Automatic Bone Marrow White Blood Cell Classfication Using .
International Journal of Scientific & Technology Research Volume 1, Issue 4, May 2012ISSN 2277-8616Automatic Bone Marrow White Blood CellClassfication using MorphologicalGranulometric Feature of NucleusShraddha Shivhare, Rajesh ShrivastavaAbstract— The differential counting of white blood cell provides invaluable information to doctors for diagnosis and treatment of many diseases.manually counting of white blood cell is a tiresome, time-consuming and susceptible to error procedure due to the tedious nature of this process, anautomatic system is preferable. in this automatic process, segmentation and classification of white blood cell are the most important stages. Anautomatic segmentation technique for microscopic bone marrow white blood cell images is proposed in this paper. The segmentation techniquesegments each cell image into three regions, i.e., nucleus, cytoplasm, and background. In this paper, we investigate whether information about thenucleus alone is adequate to classify white blood cells. This is important because segmentation of nucleus is much easier than the segmentation of theentire cell, especially in the bone marrow where the white blood cell density is very high. Even though the boundaries between cell classes are not welldefined and there are classification variations among experts, we achieve a promising classification performance using neural networks with fivefoldcross validation in which Bayes’ classifiers and artificial neural networks are applied as classifiers.The classification performances are evaluated by twoevaluation measures: traditional and classwise classificationrates. we compare our results with other classifiers and previously proposed nucleus-basedfeatures. The results showthat the features using nucleus alone can be utilized to achieve aclassification rate of 77% on the test sets. Moreover, theclassification performance is better in the class wise sense when the a priori information is suppressed in both the classifiers.Index Terms—Automatic white blood cell classification, granulometric moments, mathematical morphology, pattern spectrum, white blood celldifferential counts.—————————— ——————————I. INTRODUCTIONIn the traditional process, doctors analyze human blood bymicroscope. This manual process is time consuming andsusceptible to error procedure thus, an automatic systemseems necessary and helpful. The automatic DBC systemmay require four stages: 1) acquisition, 2) detection, 3)feature extraction, and 4) classification.Fig. 1. Cell samples and corresponding manuallysegmented images in the myelocytic or granulocyticseries.Author – Shraddha ShivhareRGPV university, M.E. SS (IV SEM)Shri Ram institute Of TechnologyJabalpur (M.P.),482001, Indiashraddhashivhare@rediffmail.comCo Author – Rajesh ShrivastavaRGPV University Asst Professor: CS DeptShri Ram institute of technologyJabalpur (M.P.),482001, Indiarajesh 5479@rediffmail.comIn the first stage, the blood smear is magnified to a suitablescale under the microscope, and then transformed to adigital image. In the sec- ond stage, cell segmentation isused to produce a number of single-cell images. Then eachsingle-cell image is segmented into three regions: 1)nucleus, 2) cytoplasm, and 3) background. In the third step,feature vectors of color, texture, and shape of the segmented cell and its nucleus are extracted. In the last stepaccording to the extracted feature vectors, each WBC islabeled by a classifier. The most important stage is the cellsegmentation because the accuracy of segmentation playsa crucial role in the subsequent stages .Bone marrowcellsare normally diagnosed by light microscopy. Flowcytometry, which is normally used for differential blood cellcounting of peripheral blood, is not suitable for bonemarrow cells. This is because, in addition to its high priceand complicated structure, markedly hypercellular orpacked bone marrow and sclerotic bone marrow may yieldtoo few cells for adequate analysis by flow cytometry Datafrom flow cytometry should always be correlated with thatfrom light microscopy. White blood cells in bone marrow areclassified according to their maturation stages.When awhite blood cell becomes older, its size, the size and shapeof the nucleus, and many other characteristics change.White blood cells in the myelocytic or granulocytic seriescan be classified into six classes, i.e., myeloblast,promyelocyte, myelocyte, metamyelocyte, band, andpolymorphonuclear (PMN) in that order from the youngestto the oldest cells Samples of white blood cells in each ofthese classes are shown in Fig. 1. Three samples of eachclass are shown to illustrate the possible variation withineach class. Due to the tediousness of manual systems,several methods have been proposed for automatic orpartially automatic counting systems. Most of them,however, are for the applications in peripheral blood ratherthan for bone marrow. It should be noted that white bloodcells in bone marrow are much denser than those inperipheral blood; therefore, segmentation of white blood125IJSTR 2012www.ijstr.org
