Using Biased-Output Neural Networks - CiteSeerX

6 downloads 0 Views 679KB Size Report
(d) Metamyelocyte; () Band; and (0 PMN. 2.1.1 Morphological Operations. Morphological operations are non-linear, translation invariant transformations.
2003 ThammasatInt. J. Sc. Tech.,Vol. 8, No. l, January-March

Automatic White Blood Cell Classification

NeuralNetworks UsingBiased-Output with MorphologicalFeatures Nipon Theera-Umpon of ElectricalEngineering, Department ChiangMai University, Facultyof Engineering, ChiangMai 50200Thailand E-mail: [email protected] Abstract Numbersof white blood cells in differentclasseshelp doctorsto diagnosepatients. A new set of featuresbasedon the mathematicalmorphologyin the white blood cell classificationproblem is proposedin this paper. The proposedfeaturesare the maximum value of a pattern spectrum,the iocationwhere the maximumvalue of a patternspectrumoccurs,the first and secondgranulometric moments.We also proposea methodto unbiasneuralnetworksby biasingthe desiredoutputusinga priori informationof the numberof samplesin eachclass. Regularartificial neuralnetworksand the usingthe five-foldcrossvalidationasthe neuralnetworksareappliedin the experiments biased-output neural usingour biased-output performances classifiers good of the results show testingmethod. The features. networksand our proposedmorphology-based Keywords: Automatic white blood cell classification,Featureextraction,Biased-outputneural network, Morphologicalfeatures,Granulometricmoments

ally would requirea trainedexpert. Moreover, it is a verytediousjob. White blood cells in bone marrow are classifiedaccordingto their maturationstages. When a white blood cell becomesolder, its size, nuclei shape and many other features change. White blood cells in the myelocl4ic seriescan be classifiedinto six classes,i.e., myeloblast, promyelocyte,myelocyte, metamyelocyte, band, and polymorphonuclear (PtvfN)orderedfrom the youngestto the oldest cells [1,2]. Figure I showssamplesof white bloodcellsin this series. Although there are some commercial automaticsystemsavailablefor countingwhite blood,thereis no such bloodcellsin peripheral system for bone marrow. There have been severalattemptsproposedto solve the problem. Most methods follow the traditional manualmaneuver,i.e., to detecta cell, extract its features,classifythe cell, and then update the count [3-7]. Some are based on neural

1. Introduction There are two methods to count white blood cells in bone marrow namely the total countand the differential counts. The total countis the total numberof white bloodcells without any classification. In contrast,the differentialcountsare the countsof different cell classesin bone marrow. One important stepto achievethe whitebloodcell differential countingis to classi0'eachwhite bloodcell in bonemarrow. The differentialcountingprovidesinvaluable informationto doctors in diagnosisof diseasessuch as leukemiaor cancers. The traditionalmethodfor an expertto achievethe differential counting is to look through a microscopeto select an areaof interestin a bone marrow slide, detecta white blood cell, classifo it based on his knowledge,and increase the count of the correspondingcell manuclass.Conductingall oftheseprocesses

A part of the content has appearedand been discussedat the 25'h Electrical EngineeringConference (EECON-25),on2l-22 Nov.2002heldat Princeof SongklaUniversity,Hat Yai, Songkla,Thailand. 64

ThammasatInt. J. Sc.Tech.,Vol. 8, No. 1, January-March 2003

networks with features such as area of cell, area of nuclei, ratio of area of nuclei to cytoplasm, Fourier descriptors of nuclei, some textural features, etc [7]. Some are developed under the mixing theoriesof the mathematicalmorphology [8 l0]. Some develop new training algorithm for neural networks in order to count numbersof different cell classes. without classification

l l r ,1 2 1 . In this paper, we apply artificial neural networksto the white blood cell classification for single-cellimagesunderthe assumption that the cell segmentation is available.We propose new featuresof a white blood cell basedon morphologicalgranulometries. We also apply the bias to the desiredoutputsof neuralnetworks basedon the prior informationof number o f c e l l si n e a c hc l a s si n a t r a i n i n gs e t . This paperis organizedas follows.Section 2 introduces the mathematical morphology, neural networks, and how to extract the proposed features.The data set is describedin section3. A descriptionof the experiments and their resultsare shown in section4. Section5 concludes this paper. 2. Methodology In this research, artificialneuralnetworksare usedas our classifiersin the six-classproblem. The input featuresare mainly extractedfrom patternspectraof nucleus.To be morespecific, there are six features- two are area-based, the remainingfour are morphology-based. 2.1 Mathematical Morphology Mathematical morphology is a branch of nonlinearimageprocessing andanalysis.It was first introducedby Matheron in the context of randomsets !3,141. Morphologicalmethods are usedin many ways in imageprocessing,for example, enhancement,segmentation,restoration, edge detection,texture analysis,shape a n a l y s i se,t c . [ 1 5 , 1 6 ] . I t i s a l s oa p p l i e dt o s e v eral researchareas,such as, medicalimaging, remotesensing, militaryapplications, etc.

n {

(a)

(b)

(c)

65

# i l ' r*, (d)

(e)

(0

Fig. l Cellsamples in themyelocytic series: (a)Myeloblast; (b) Promyelocyte; (c) Myelocyte; (d) Metamyelocyte; (e) Band;and(0 PMN.

2.1.1 MorphologicalOperations Morphological operations are non-linear, translationinvarianttransformations. This paper describes binarymorphological operations only. Binary imagescan be consideredas functionson gridswith valuesof 0 or 1 or, two-dimensional equivalently, as characteristic functions of subsetsof the two-dimensionalplane. The conceptof structuringan elementis fundamental in morphology;it is the analogueof a convolution maskin linearimageprocessing.The basic morphologicaloperationsinvolving an imageS and a structuringelementE are erosion'..SOt:n {S- e:e e E\ dilation:,SO ,8: w {E + s. s e S},

where n and v denotethe set intersectionand union, respectively.A * x denotesthe translat i o no f a s e tI b y a p o i n tx , i . e . A+r={a+x.aeA}

The closing and opening operations,derived from the erosionand dilation,are definedby closing:,SOt: (.SO (-E)) e (-t) opening:,SO,=(SOE)OE

where-E = 1*e: e e E) denotesthe 180orotat i o no f E a b o u t h eo r i g i n . The examplesof an image S, a structuring elementE, and outputsof the erosion,dilation, closing and opening operatorsare shown in Figure2.

2003 Int. J. Sc.Tech.,Vol. 8, No. 1, January-March Thammasat

ImageS

Dilation:S@E

o Structuring ElementE

The goal is to find the bestsetof weights(w) &reds closeto the desired so that the outputso1,n for a given input pattern possible as d,, outputs x , . n i,- - 1 , . . . ,P a n d T= 1 , . . . , Q . P a n dQ a r e the numberof input featuresand the numberof classes, respectively. There are many approachesusedto find the best set of weights. However, in this research, the LevenbergMarquardt(LM) algorithmwas chosenbecause it providesfasterconvergence [18].

2.3 Feature Extraction This study focusesits featureextractionon the morphology-basedfeatures. Hence, their Fig.2 Samplesof an imageS, structuringelement derivationsare introducedhere. For a random E, and outputs of the erosion,dilation, closing and set,S,O(l) is a randomfunction.The normalized openingoperators. size distribution