatorsand thresholding method were used over S channel forcell segmentation They used Circular Hough Transform toinvestigate the circularity feature of the red blood cells inorder to perform detection and counting Their proposedmethod achieved approximately 96 of accuracy rate incomparison to manual counting Extracting and countingnormal cells are simple tasks if the detected cells are normalcells and consist of small number of overlapping cells withregular shape For this kind of easy case the idea of usingHough Transform can be very helpful because it can producea good performance

With similar inspiration to [23] Venkatalakshmi andThilagavathi [24] have also applied circularHoughTransformmethod to count the RBCs from microscopic images afterperforming preprocessing steps such as HSV transformationS channel extraction histogram thresholding morphologicaloperations XOR logical operation and Canny edge detec-tion However again this proposed idea is less tolerant toany overlapping cells or irregular cellsrsquo shape Later Maitraet al [25] presented a composition method to extract redcell from five microscopic images these steps include spatialsmoothing and filtering adaptive histogram equalizationand edge detection Similar to [23 24] they used basicCircular Hough Transform to detect the red cells based onprior information such as size and shape features Those


Counting using iterative structured circle algorithm

Show results with automatic

counting

thresholding

Read image

thresholding

preprocessing preprocessing

image

Extract WBCs by Extract RBCs by

WBCs RBCs

Subtract WBCs

Figure 1 General methodology for the proposed method

methods [23ndash25] employ classic Hough Transform methodto detect the blood cells which inherits some drawbacksfor example required more computation high memoryconsumption and less ability in detecting overlapping cellsor irregular cells

Finally Cuevas et al [27] presented a method whichis not Hough Transform based to detect white bloodcells They used a combination of circle detection withelectromagnetism-like optimization from the edge mapimage This method tolerated noise In blood smear imagesthe number of white cells was small compared with thenumber of red cells therefore small degree of overlappingwhite cells can be detected with their method Howeverthey did not test the method on clumping cells Theirresults were compared with othermethods and their methoddemonstrated good accuracy

3 Methodology

The proposed method was developed to analyze microscopicimages of blood smears by segmenting and counting bothWBCs and RBCsThe segmentation is based on thresholdingandmorphological operations and then counting is based onthe circularity feature of the blood cells extracted using aniterative structured circle detection algorithm

A new technique for binary images based on the funda-mentals of RCD has been proposed and used for countingRBCs and WBCs Therefore the original image is separatedinto two images the first image contains RBCs only and thesecond image contains WBCs this step has been done usingthresholdingWe study the histogram for 20 sample gray scaleimages andwe find out the best thresholding values to extractWBCs and RBCs values were 64 and 140 respectively Aftercells separation each image is preprocessed using morphol-ogy operators to obtain the edge image using Canny operatorThen an iterative structured circle detection algorithm is

Convert the image to gray

Convert image to binary using 64

thresholding values

Complement the image

Removing extra not connected pixels

Edge image using canny

Erode the image

Fill holes

Figure 2 Preprocessing steps for WBCs

used to count cells in each image Figure 1 shows the proposedmethod for WBC and RBC segmentation

31 Preprocessing The proposed method for cell segmenta-tion works with edge images Microscopy images of bloodsmears are colored images and several steps are required toprepare the image before extraction of the edge image Inour proposed method the cells were separated by type anddistinct preprocessing steps were developed for WBCs andRBCs separately

Preprocessing for WBCs At this stage white blood cellsare extracted as a separate image and the red blood cellshave varying intensities therefore it is preferable to developseparate preprocessing steps for each cell type

Figure 2 shows the overall preprocessing steps forWBCsTo remove RBCs from the image the RGB image wasconverted into grayscale image by eliminating the Hue andsaturation information while retaining its luminance asshown in Figure 3(b) and then converted the image to binaryusing thresholding using a threshold value (64) visualizethe WBCs as shown in Figure 3(c) Some undesired holesappeared in the cells and the morphology operator fillholes were used to remove themThe complementary imagesbefore the holes were filled as shown in Figure 3(d)

This image eroded to reduce the number of overlappingcells as shown in Figure 4(a) Because the boundary of thecells is required the holes were filled to improve the edgedetection Figure 4(b) shows the image after the holes filledThe Canny operator was used to visualize the cell edges asshown in Figure 4(c) After edge detection some undesiredpixels appeared that affect the segmentation process Thesepixels were removed using an open morphology operator asshown in Figure 4(d)

311 Preprocessing for RBCs In this stage preprocessingwas performed on the red blood cells after removing thewhite blood cells from the image Figure 5 shows the overallpreprocessing steps for the RBCs First the original image isconverted into grayscale image by eliminating the Hue andsaturation information while retaining its luminance


(a) (b)

(c) (d)

Figure 3 (a) Original blood image (b) gray image (c) binary image using thresholding and (d) complement image

Then image was converted to binary using thresholdingvalue of (140) to visualize all of the red and white blood cellsin the image as shown in Figure 6(a) To remove the whiteblood cells from the image the complementary white cellsimage was taken and subtracted from the first image to obtainonly the red cells as shown in Figure 6(b)

When the image was converted to binary and the whiteblood cells were removed undesired holes were createdThese holes disturb the solidness of the object Thereforemorphological operators are used to fill the holes as shownin Figure 6(c)

A morphological step using an erosion operator is usedto reduce the overlap between cells as shown in Figure 7(a)Canny edge detection method is used to obtain the edgeimage as shown in Figure 7(b) Some undesired pixelsappeared after edge detection These pixels represent eitherplatelets or noise and it will subsequently affect the seg-mentation process Therefore these pixels are removed usinga morphology operator on the binary image as shown inFigure 7(c) After these preprocessing steps the image wasprepared for input to our proposed circle detection algorithm

32The ProposedMethod for Counting Thebasic idea for theproposed method was derived from the RCD [21] algorithm

RCD algorithm ignores accumulator capability that has beenintroduced by Hough Transform method Several modifica-tions were made to the basic RCD algorithm to solve theinitialization problemwhen using big images with high num-ber of pixels The methods modified to detecting irregularcircles selecting the optimal circle from the candidate circlesdetermining the number iterations in a fully dynamic wayto enhance the algorithm detection and running time andimproving the detection of overlapping cells The workflowof the proposed method is shown in Figure 8

Initialization Problem The proposed method partitions theedge image based on 8-neighbor connected components inorder to overcome the initialization problem caused in basicRCD [21] We eventually divide the whole image into smallpartitions and we consider each partition as an input imagebefore entering into our iterative structured circle detectionalgorithm that employs local randomization step

Regarding to this initialization problem as we high-lighted earlier that selecting four pixels globally (from thewhole image) can reduce the probability of finding truecircle time consuming and needs of a high number ofiterations to find all true circles This can be easily illustratedas in Figure 9(a) if cell A is to be detected and the fourselected pixels are chosen in Red Cross mark Therefore


(a) (b)

(c) (d)

Figure 4 (a) Eroded image (b) image after filling holes (c) edge detection using a Canny at highmagnification and (d) image after removingthe noise at high magnification


Convert image to binary using thresholding (140) value

Remove white cells

Remove extra not connected pixels

Fill holes

Detect edge using cannyErode image

Figure 5 Preprocessing steps for RBCs

we simplify this process by introducing four pixel selectionslocally (within each partition image) Assume that cell A isto be detected in the partition image and then these fourlocal pixels are chosen randomly from the partition imageas depicted in Figure 9(b) This local randomization processrepeats until all partition images are visited