International Journal of Scientific & Technology Research Volume 1, Issue 4, May 2012cells in bone marrow is a more difficult problem thansegmentation in peripheral blood. Moreover, the immaturecells are normally seen only in the bone marrow which,thus, makes cell classification in bone marrow a moredifficult and also a complex problem Sanei and Leeachieved classification rates of more than 95% for maturecells in normal blood, more than 85% for immature cells inthe blood, but only just over 70% for immature cells in bonemarrow. Most of the proposed automatic techniques followthe traditional manual process of detecting a cell, extractingits features, classifying the cell, and then updating thecounts. There are three types of cells in normal humanblood: red cells, leukocyte or white cells and blood platelets.Generally, red cells are simple and similar. While white cellscontain nucleus and cytoplasm and there are different typesof them. In our paper we are considering only the nucleus.In blood smear, number of red cells is many more thanwhite cells. For example, an image may contain up to 100red cells and only 1 to 3 white cells. In laboratories,haematologists analyse human blood by microscope. Theirmain tasks in this area are: red cell count, white cell countand blood disorder detection. It is tedious task to locate,identify and count these classes of cells. Due to theimportance of these processes, an automated systemseems necessary and helpful. White cells are clinicallymore important than red cells and many of blood disordersare related to them. Thus, accurate segmentation of thesecells is very important.white bloodcells count is used todetermine the presence of an infection in the human body.The segmentation step is very crucial because the accuracyof the subsequent feature extraction and classificationdepends on the correct segmentation of white blood cells. Itis also a difficult and challenging problem due to thecomplex nature of the cells and uncertainty in themicroscopic image. Therefore, this step is the mostimportant challenge in many literatures and improvement ofcell segmentation has been the most common effort inmany research works. Most of the proposed automatictechniques follow the traditional manual process ofdetecting a cell, extracting its features, classifying the cell,and then updating the counts. Our previous researchconcentrated on the counting of white blood cellsspecifically in bone marrow We developed the mixingtheories of mathematical morphology, and applied them tothe counting problem without classification We alsodeveloped a new training algorithm for neural networks inorder to count the number of different cell classes withoutclassification In this paper, we propose a method for theclassification of white blood cells using only their nucleusinformation. This idea is very useful in practice because oneof the difficulties in the differential counting in bone marrowis how to deal with the cells that touch each other. Thisproblem occurs frequently in cells of the bone marrowbecause there the white blood are very dense. If the cellclassification is based only on the information about thenucleus, then we do not need to segment the entire cell,and only nucleus segmentation is adequate. Although manytechniques have been applied to cell segmentation, thisproblem is not solved, especially in touching cells. Todecouple the effects of segmentation errors, we extractfeatures from manually segmented nucleus of a white bloodcell based on morphological granulometries. We applyBayes classifiers and artificial neural networks to theISSN 2277-8616problem of white blood cell classification of single-cellimages and compare their results to those obtained by anexpert.We also compare the results to other classifiers andother previously proposed features.in this paper we alsopropose an algorithm that keeps the original shape of theblood cell and uses