In each partition image our proposed method randomlyselects up to four edge pixels and checks the existence of

a circle using (1) which is based on three noncollinear pixels(v1 v2 and v3)

(119909119895 minus 119909119894) (119910119896 minus 119910119894) minus (119909119896 minus 119909119894) (119910119895 minus 119910119894) = 0 (1)

If the result is zero the three-edge pixels are collinear andcannot form a circle they are returned to the edges arrayfor the chosen partition and four new subsequent pixels areselected from the same partition

If the pixels are not collinear they are able to form acircle Four candidate circles C123 C124 C134 and C234 canbe formed using the four edge pixels (v1 v2 v3 and v4) asshown in Figure 10 Each circle is formed from three edgepixels and the center (119886 119887) and radii (119903)were calculated using(2) (3) and (4) respectively Consider the following

119886119894119895119896 =

1003816100381610038161003816100381610038161003816

1199092119895+1199102119895minus(1199092

119894+1199102119894 ) 2(119910119895minus119910119894)

1199092119896+1199102119896minus(1199092

119896+1199102119894 ) 2(119910119896minus119910119894)

1003816100381610038161003816100381610038161003816

4 ((119909119895 minus 119909119894) (119910119896 minus 119910119894) minus (119909119896 minus 119909119894) (119910119895 minus 119910119894))

(2)

119887119894119895119896 =

1003816100381610038161003816100381610038161003816

2(119909119895minus119909119894) 1199092

119895+1199102119895minus(1199092

119894+1199102119894 )

2(119909119896minus119909119894) 1199092

119896+1199102119896minus(1199092

119894+1199102119894 )

1003816100381610038161003816100381610038161003816


(3)

119903119894119895119896 =radic(119909119894 minus 119886119894119895119896)

2+ (119910119894 minus 119887119894119895119896)

2 (4)


(a) (b)

(c)

Figure 6 (a) Binary image using thresholding (b) RBCs after removing theWBCrsquos by subtracting the two images and (c) image after fillingholes

In basic RCD algorithm circle can be formed with threeprior pixels the fourth pixel is used to obtain four candidatecircles at a time for each iteration which is better thanforming one candidate circle for each iteration However inour proposedmethod one of the candidate circles C123 C124C134 and C234 is selected to be possible circle This based onthe candidate with the highest probability of being a possiblecircle This can be achieved by checking the number of edgepixels located on the boundary of each candidate circle Weassumed that candidate circle with the maximum number ofedge pixels on its boundary is considered as a possible circleThe following calculates the distance between each edge pixel(119911) in the partition and the boundary of the candidate circle

119889119911rarr 119894119895119896 =

1003816100381610038161003816100381610038161003816100381610038161003816

radic(119909119911 minus 119886119894119895119896)2+ (119910119911 minus 119887119894119895119896)

2minus 119903119894119895119896

1003816100381610038161003816100381610038161003816100381610038161003816

(5)

After checking all of the edge pixels in the partition andselecting the possible circle the number of edge pixels locatedon its boundary is determined If this number is greater thanthe value obtained from (2times120587times119903times119879119903) then the possible circleis considered as a real circle Where 119879119903 = 07 for WBCs and119879119903 = 05 for RBCs (2 times 120587 times 119903) is the circle perimeter and thisvalue should be equal to the estimatednumber of pixels on thecircle boundary 119879119903 is the accepted ratio of the total numberof pixels on the possible circle boundary that determinesthat the circle is a real circle Therefore if 70 or 50 ofedge pixels are located on the boundary of possible circle for

WBCs or RBCs respectively the possible circle is considereda real circle Then the edge pixels are removed from the setof edges in that partition and the process starts again tofind other circles within the same partition If there are notenough remaining edge pixels or less than the threshold valueof 119879min = (2 times 120587 times 119903min times 119879119903) where 119903 is the minimumradius possible for red or white blood cells and 119879119903 is oneof the values mentioned above then the next partition isselected and the process begins again Unlike basic RCD ourproposed method applies dynamic 119879min thresholding valuewhich depends on minimum acceptable radius (119903min) valueand 119879119903 value pertaining to the type of cells to be detected

If the number of edge pixels on the boundary of thepossible circle is less than the threshold value 119879119903 fournew edge pixels are randomly selected The 119879min value isdetermined based on theminimumnumber of pixels that canform a circle boundary with a radius 119903 and is estimated to be(2 times 120587 times 119903 times 119879119903) with a width of one pixel

By adding the 8-neighbor connected component stepthe algorithm is guaranteed to randomly select edge pixelsfrom every part of the image check the circle possibilitiesfor all image edge pixels and locate more true circles whichimproves the performance of the method especially for largeimages with a high number of pixels

Another drawback of basic RCD algorithm it choosesfour random edge pixels from the entire or global imagethat may cause chosen pixels from different actual shapes asdepicted in Figure 11(a) These edge pixels can form a circle


(a) (b)

(c)

Figure 7 (a) The eroded image (b) image after edge detection using thresholding and (c) after removing extra unconnected pixels

based on three threshold values (119879119886 119879119889 and 119879119903) Figure 11(b)shows some basic RCD threshold values 119879119886 represents thethreshold value of distance between randomly selected edgepixels 119879119889 represents the threshold value of distance betweenthe fourth edge pixel and the boundary of the possible circleand 119879119903 threshold value of accepted ratio of the number ofpixels on the possible circle boundary as (6) Based on thesethreshold values basic RCD decided that this circle is a truecircle even if it is untrue Figure 11(a) demonstrates how thebasic RCD algorithm selects four pixels randomly from thenumbered shapes (1 2 3 and 5) and forms a possible circlebased on 119879119886 119879119889 and 119879119903 thresholding values

119862119894119895119896 =

accept as a possible circleif (119889(119894119895) cap 119889(119894119896) cap 119889(119895119896) gt 119879119886)cap (119889119911rarr119862119894119895119896 lt 119879119889)

accept as a true circleelseif Number of edge pixels on119862119894119895119896 boundary gt 119879119903

choose another random pixelselse

(6)

Taking into account the above problems we introducepartition based on 8-nieghbor connected components in ourproposed method We restrict up to four random selectedpixels from a particular partition Next we also check thenumber of edge pixels located on the possible circle boundarywithin that particular partition Having known these pieces

of information it guarantees (1) better improvement perfor-mance and (2) higher probability of finding all existed truecircles in that source image and reduces true-negative circlesIrregular Circle Detection Normal red and white bloodcells are approximately circular however not all blood cellsare regular circles Therefore we detected irregular circleswhenever a possible circle was found We defined a virtualcircle with the same center and different radii as shown inFigure 12(a) When the algorithm locates a possible circle wesuperimpose these two circles on the grid and check all of theedge pixels in the segment and detect all edges lying betweenthe two circles Figure 12(b) shows the irregular circle pixelswhich are located between the two virtual circles in gray colorare detected

The red pixels are not included in the boundary ofcandidate circle By partitioning the image based on con-nected components we increase the distance between the twocirclesThis ensures that the detection edges on the boundaryof the possible circle will be correct without using edges fromother shapes as observed in Figure 11(a)Our Proposed Iterative StructuredThenumber of iterations isextremely important to find all circles in the partition imageTherefore determining the number of iterations is not aneasy task for several reasons Firstly the number of iterationspredominately depends on two factors the number of edgepixels in the image and the distance between the edge pixelsthat are required to form a circle in each iteration As aresult this may cause huge numbers of iteration to find all


Picks four random edge pixels

Four candidate circles

Draw that circle and remove its edge pixels

from edges set

8-neighbor connected componentsfor binary image

Loop

number of partitions

Check candidate circle that has a max pixels

number on its boundary

This candidate is a possible circle

If number of edge pixels on its boundary greater than

Start

End

No

No

Yes

YesIf remaining edges greater

new random pixels

If finished all partitions

No

Yes

current partition lt

(2 times 120587 times r times Tr)

than (2 times 120587 times r times Tr) pick

Noncollinear

Figure 8 Proposed method workflow

(a) (b)

Figure 9 (a) Initialization problem in basic RCD in real image and (b) partition from an image

v4

v3

v1

v2

Figure 10 Four candidate circles formed using four edge pixels

possible circles in the imageOur proposedmethod suggests adynamicway to decide number of iterations for each partitionimage We introduce two other factors the number of pixelsin the partition image and the cell radius

For example assume we have a partition image as shownFigure 13(a) it has 1000 edge pixels and we search for the

circle with a minimum circumference or perimeter value(2times120587times119903times07) If 119903 = 100 and the perimeter has a width of onepixel size then it is about 440 edge pixels can be formed as acircle perimeter (2times120587times100times07) Therefore we can assumethat Figure 13(a) contains at most (2 asymp 1000440) circles

According to the information above the number ofiterations can be determined as detailed below Supposerandom pixels were selected on the first iteration such asin Figure 13(b) This case cannot represent the acceptedcircle because it does not meet with the proposed methodthreshold value 119879119889 and 119903max parameter in order to detectany of the two cells (A B) If v2 v3 and v4 form a circleas in Figure 13(b) then the proposed method rejects thatcandidate circle because the distance between v1 and the circleis greater than threshold value 119879119889 and the candidate circlersquosradius exceeds 119903max This will be similar to cases in Figures13(c) and 13(d)

Figure 14 explains in detail the subsequent processes asshown in Figure 13(b) Assuming that there are approximately


1

2 3

5

4

v4

v1

v2v3

(a)

v4

v1

v2

v3

Ta

Td

(b)

Figure 11 (a) Wrong circle detection in basic RCD algorithm and (b) some basic RCD thresholding values

Irregular circle

(a b)

r1r2

(a) (b)

Figure 12 (a) The irregular circle detection method and (b) the detection of gray pixels of the irregular cell

B

A

(a)

v3

v1

v2

v4

(b)

v3

v1

v2

v4

(c)

v3

v1

v2

v4

(d)

v3

v1 v2

v4

(e)

Figure 13 Example of (a) partition image (b) first (c) second (d) and third random pixel distribution cases demonstrating unacceptablecircles and (e) the accepted case of random pixel distribution


B

A

v3

v1

v2

v4B

A

v3

v1

v2

v4B

A

v3

v1

v2

v4B

A

v3

v2

v4

v1

B

A

v3

v2

v4

v1

B

A

v3

v2

v4

v1

Figure 14 Some cases for unacceptable circles for the case shown in Figure 13(b)

500 edge pixels forming each cell (A and B) and the cell (A)is to be detected then about 500 number of iterations arerequired to find all unacceptable circles for each distributionof random pixels as shown in Figures 13(c) and 13(d) Thisis due to found circles that do not meet requirement of 119879119889threshold and 119903max parameter values Therefore the totalnumber of iterations for this partition image may arise upto 1500 In other word 500 multiply with 3 cases as shownin Figures 13(b) 13(c) and 13(d) The only acceptable caseis shown in Figure 13(e) Here the random pixel distribu-tion meets our proposed methodrsquos condition dealing with119879119889 thresholding and 119903max parameter values in relation toaccepted circle formation Based on the previous case theproposed method requires 1500 iterations at most to find allof the unaccepted circles and the accepted circle should belocated by iteration 1501

Because the proposed method requires a certain dis-tance between random edge pixels on the same circle thenumber of iterations was increased slightly to tolerate theproposed method conditions and ensure that all possibilitiesare includedThe final expression for the number of iterationsis as follows

Iterations Number

= (Partition Edge Pixels Number

times (Number of Randomly Edge Pixels Selected minus 1))

(7)

For the previous example the proposed method requires(1000 times (4 minus 1)) = 3000 iterations to locate all of the circlesin Figure 13(a) For this case there is no fixed number ofiterations for any of the partitions The number of iterationsis fully dynamic based on the relation between the number ofedge pixels in the partition and cell radius

33 Steps of the Proposed Method

Step 1 Divide the entire binary image to a small partitionsbased on 8-neighbor connected components Each partitionis denoted by 119875119894 so that each image is defined by the set

1198751 1198752 119875119899 where 119875 is the number of edge pixels for eachpartition 119899 is the number of partitions and119891 is the number offailures that can be toleratedThe failure counter is initializedso that 119891 is set to 0

Step 2 The algorithm loops from 1198751 to 119875119899

Step 3 If |119875119894| le 119879min or 119891 ge ((V minus 1) times |119875119894|) where V is thenumber of randomly selected pixels and119879min is theminimumnumber of pixels allowed to start the method the algorithmproceeds to Step 2 otherwise four edge pixels 1199071 1199072 1199073 and1199074 are randomly selected and removed from 119875119894

Step 4 Determine the four candidate circles from the fouredge pixels such that the distance between any two of thethree pixels that forms the circle is greater than 119879119886 119879119889 thedistance between the possible circle and virtual circle Thefourth pixel must come between the two circles the distancebetween fourth edge pixel and the boundary of the virtualcircle is less than 119879119889 Figure 15 shows the proposed methodthresholding values and parameters and the radius 119903min le

119903 le 119903max If this is true the algorithm proceeds to Step 5otherwise the four edge pixels are returned to 119875119894 Then 119891 =

119891 + 1 and the algorithm continues to Step 3

Step 5 For each candidate circle the number of pixels onthe boundary is calculated using (5) and the circle with themaximum number of pixels on its boundary is assumed apossible circle because it has the highest probability

Step 6 If the number of pixels on the boundary of thepossible circle is greater than (2120587119903119879119903)Where119879119903 is a thresholdvalue equal to 05 for RBCs and 07 forWBCsThen it is a realcircle and all of the pixels found on its boundary are removedfrom the set of pixels 119875119894 and 119891 is set to 0 The algorithmreturns to Step 3 to find other circles Otherwise this possiblecircle is false 119891 = 119891 + 1 and the algorithm returns to Step 3

The proposed method has a few thresholding values(119879119886 119879119889 and 119879min) and parameters (119903min and 119903max) to bedetermined because they have direct correlation on thedetected cell results For RBCrsquos and WBCrsquos the radiuses of[119903min 119903max] were measured using simple distance measureranging from [10 50] and [20 70] pixels respectively This


v1

v2

v3

v4

Ta

Td

Figure 15 Proposed method parameters

Figure 16 RBCs detected by the proposed method

helps to increase both the probability of detecting correct cellsand performance of overlapping cells

Specifying119879min thresholding with the value (2times120587times119903mintimes119879119903) it improves the proposed method performance It stopssearching for a circle when the number of edge pixels inthe partition is less than 119879min 119879119886 thresholding values alsoinfluence the proposed method performance As 119879119886 valuedecreases both number of candidate circles and probabilityof detecting wrong circles are increasing (FP) otherwise thenumber of detected circles is reducing Apart from that the119879119889thresholding value represents the distance between the twovirtual circles and this condition helps our proposed methodto detect irregular circles (119879119889 = 20) and increases the numberof edge pixels to be included on the boundary of the possiblecircle