infor,mation of this shape to split theoverlapped regions by drawing a conical curve.II METHODOLOGYIn this research, we use artificial neural networks as ourclassifiers in the six-class problem. The input features aremainly extracted from pattern spectra of nucleus. To bemore specific, we use six features – two are area-based,the remaining four are morphologybased.1.Mathematical MorphologyMathematical morphology is a branch of nonlinear imageprocessing and analysis. Morphological methods are usedin many ways in image processing, for example,enhancement, segmentation, restoration, edge detection,texture analysis, shape analysis, etc It is also applied toseveral research areas, such as, medical imaging, remotesensing, military applications, etc.2.Morphological OperationsMorphological operations are non-linear, rymorphological operations only. Binary images can beconsidered as functions on two dimensional grids withvalues of 0 or 1 or, equivalently, as characteristic functionsof subsets of the two-dimensional plane. The concept ofstructuring element is fundamental in morphology; it is theanalogue of a convolution mask in linear image processing.The basic morphological operations involving an image Sand a structuring element E areerosion: S E {S – e: e E}dilation: S E {E s: s S},where and denote the set intersection andunion, respectively. A x denotes the translationof a set A by a point x, i.e.A x {a x: a A}.The closing and opening operations, derived from theerosion and dilation, are defined byclosing: S E (S (–E)) (–E)opening: S E (S E) Ewhere –E {–e: e E} denotes the 180 rotation of E aboutthe origin.126IJSTR 2012www.ijstr.org
International Journal of Scientific & Technology Research Volume 1, Issue 4, May 2012ISSN 2277-86163. Pattern SpectrumWe successively apply the opening operation to an imageand increase the size of structuring element in order todiminish the image. Let Ω(t) be area of S tE where t is a realnumber and Ω(0) is area of S. Ω(t) is called a sizedistribution. The normalized size distribution Φ(t) 1 –Ω(t)/Ω(0), and dΦ(t)/dt are called granulometric sizedistribution or pattern spectrum of S.4.Feature ExtractionWe focus our feature extraction on themorphology - basedfeatures. Hence, we will introduce their derivations here.For a random set S, Ω(t) is a random function. Thenormalized size distribution Φ(t) 1 – Ω(t)/Ω(0), the socalled pattern spectrum of S, is a probability distributionfunction. Its moments, μ(1)(S), μ(2)(S), , are thereforerandom variables namely granulometric moments. In thisresearch, we consider nuclei as an object of interest. Wecalculate a pattern spectrum of each cell’s nuclei andcalculate the first and second granulometric moments of thepattern spectrum to achieve our features. We also extracttwo other features from each nucleus, i.e., the area of thenucleus and the location of its pattern spectrum’s peak. We,therefore, determine four feature values for each cell image.To form an input feature vector to a neural network, weextract six features from each cell,i.e.,o the area of cell,o the nuclei-to-cytoplasm ratio,o the maximum value of a pattern spectrum,o the location where the maximum valueof a pattern spectrum occurs,o the first granulometric moments ando the second granulometric moments.We select a small digital disc as the structuringelement in our experiments. The structuringelement is shown in Figure 3.0110111111110110Fig. 3 Structuring element used in the experiments.5.Bayes ClassifierBayes classifier is a traditional statistical-based classifierthat analyzes discriminant functions by using Bayes’theorem. Consider a classifier, we assign an input vector xto class Ck if yk(x) yj(x) for all j 6 k. By choosing yk(x) P(Ck x), this posterior probability is the probability of patternbelonging to class Ck when we observe the input vector x.Bayes’ theorem yields yk(x) P(Ck x) p(x Ck)P(Ck) p(x) ,(9) where p(x) is the unconditional density and P(Ck) is theprior probability of the kth class. Assuming the conditionalprobability density is normal, i.e.,A theorem describing how the conditional probability of aset of possible causes for a given observed event can becomputed from knowledge of the probability of each causeand the conditional probability of the outcome of eachcausePosterior likelihood priornormalizing constantFigure 4.Objects classified as RED and GREENHere the objects can be classified as either GREEN orRED. Our task is to classify new cases as they arrive, i.e.,decide to which class label they belong, based on thecurrently exiting objects. Since there are twice as manyGREEN objects as RED, it is reasonable to believe that anew case (which hasn't been observed yet) is twice as likelyto have membership GREEN rather than RED. In theBayesian analysis, this belief is known as the priorprobability. Prior probabilities are based on previousexperience, in this case the percentage of GREEN andRED objects, and often used to predict outcomes beforethey actually happen.Thus, we can write:Prior probability for GREEN α Number ofGREEN objectsTotal number of objects(1)Prior probability for RED αNumber ofRED objectsTotal number of objects(2)127IJSTR 2012www.ijstr.org