As conclusion several improvements were made to thebasic RCD algorithm including the initialization step solvedby addition of 8-neighbor connected component stepWhichdivide the image into smaller partitions and the ability toselect random edge pixels locally from the smaller partsinstead of choosing random edge pixels globally from theentire image This increased the probability of locating morecircles Not all blood cells have a regular circular shape Byincreasing the thickness of the circle perimeter the proposedmethod is also able to detect irregular cells Partitioningthe image and specifying the cell radius in the proposed

model improved the detection of overlapping cells Finallywe determined a relationship between the cell radius and thenumber of edge pixels in the partitioned image to control thenumber of iterations

4 Experimental Results and Discussion

41 Dataset The dataset used in this paper consisted of100 actual microscopy images of blood samples The imagescaptured with an optical laboratory microscope coupled witha Canon Power Shot G5 camera All of the images are in JPGformat with 24-bit color depth and a resolution of 2592 times1944 pixels The images were taken at different microscopemagnifications ranging from 300 to 500x [26]

42 Ground Truth The ground truth was determined by anexpert and used to validate our proposed method results

43 Evaluation Methods The results from the proposedmethod were quantitatively analyzed based on the groundtruth Precision Recall and 119865-measurements were used todetermine segmentation and counting accuracy with thefollowing equations

PR =TP

(TP + FP)


Figure 17 Overlapping cells detected by the proposed method

Table 1 Summary of results for WBCs

Set Manual count Proposed method count TP FN FP PR RC FM1 117 118 116 1 2 983 991 9872 192 199 191 1 8 959 994 9763 53 57 53 0 4 922 100 9634 15 13 13 2 0 100 866 9285 99 108 95 4 13 879 959 9176 298 350 297 1 53 848 996 9167 73 82 72 1 10 878 986 9298 28 31 26 2 5 838 928 8819 41 47 40 1 7 851 975 90910 34 37 32 2 5 864 941 901Total 950 1042 935 15 107 90 98 94

Figure 18 Real case example of irregular cells detected

RC =TP

(TP + FN)

119865-Measure = 2

(1PR + 1RC)

(8)

We divided the dataset into 10 sets For each set wedetermined the average true positive value (TP) which iswhen the agreement between the expert and the proposedmethod for the detected cells The false negative (FN) whichis when the proposedmethodwas unable to detect the cell butthe expert detected a cell and the false positive (FP) whichis when the proposed method detected a cell but the expert

did not Then we calculated the Precision (PR) Recall (RC)and 119865-measures (FM) of each set The overall PR RC and 119865-measures for the proposed method are listed on the far rightin Tables 1 and 2

Table 1 summarizes the segmentation and counting accu-racy of the proposed method for WBCs The followingsegmentation and counting accuracies were calculated usingthe proposed method for WBCs PR = 90 RC = 98 and119865-measure = 94

Table 2 summarizes the segmentation and counting accu-racy results of the proposed method for RBCs The followingis the segmentation and counting accuracies of RBCs thatwere calculated using our proposed method PR = 95 RC= 98 and 119865-measure = 96

Figure 16 shows sample images from the dataset andRBCs detected using the proposed method These samplesdid not have many overlapping cells Our proposed methodwas able to tolerate a variable degree of overlapping cellsFigure 17 shows sample images from the dataset that containoverlapping cells and the RBCs that were detected using theproposed method

Additionally our proposed method was able to performfor cases with irregularly shaped cells Figure 18 shows anexample of the irregularly shaped cells that were detectedusing the proposed method The RBCs had irregular shapesand were detected with the proposed method


Table 2 Summary of results for RBCs


(a) (b)

Figure 19 (a) Sample from our results for normal cases and (b) staining problem in some cases

Tables 1 and 2 show some testing groups presenting aquite high FP rate This is due to microscopic images usedfor Leukemia blood cells patients These images in somecases (healthy cases) and the cells are not crowded andclumped together and our proposed methodrsquos performancewas acceptable with low FP rate as shown in Figure 19(a) Insome other cases some images have problems in staining asshown in Figure 19(b) Unfortunately our proposed methodhas detected some true negative cells It finds random edgepixels that can form a circle and the number of edges onthe boundary of the possible circle is adequate however thisleads to detect the wrong circle as shown in Figure 20(a)In some other cases error in preprocessing steps such asincorrect holes filling problem creates negative effect on edgedetection phase as shown in Figure 20(b)

We evaluate our proposed methodrsquos result using a simpleleast square method in RANSAC algorithm We take 40random samples from our automatic and its correspondingmanual counting results and apply them on the RANSACalgorithm Our findings show that the selected results con-sistently fit a regression line as shown in Figure 21Thereforeour data can be classified as a consensus which meansthat our results are reasonably acceptable because most ofthe points have been classified as part of the consensusset

44 Comparison of the Results with Other Methods Theaverage accuracy of the proposedmethodwas 975 for RBCsand 984 for WBCs We compared our proposed methodwith the method presented in [22] They used the samedataset performed segmentation and counted WBCs andRBCs They used color-based segmentation using 119871 lowast 119886 lowast 119887color space and the Hough Transform for cell extractionand counting The results for all of the images were basedon a user specified CIELAB range The accuracy of theirresults ranged from 64 to 87 In our proposedmethod weperformed segmentation on the WBCS and RBCs separatelyand developed distinct preprocessing steps for each Thisstep helped us to obtain more accurate results where asthey implemented the same preprocessing for both cell typeswhich were not separatedThey used the 119887 channel forWBCsand the 119871 channel for RBCs Three sample images were usedto develop general parameter values for CIELAB which wasused to filter the cells Only three sample images were testedto justify the thresholding values which are insufficient togeneralize for all images Hence it leads to some informationloss which affected their performance Unlike basic CHTour proposed method was able to detect a variable degree ofoverlapping and irregular cells

Putzu and di Ruberto [28] used the same data set andpresented a method that only detected and counted WBCs


(a) (b)

Figure 20 (a) Wrong detection in some case from our proposed method and (b) preprocessing problem

Table 3 Accuracy of the proposed method compared with other methods

Method RBCs WBCs RBCrsquosAverage accuracy Average accuracy PR RC FM

Color-based and Hough transform [22] 64 81Morphology operators and watershed algorithm [28] mdash 935Basic Circular Hough Transform (CHT) 67 mdash 71 68 71Our proposed method 975 984 95 98 96

200 250 300 350 400 450 500 550 600200

250

300

350

400

450

500

550

600

650

Automatic counting

Man

ual c

ount

ing

Figure 21 The outliers have no influence on our proposed result asit consistently fits in a regression line

They selected 33 images to test The watershed techniquewas used to segment and separate overlapping cells and theaverage accuracy for detecting and counting the cells in the33 images was 935 Table 3 compares the accuracy of ourproposed method to other methods

Furthermore we also have conducted similar experimentusing the same dataset and the same preprocessing stepsusing the classic circular Hough Transform CHT [19] todetect RBCsOur findings justified that our proposedmethodoutperforms the classic CHT The comparison results can

be shown in Table 3 the basic CHT achieves 67 averageaccuracy for detecting RBCs whereas our proposed methodachieves 975 Our proposed method has shown betterresults in sensitivity (PR) and specificity (RC)With the intro-duction of iterative structured circle detection algorithm ourproposed method is able to detect irregular cells and toleratewith a certain degree of overlapping cells which it leadsto produce higher probability to detect more cells that arecorrect Unlike our proposed method basic CHT methodgains slightly lower performance when detecting overlappingcells as shown in Figure 22

In case of other cases irregular cells our proposedmethodpresents a better performance thanCHT in that case as shownin Figure 23