International Journal of Scientific & Technology Research Volume 1, Issue 4, May 2012Since there is a total of 60 objects, 40 of which are GREENand 20 RED, our prior probabilities for class membershipare:Prior probability for GREENα4060ISSN 2277-8616centroid and we provide some concave points as points ofseparation .Then we transform edges from a Cartesian to apolar space ,and we interpolate discontinuous points usinga linear interpolation this allows completing cell borers witha conical shape .Finally we join some edges discontinuitiesby applying morphological operations.(3)Prior probability for REDα2060(4)Having formulated our prior probability, we are now ready toclassify a new object (WHITE circle). Since the objects arewell clustered, it is reasonable to assume that the moreGREEN (or RED) objects in the vicinity of X, the more likelythat the new cases belong to that particular color. Tomeasure this likelihood, we draw a circle around X whichencompasses a number (to be chosen a priori) of pointsirrespective of their class labels. Then we calculate thenumber of points in the circle belonging to each class label.From this we calculate the likelihood:Likelihood of X given GREEN α Number of GREEN inVicinity of XTotal number of GREEN ases(5)Likelihood of X given RED αNumber of RED inVicinity of XTotal number of RED cases(6)6. Morphological SeparationOverlapped and cluttered cells are an inevitable, unsolved,and usually ignored problem in blood slideanalysis. It is upto the technician to chose an ideal work area in the smearwhere the cells are neither too cluttered nor to disperse. Inthe more disperse area the cells extend due to the lack ofpressure and lose characteristic morphology and in thecluttered area they are indistinguishable one from the other.have proposed automated criteria for the choosing of anideal area. Our approach has been to use the morphologyof the background-cell border as an initial approach to thecells forms, using a priori knowledge of the cell. We latermake use of local information, such as edges or greyscaleconnectivity, in a top down segmentation scheme to refinethe classifying and find cells deeper down the cluster. Thewatershed algorithm has been widely used as it subdividesthe image in catchment basins and clusters together pixelsbased on spatial proximity and similarity of the gradient.7. Seperation of overlapped blood cellsFigure 5 summarizes the process of cell separation oncethe overlapped region is identified. in order to split theoverlapped regions we obtain the edges of the region an itsFigure 5: Cell Seperation ProcedureIII. DATA DESCRIPTIONWe used the gray-scale bone marrow images The imageswere taken from a slide of a patient’s bone marrow smear,without any information about his/her health condition, byan Olympus BX50 microscope, a Sony B/W charge-coupleddevice (CCD) camera, and an 8-bit digitizer (PDI IMAXX.)Magnification of 600 was used without any special filters.Each white blood cell image was cropped manually to forma single-cell image. Then, a single-cell image wassegmented manually into nucleus, cytoplasm, andbackground region. The data set consists of six classes ofwhite blood cells—myeloblast, promyelocyte, myelocyte,metamyelocyte, band, and PMN.IV. EXPERIMENTAL FRAMEWORKFour features are extracted from each cell’s nucleus. Thefeatures are used as the inputs to two types of classifiers,i.e., a Bayes classifier and an artificial neural networkclassifier.Afivefold cross validation is applied to let usperform the training and testing on the data set. Theclassification results are evaluated in terms of the traditionalclassification rate and the classwise classification rate. Inthis section, we describe the nucleus feature extraction, theclassification performance evaluation, and the experimentalresults and analysis.1. Classwise Classification RatesIn a