5 Conclusions

This paper proposed a method to automate the segmentationand counting of red and white blood cells using iterativestructured circle detection algorithm Several improvementswere made to the RCD algorithm including an initializationstep to find 8-neighbor connected component Additionallythe proposed model features an enhanced probability ofselecting the correct circle from four candidate circles thecapability to detect irregular cells the use of dynamic numberof iterations and improved detection of overlapping cellsTheproposed method performed the segmentation and countingof WBCs and RBCs well when results were compared withthe ground truth which was determined by experts The fol-lowing segmentation and counting accuracies were achievedusing the proposed method PR = 897 RC = 984 and


(a) (b)

Figure 22 (a) CHT RBCs detection results and (b) our proposed method RBCs detection results in overlapping cells

(a) (b)

Figure 23 (a) CHT performance when detecting irregular cells and (b) the proposed method performance when detecting irregular cells

119865-measure = 939 for WBC and PR = 953 RC = 975and 119865-measure = 964 for RBCs

Conflict of Interests

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

Acknowledgments

The authors would like to thank to Universiti KebangsaanMalaysia and Ministry of Education for providing facilitiesand financial support under ldquoAP-2012-019 Automated Med-ical Imaging Diagnostic Based on Four Critical DiseasesBrain Breast Prostate and Lung Cancerrdquo and FRGS12012SG05UKM028 Grant entitled ldquoGeneric Object Localiza-tion Algorithm for Image Segmentationrdquo Special thanks toDr Reena Rahayu Md Zin who had worked hard on givingideas

References

[1] M Habibzadeh A Krzyzak and T Fevens ldquoApplication ofpattern recognition techniques for the analysis of thin bloodsmears imagesrdquo Journal of Medical Informatics amp Technologiesvol 18 2011

[2] MM Osman ldquoNormal reference value of blood cell count redwhite and platelet of KhartoumNorth AreardquoAl NeelainMedicalJournal vol 3 no 8 2013

[3] (NIH) National Institutes of Health ldquoA service of the USNational Library of Medicinerdquo httpwwwnlmnihgovmed-lineplusencyarticle003642htm

[4] httpwwwbuzzlecomarticleshigh-white-blood-cell-counthtml

[5] H Berge D Taylor S Krishnan and T S Douglas ldquoImprovedred blood cell counting in thin blood smearsrdquo in Proceedings ofthe 8th IEEE International Symposium on Biomedical ImagingFrom Nano to Macro (ISBI rsquo11) pp 204ndash207 Chicago Ill USAApril 2011

[6] S S Savkare and S P Narote ldquoAutomatic classification ofnormal and infected blood cells for parasitemia detectionrdquoInternational Journal of Computer Science andNetwork Securityvol 11 no 2 pp 94ndash97 2011

[7] J M Sharif M F Miswan M A Ngadi S Hj and MMahadi ldquoRed blood cell segmentation using masking andwatershed algorithm a preliminary studyrdquo in Proceedings of theInternational Conference on Biomedical Engineering (ICoBE rsquo12)pp 258ndash262 Penang Malaysia February 2012

[8] L B Damahe P G Student R G College and R G C EnggldquoSegmentation based approach to detect parasites and RBCs inblood cell imagesrdquo International Journal of Computer Scienceand Applications vol 4 no 2 2011

[9] L B D Panchbhai and V Vishal ldquoRBCs and parasites segmen-tation from thin smear blood cell imagesrdquo International Journalof Image Graphics and Signal Processing vol 10 no 10 pp 54ndash60 2012

[10] S Khan A Khan and A Naseem ldquoAn accurate and costeffective approach to blood cell countrdquo International Journal ofComputer Applications vol 50 no 1 pp 18ndash24 2012


[11] N Nguyen A Duong and H Vu ldquoCell splitting with highdegree of overlapping in peripheral blood smearrdquo InternationalJournal of Computer Theory and Engineering vol 3 no 3 pp473ndash478 2011

[12] A Rhodes and L Bai ldquoCircle detection using a gabor annulusrdquoinProcedings of the BritishMachine Vision Conference pp 1081ndash10811 September 2011

[13] K-L Chung and Y-H Huang ldquoA pruning-and-voting strategyto speed up the detection for lines circles and ellipsesrdquo Journalof Information Science and Engineering vol 24 no 2 pp 503ndash520 2008

[14] R K K Yip P K-S Tam and D N K Leung ldquoModificationof Hough transform for circles and ellipses detection using a 2-dimensional arrayrdquo Pattern Recognition vol 25 no 9 pp 1007ndash1022 1992

[15] E P Lyvers and O R Mitchell ldquoPrecision edge contrast andorientation estimationrdquo IEEE Transactions on Pattern Analysisand Machine Intelligence vol 10 no 6 pp 927ndash937 1988

[16] C-T Ho and L-H Chen ldquoA fast ellipsecircle detector usinggeometric symmetryrdquo Pattern Recognition vol 28 no 1 pp 117ndash124 1995

[17] C-T Ho and L-H Chen ldquoA high-speed algorithm for ellipticalobject detectionrdquo IEEE Transactions on Image Processing vol 5no 3 pp 547ndash550 1996

[18] L Xu E Oja and P Kultanen ldquoA new curve detection methodrandomized Hough transform (RHT)rdquo Pattern RecognitionLetters vol 11 no 5 pp 331ndash338 1990

[19] L Xu and E Oja ldquoRandomized Hough transform (RHT)basic mechanisms algorithms and computational complexi-tiesrdquo CVGIP Image Understanding vol 57 no 2 pp 131ndash1541993

[20] S-H Chiu J-J Liaw and K-H Lin ldquoA fast randomized Houghtransform for circlecircular arc recognitionrdquo InternationalJournal of Pattern Recognition and Artificial Intelligence vol 24no 3 pp 457ndash474 2010

[21] T-C Chen and K-L Chung ldquoAn efficient randomized algo-rithm for detecting circlesrdquo Computer Vision and Image Under-standing vol 83 no 2 pp 172ndash191 2001

[22] N H Mahmood P C Lim S M Mazalan and M A ARazak ldquoBlood cells extraction using color based segmentationtechniquerdquo International Journal of Life Sciences Biotechnologyand Pharma Research vol 2 no 2 2013

[23] N HMahmood andM AMansor ldquoRed blood cells estimationusing Hough transform techniquerdquo Signal amp Image Processingvol 3 no 2 pp 53ndash64 2012

[24] B Venkatalakshmi and K Thilagavathi ldquoAutomatic red bloodcell counting using Hough transformrdquo in Proceedings of theIEEE Conference on Information and Communication Technolo-gies (ICT rsquo13) pp 267ndash271 JeJu Island Korea April 2013

[25] M Maitra R K Gupta and M Mukherjee ldquoDetection andcounting of red blood cells in blood cell images using Houghtransformrdquo International Journal of Computer Applications vol53 no 16 pp 18ndash22 2012

[26] R D Labati V Piuri and F Scotti ldquoAll-IDB the acutelymphoblastic leukemia image database for image processingrdquoin Proceedings of the 18th IEEE International Conference onImage Processing (ICIP rsquo11) pp 2045ndash2048 Brussels BelgiumSeptember 2011

[27] E Cuevas D Oliva M Dıaz D Zaldivar M Perez-cisnerosand G Pajares ldquoWhite blood cell segmentation by circle detec-tion using electromagnetism-like optimizationrdquoComputational

and Mathematical Methods in Medicine vol 2013 Article ID395071 15 pages 2013

[28] L Putzu and C di Ruberto ldquoWhite blood cells identificationand counting from microscopic blood imagerdquo World Academyof Science Engineering and Technology vol 7 no 1 pp 363ndash3702013