classification problem,we generally evaluate aclassifier’s performance using the traditional classificationrate, which is the ratio of the total correct classifications tothe total number of samples classified. In addition to thetraditional classification rate calculation, we consideranother rate called the classwise classification rate.Basically, the classwise classification rate is the average ofthe classification rates of all classes, i.e.,(7)Where C is the number of classes. The basic idea of theclasswise rate is to take out the effects of the number ofsamples in the training. While the traditional classificationrate may be high if a large number of correct classificationsoccur in a class consisting of a large number of samples,the classwise rate is high only if all the classes have largenumbers of correct classifications compared to theircorresponding total number of samples. Therefore, we128IJSTR 2012www.ijstr.org
International Journal of Scientific & Technology Research Volume 1, Issue 4, May 2012prefer to have a classifier that provides good classificationperformance in both the traditional and class wise senses.2. Experimental MethodsBoth the Bayes classifier and artificial neural networkclassifier require supervised learning, i.e., training andtesting with known classified samples. From the datadescription, the available data set is not divided into trainingand test sets; however,we need to have training and testsets to train and test our classifiers to evaluate theirgeneralization properties. We, therefore, apply the crossvalidation method, which is a standard solution of theaforementioned limitation. The experiments are performedusing the fivefold cross validation method.2.1) Bayes’ Classifiers:We initially performed the cell classification using Bayesclassifiers due to their simpl
thus, makes cell classification in bone marrow a more difficult and also a complex problem Sanei and Lee achieved classification rates of more than 95% for mature cells in normal blood, more than 85% for immature cells in the blood, but only just over 70% for immature cells in bone marrow.
BM Bone marrow BMAT Bone marrow aspiration and trephine BMB Bone marrow biopsy BMD Bone Mineral Density BMH Benign monoclonal hypergammaglobulinaemia BMI Body mass index BMPR1A Bone morphogenetic protein receptor, type 1A gene BMR Basal metabolic rate BMS Bare metal stent BMT Bone marrow transplant
bones. By the age of 20 years most of the appendicular skeleton contains fatty bone marrow, while the central skeleton including proximal femur and humerus contain largely hematopoietic bone marrow. In the 6th decade of life a substantial amount of fatty bone marrow is also found in the axial skeleton. Please note also that reconversion of fatty to
Keywords: Benign bone tumors of lower extremity, Bone defect reconstruction, Bone marrow mesenchymal stem cell, Rapid screening-enrichment-composite system Background Bone tumors occur in the bone or its associated tissues with a 0.01% incidence in the population. The incidence ratio among benign bone tumors, malignant bone tu-
bone vs. cortical bone and cancellous bone) in a rabbit segmental defect model. Overall, 15-mm segmental defects in the left and right radiuses were created in 36 New Zealand . bone healing score, bone volume fraction, bone mineral density, and residual bone area at 4, 8, and 12 weeks post-implantation .
The compact bone is the dense and hard part of the long bone. The spongy bone is the tissue filled cavity of the bone which is comparatively less hard and contains the red bone marrow. The gross structure of the long bone consists of many parts; proximal and distal epiphysis, the spongy bone and the diaphysis consisting of the medullary cavity, endosteum, periosteum and the
What is the difference between compact bone and spongy bone? A They have different bone marrow B They are made of different materials C They have different sized cells D They have a different arrangement of bone cells Question 6 Refer to the table below. Bone X refers to the Type of bone Example Longbone Humerus Flat bone X
CaseReportsinSurgery (a) (b) (c) (d) F : Bone-marrow-derived mononuclear cells preparation: (a) the bone marrow
The Academic Phrasebank is a general resource for academic writers. It aims to provide the phraseological ‘nuts and bolts’ of academic writing organised according to the main sections of a research paper or dissertation. Other phrases are listed under the more general communicative functions of academic writing.