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


10000 WBCs per microliter of blood [4] High WBC counts(above 30000 cells per microliter) indicate an infectionsystemic illness inflammation allergy leukemia or burn-induced tissue injury If leukemia is suspected analysis ofblood smear is done to look for morphology of the leukemiccells and followed by bone marrow examinations [2ndash4] Forplatelets which are small blood cell fragments that assist inblood clotting normal counts range from 150000 to 450000platelets per microliter In patients with low platelet countsuch as in patients with dengue infection their platelet countis monitored closely and the value is within critical levelthe patient might need platelet transfusion [2] Generallyany abnormal blood smear reading indicates an infection ordisease

Malaria and Babesia are parasites that infect RBCs theanalysis of thin blood smear remains the gold standard fordiagnosis in such disease [5] Microscopy images are alsostill used for early diagnosis analysis and count of someblooddisorder such as sickle cell anemia and leukemia beforeconfirming it with other laboratory tests However manualor visual quantification of parasitemia in thin blood filmsand WBCs in leukemia is an extremely tedious subjectiveand time-consuming task with a high probability of count-ing error [6 7] An accurate segmentation and countingmechanism that gathers information about the distribution ofmicroscopic particles may help diagnose abnormalities dur-ing clinical analysis Our objective in this paper is to developand validate an algorithm that segments and automaticallycounts red and white blood cells in microscopy imagesThe ground truth of the images was determined by expertsFor the evaluation quantitative analyses were performedon the segmentation results based on the ground truthand the 119865-measure method was used to confirm accuracyIn the following sections of this paper we will summarizerelated work on the segmentation and counting of RBCs andWBCs (Section 2) present themethodology used (Section 3)discuss the results and experiments (Section 4) and reviewthe conclusions

2 Related Work

Many researchers have investigated blood cell segmentationand counting Some researchers [5 8ndash10] usedmorphologicaloperations and thresholding to do the segmentation andcounting Berge et al [5] presented an approach basedon a morphological method and iterative threshold tech-niques Segmentation was performed on red blood cellswhich included clumped cells and boundary curvatureswere used to construct a Delaunay triangulation They usedreal microscopy images prepared in the laboratory andthe ground truth was determined by a laboratory expertA 28 difference was calculated between the manual andautomatic counting of red blood cellsTheirmethod tolerateda degree of overlapping but in cases with a high degreeof overlapping cells the cells were unable to be detectedAdditionally the iterative threshold method was unable todetect faint red cells Damahe et al [8] used the S andV imagecomponents of a HSV color model with Zackrsquos thresholdingtechnique for cell segmentationThresholding combinedwith

a sequential edge-linking algorithm was used to increasesegmentation accuracy The experimental dataset and bloodcell images were collected from the Dhruv Pathology Laband the CDC-DPDx respectively and the ground truth wasdetermined by experts The dataset size was limited SeveralRBCs that were detected possessed holes Additional prepro-cessing steps should have been implemented to increase theaccuracy Panchbhai and Vishal [9] proposed an automatedanalysis that counted red blood cells and detected malariaparasites in thin blood smear samples The green color layerwas processed to count all of the RBCs segmentation wasperformed on the infected RBCs using Otsu thresholding Ahistogramwas used to determine the optimal threshold CDCdatasets were used for the experimental part and the groundtruth was determined by a pathology expert who comparedtheir results with the manual counts However becausethe detection algorithm used morphology and thresholdingtheir method was unable to detect clumped and overlappingcells Khan et al [10] proposed a method to count WBCsRBCs and platelets Several preprocessing steps were per-formed before converting the image to binary Segmentationand cell counting were performed based on the optimalthreshold value which was determined from a histogramThey achieved 95 accuracy with their proposed methodcompared to manual counting and a hematology analyzerHowever this method is unable to detect overlapping cellsWhen using iterative thresholds the probability of losinguseful information from the image is high this decreases theaccuracy of segmentation

Nguyen et al [11] used distance transform to solve theoverlapping cells problem they proposed a method thatconcentrated on clumped cells First they assigned centralpoints based on a distance transform The optimal centerpoints were selected by checking the degree of boundarycovering the center point and the average size of a cell wasestimated by the extraction of a single cellThen an algorithmwas developed that used a single cell mask to split the cellsTheir dataset ground truth was labeled by experts and theaccuracy of the proposed method using 119865-measurementswas 935 and 82 Clumped cells were tolerated but thecells had to be regularly shaped and focused at a highmagnification Not all blood cells have regular cell shapesand this is especially true if the blood cells are diseasedTherefore cell detection methods should be able to detectirregular cells Additionally because of noise the performanceof their method was not good

Rhodes and Bai [12] presented a circle detection methodusing specific properties of Gabor wavelet filter to detect theimage features such as circularity It is able to extract radiuswraps around the origin and the plane wave radiates fromthe center of the filter They test their proposed method onsynthetic images and real microscopic images which allow acertain degree of overlapping cellsThe results were 913 and87 of the cells detected in the two microscopic test images

Since blood cells are approximately circular shape circledetection algorithms can still handle the challenge of bloodcells detection Hough Transform is considered as one of themost known algorithms for line and circle detection It wasdeveloped by Richard Duda and Peter Hart in 1972 [13] they













counting

thresholding

Read image

thresholding


image


WBCs RBCs

Subtract WBCs




3 Methodology





thresholding values




Erode the image

Fill holes









(a) (b)

(c) (d)










(a) (b)

(c) (d)




Remove white cells


Fill holes









119886119894119895119896 =

1003816100381610038161003816100381610038161003816

1199092119895+1199102119895minus(1199092

119894+1199102119894 ) 2(119910119895minus119910119894)

1199092119896+1199102119896minus(1199092

119896+1199102119894 ) 2(119910119896minus119910119894)

1003816100381610038161003816100381610038161003816


(2)

119887119894119895119896 =

1003816100381610038161003816100381610038161003816

2(119909119895minus119909119894) 1199092

119895+1199102119895minus(1199092

119894+1199102119894 )

2(119909119896minus119909119894) 1199092

119896+1199102119896minus(1199092

119894+1199102119894 )

1003816100381610038161003816100381610038161003816


(3)

119903119894119895119896 =radic(119909119894 minus 119886119894119895119896)

2+ (119910119894 minus 119887119894119895119896)

2 (4)


(a) (b)

(c)



119889119911rarr 119894119895119896 =

1003816100381610038161003816100381610038161003816100381610038161003816


2minus 119903119894119895119896

1003816100381610038161003816100381610038161003816100381610038161003816

(5)







(a) (b)

(c)



119862119894119895119896 =




(6)








from edges set


Loop






Start

End

No

No

Yes


new random pixels


No

Yes




Noncollinear


(a) (b)


v4

v3

v1

v2








1

2 3

5

4

v4

v1

v2v3

(a)

v4

v1

v2

v3

Ta

Td

(b)


Irregular circle

(a b)

r1r2

(a) (b)


B

A

(a)

v3

v1

v2

v4

(b)

v3

v1

v2

v4

(c)

v3

v1

v2

v4

(d)

v3

v1 v2

v4

(e)



B

A

v3

v1

v2

v4B

A

v3

v1

v2

v4B

A

v3

v1

v2

v4B

A

v3

v2

v4

v1

B

A

v3

v2

v4

v1

B

A

v3

v2

v4

v1




Iterations Number



(7)














v1

v2

v3

v4

Ta

Td











PR =TP

(TP + FP)






RC =TP

(TP + FN)

119865-Measure = 2

(1PR + 1RC)

(8)










(a) (b)







(a) (b)





200 250 300 350 400 450 500 550 600200

250

300

350

400

450

500

550

600

650

Automatic counting

Man

ual c

ount

ing






5 Conclusions



(a) (b)


(a) (b)





Acknowledgments


References




































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of




























counting

thresholding

Read image

thresholding


image


WBCs RBCs

Subtract WBCs




3 Methodology





thresholding values




Erode the image

Fill holes









(a) (b)

(c) (d)










(a) (b)

(c) (d)




Remove white cells


Fill holes









119886119894119895119896 =

1003816100381610038161003816100381610038161003816

1199092119895+1199102119895minus(1199092

119894+1199102119894 ) 2(119910119895minus119910119894)

1199092119896+1199102119896minus(1199092

119896+1199102119894 ) 2(119910119896minus119910119894)

1003816100381610038161003816100381610038161003816


(2)

119887119894119895119896 =

1003816100381610038161003816100381610038161003816

2(119909119895minus119909119894) 1199092

119895+1199102119895minus(1199092

119894+1199102119894 )

2(119909119896minus119909119894) 1199092

119896+1199102119896minus(1199092

119894+1199102119894 )

1003816100381610038161003816100381610038161003816


(3)

119903119894119895119896 =radic(119909119894 minus 119886119894119895119896)

2+ (119910119894 minus 119887119894119895119896)

2 (4)


(a) (b)

(c)



119889119911rarr 119894119895119896 =

1003816100381610038161003816100381610038161003816100381610038161003816


2minus 119903119894119895119896

1003816100381610038161003816100381610038161003816100381610038161003816

(5)







(a) (b)

(c)



119862119894119895119896 =




(6)








from edges set


Loop






Start

End

No

No

Yes


new random pixels


No

Yes




Noncollinear


(a) (b)


v4

v3

v1

v2








1

2 3

5

4

v4

v1

v2v3

(a)

v4

v1

v2

v3

Ta

Td

(b)


Irregular circle

(a b)

r1r2

(a) (b)


B

A

(a)

v3

v1

v2

v4

(b)

v3

v1

v2

v4

(c)

v3

v1

v2

v4

(d)

v3

v1 v2

v4

(e)



B

A

v3

v1

v2

v4B

A

v3

v1

v2

v4B

A

v3

v1

v2

v4B

A

v3

v2

v4

v1

B

A

v3

v2

v4

v1

B

A

v3

v2

v4

v1




Iterations Number



(7)














v1

v2

v3

v4

Ta

Td











PR =TP

(TP + FP)






RC =TP

(TP + FN)

119865-Measure = 2

(1PR + 1RC)

(8)










(a) (b)







(a) (b)





200 250 300 350 400 450 500 550 600200

250

300

350

400

450

500

550

600

650

Automatic counting

Man

ual c

ount

ing






5 Conclusions



(a) (b)


(a) (b)





Acknowledgments


References




































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















(a) (b)

(c) (d)










(a) (b)

(c) (d)




Remove white cells


Fill holes









119886119894119895119896 =

1003816100381610038161003816100381610038161003816

1199092119895+1199102119895minus(1199092

119894+1199102119894 ) 2(119910119895minus119910119894)

1199092119896+1199102119896minus(1199092

119896+1199102119894 ) 2(119910119896minus119910119894)

1003816100381610038161003816100381610038161003816


(2)

119887119894119895119896 =

1003816100381610038161003816100381610038161003816

2(119909119895minus119909119894) 1199092

119895+1199102119895minus(1199092

119894+1199102119894 )

2(119909119896minus119909119894) 1199092

119896+1199102119896minus(1199092

119894+1199102119894 )

1003816100381610038161003816100381610038161003816


(3)

119903119894119895119896 =radic(119909119894 minus 119886119894119895119896)

2+ (119910119894 minus 119887119894119895119896)

2 (4)


(a) (b)

(c)



119889119911rarr 119894119895119896 =

1003816100381610038161003816100381610038161003816100381610038161003816


2minus 119903119894119895119896

1003816100381610038161003816100381610038161003816100381610038161003816

(5)







(a) (b)

(c)



119862119894119895119896 =




(6)








from edges set


Loop






Start

End

No

No

Yes


new random pixels


No

Yes




Noncollinear


(a) (b)


v4

v3

v1

v2








1

2 3

5

4

v4

v1

v2v3

(a)

v4

v1

v2

v3

Ta

Td

(b)


Irregular circle

(a b)

r1r2

(a) (b)


B

A

(a)

v3

v1

v2

v4

(b)

v3

v1

v2

v4

(c)

v3

v1

v2

v4

(d)

v3

v1 v2

v4

(e)



B

A

v3

v1

v2

v4B

A

v3

v1

v2

v4B

A

v3

v1

v2

v4B

A

v3

v2

v4

v1

B

A

v3

v2

v4

v1

B

A

v3

v2

v4

v1




Iterations Number



(7)














v1

v2

v3

v4

Ta

Td











PR =TP

(TP + FP)






RC =TP

(TP + FN)

119865-Measure = 2

(1PR + 1RC)

(8)










(a) (b)







(a) (b)





200 250 300 350 400 450 500 550 600200

250

300

350

400

450

500

550

600

650

Automatic counting

Man

ual c

ount

ing






5 Conclusions



(a) (b)


(a) (b)





Acknowledgments


References




































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















(a) (b)

(c) (d)




Remove white cells


Fill holes









119886119894119895119896 =

1003816100381610038161003816100381610038161003816

1199092119895+1199102119895minus(1199092

119894+1199102119894 ) 2(119910119895minus119910119894)

1199092119896+1199102119896minus(1199092

119896+1199102119894 ) 2(119910119896minus119910119894)

1003816100381610038161003816100381610038161003816


(2)

119887119894119895119896 =

1003816100381610038161003816100381610038161003816

2(119909119895minus119909119894) 1199092

119895+1199102119895minus(1199092

119894+1199102119894 )

2(119909119896minus119909119894) 1199092

119896+1199102119896minus(1199092

119894+1199102119894 )

1003816100381610038161003816100381610038161003816


(3)

119903119894119895119896 =radic(119909119894 minus 119886119894119895119896)

2+ (119910119894 minus 119887119894119895119896)

2 (4)


(a) (b)

(c)



119889119911rarr 119894119895119896 =

1003816100381610038161003816100381610038161003816100381610038161003816


2minus 119903119894119895119896

1003816100381610038161003816100381610038161003816100381610038161003816

(5)







(a) (b)

(c)



119862119894119895119896 =




(6)








from edges set


Loop






Start

End

No

No

Yes


new random pixels


No

Yes




Noncollinear


(a) (b)


v4

v3

v1

v2








1

2 3

5

4

v4

v1

v2v3

(a)

v4

v1

v2

v3

Ta

Td

(b)


Irregular circle

(a b)

r1r2

(a) (b)


B

A

(a)

v3

v1

v2

v4

(b)

v3

v1

v2

v4

(c)

v3

v1

v2

v4

(d)

v3

v1 v2

v4

(e)



B

A

v3

v1

v2

v4B

A

v3

v1

v2

v4B

A

v3

v1

v2

v4B

A

v3

v2

v4

v1

B

A

v3

v2

v4

v1

B

A

v3

v2

v4

v1




Iterations Number



(7)














v1

v2

v3

v4

Ta

Td











PR =TP

(TP + FP)






RC =TP

(TP + FN)

119865-Measure = 2

(1PR + 1RC)

(8)










(a) (b)







(a) (b)





200 250 300 350 400 450 500 550 600200

250

300

350

400

450

500

550

600

650

Automatic counting

Man

ual c

ount

ing






5 Conclusions



(a) (b)


(a) (b)





Acknowledgments


References




































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















(a) (b)

(c)



119889119911rarr 119894119895119896 =

1003816100381610038161003816100381610038161003816100381610038161003816


2minus 119903119894119895119896

1003816100381610038161003816100381610038161003816100381610038161003816

(5)







(a) (b)

(c)



119862119894119895119896 =




(6)








from edges set


Loop






Start

End

No

No

Yes


new random pixels


No

Yes




Noncollinear


(a) (b)


v4

v3

v1

v2








1

2 3

5

4

v4

v1

v2v3

(a)

v4

v1

v2

v3

Ta

Td

(b)


Irregular circle

(a b)

r1r2

(a) (b)


B

A

(a)

v3

v1

v2

v4

(b)

v3

v1

v2

v4

(c)

v3

v1

v2

v4

(d)

v3

v1 v2

v4

(e)



B

A

v3

v1

v2

v4B

A

v3

v1

v2

v4B

A

v3

v1

v2

v4B

A

v3

v2

v4

v1

B

A

v3

v2

v4

v1

B

A

v3

v2

v4

v1




Iterations Number



(7)














v1

v2

v3

v4

Ta

Td











PR =TP

(TP + FP)






RC =TP

(TP + FN)

119865-Measure = 2

(1PR + 1RC)

(8)










(a) (b)







(a) (b)





200 250 300 350 400 450 500 550 600200

250

300

350

400

450

500

550

600

650

Automatic counting

Man

ual c

ount

ing






5 Conclusions



(a) (b)


(a) (b)





Acknowledgments


References




































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















(a) (b)

(c)



119862119894119895119896 =




(6)








from edges set


Loop






Start

End

No

No

Yes


new random pixels


No

Yes




Noncollinear


(a) (b)


v4

v3

v1

v2








1

2 3

5

4

v4

v1

v2v3

(a)

v4

v1

v2

v3

Ta

Td

(b)


Irregular circle

(a b)

r1r2

(a) (b)


B

A

(a)

v3

v1

v2

v4

(b)

v3

v1

v2

v4

(c)

v3

v1

v2

v4

(d)

v3

v1 v2

v4

(e)



B

A

v3

v1

v2

v4B

A

v3

v1

v2

v4B

A

v3

v1

v2

v4B

A

v3

v2

v4

v1

B

A

v3

v2

v4

v1

B

A

v3

v2

v4

v1




Iterations Number



(7)














v1

v2

v3

v4

Ta

Td











PR =TP

(TP + FP)






RC =TP

(TP + FN)

119865-Measure = 2

(1PR + 1RC)

(8)










(a) (b)







(a) (b)





200 250 300 350 400 450 500 550 600200

250

300

350

400

450

500

550

600

650

Automatic counting

Man

ual c

ount

ing






5 Conclusions



(a) (b)


(a) (b)





Acknowledgments


References




































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of




















from edges set


Loop






Start

End

No

No

Yes


new random pixels


No

Yes




Noncollinear


(a) (b)


v4

v3

v1

v2








1

2 3

5

4

v4

v1

v2v3

(a)

v4

v1

v2

v3

Ta

Td

(b)


Irregular circle

(a b)

r1r2

(a) (b)


B

A

(a)

v3

v1

v2

v4

(b)

v3

v1

v2

v4

(c)

v3

v1

v2

v4

(d)

v3

v1 v2

v4

(e)



B

A

v3

v1

v2

v4B

A

v3

v1

v2

v4B

A

v3

v1

v2

v4B

A

v3

v2

v4

v1

B

A

v3

v2

v4

v1

B

A

v3

v2

v4

v1




Iterations Number



(7)














v1

v2

v3

v4

Ta

Td











PR =TP

(TP + FP)






RC =TP

(TP + FN)

119865-Measure = 2

(1PR + 1RC)

(8)










(a) (b)







(a) (b)





200 250 300 350 400 450 500 550 600200

250

300

350

400

450

500

550

600

650

Automatic counting

Man

ual c

ount

ing






5 Conclusions



(a) (b)


(a) (b)





Acknowledgments


References




































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















1

2 3

5

4

v4

v1

v2v3

(a)

v4

v1

v2

v3

Ta

Td

(b)


Irregular circle

(a b)

r1r2

(a) (b)


B

A

(a)

v3

v1

v2

v4

(b)

v3

v1

v2

v4

(c)

v3

v1

v2

v4

(d)

v3

v1 v2

v4

(e)



B

A

v3

v1

v2

v4B

A

v3

v1

v2

v4B

A

v3

v1

v2

v4B

A

v3

v2

v4

v1

B

A

v3

v2

v4

v1

B

A

v3

v2

v4

v1




Iterations Number



(7)














v1

v2

v3

v4

Ta

Td











PR =TP

(TP + FP)






RC =TP

(TP + FN)

119865-Measure = 2

(1PR + 1RC)

(8)










(a) (b)







(a) (b)





200 250 300 350 400 450 500 550 600200

250

300

350

400

450

500

550

600

650

Automatic counting

Man

ual c

ount

ing






5 Conclusions



(a) (b)


(a) (b)





Acknowledgments


References




































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















B

A

v3

v1

v2

v4B

A

v3

v1

v2

v4B

A

v3

v1

v2

v4B

A

v3

v2

v4

v1

B

A

v3

v2

v4

v1

B

A

v3

v2

v4

v1




Iterations Number



(7)














v1

v2

v3

v4

Ta

Td











PR =TP

(TP + FP)






RC =TP

(TP + FN)

119865-Measure = 2

(1PR + 1RC)

(8)










(a) (b)







(a) (b)





200 250 300 350 400 450 500 550 600200

250

300

350

400

450

500

550

600

650

Automatic counting

Man

ual c

ount

ing






5 Conclusions



(a) (b)


(a) (b)





Acknowledgments


References




































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















v1

v2

v3

v4

Ta

Td











PR =TP

(TP + FP)






RC =TP

(TP + FN)

119865-Measure = 2

(1PR + 1RC)

(8)










(a) (b)







(a) (b)





200 250 300 350 400 450 500 550 600200

250

300

350

400

450

500

550

600

650

Automatic counting

Man

ual c

ount

ing






5 Conclusions



(a) (b)


(a) (b)





Acknowledgments


References




































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of





















RC =TP

(TP + FN)

119865-Measure = 2

(1PR + 1RC)

(8)










(a) (b)







(a) (b)





200 250 300 350 400 450 500 550 600200

250

300

350

400

450

500

550

600

650

Automatic counting

Man

ual c

ount

ing






5 Conclusions



(a) (b)


(a) (b)





Acknowledgments


References




































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of



















(a) (b)







(a) (b)





200 250 300 350 400 450 500 550 600200

250

300

350

400

450

500

550

600

650

Automatic counting

Man

ual c

ount

ing






5 Conclusions



(a) (b)


(a) (b)





Acknowledgments


References




































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















(a) (b)





200 250 300 350 400 450 500 550 600200

250

300

350

400

450

500

550

600

650

Automatic counting

Man

ual c

ount

ing






5 Conclusions



(a) (b)


(a) (b)





Acknowledgments


References




































of






Disease Markers



OncologyJournal of





PPAR Research



Journal of

ObesityJournal of

















(a) (b)


(a) (b)





Acknowledgments


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 automatic detection and quantification of...

Documents