MCs Detection Approach Using Bagging and Boosting Based Twin Support Vector Machine

Transcription

1 Proceedngs of he 009 IEEE Inernaonal Conference on Sysems, Man, and Cybernecs San Anono, TX, USA - Ocober 009 MCs Deecon Approach Usng Baggng and Boosng Based Twn Suppor Vecor Machne Xnsheng Zhang School of managemen, X an Unv of Arch and Tech School of Elecronc Engneerng, Xdan Unversy X an, Chna xnshengzh@homalcom Xnbo Gao School of Elecronc Engneerng Xdan Unversy X an, Chna xbgao@lab0xdaneducn Mnghu Wang School of Auomac Conrol Norhwesern Polyechncal Unversy X an, Chna wangmnghuemal@63com Absrac In hs paper we dscuss a new approach for he deecon of clusered mcrocalcfcaons (MCs) n mammograms MCs are an mporan early sgn of breas cancer n women Ther accurae deecon s a key ssue n compuer aded deecon scheme To mprove he performance of deecon, we propose a Baggng and Boosng based wn suppor vecor machne (BB-TWSVM) o deec MCs The algorhm s composed of hree modules: he mage proprocessng, he feaure exracon componen and he BB- TWSVM module The ground ruh of MCs n mammograms s assumed o be known as a pror Frs each MCs s preprocessed by usng a smple arfac removal fler and a well desgned hgh-pass fler Then he combned mage feaure exracors are employed o exrac 64 mage feaures In he combned mage feaures space, he MCs deecon procedure s formulaed as a supervsed learnng and classfcaon problem, and he raned BB-TWSVM s used as a classfer o make decson for he presence of MCs or no The expermenal resuls of hs sudy ndcae he poenal of he approach for compuer-aded deecon of breas cancer Keywords enseble learnng, Boosng, feaure exracon, deecon, clusered mcrocalcfcaons, Baggng, wn suppor vecor machne I INTRODUCTION Bres cancer s he mos frequenly dagnosed cancer n women[, ] Dgal mammography s, a presen, one of he mos suable mehods for early deecon of breas cancer[3] One of he mporan early sgns of breas cancer s he appearance of clusered mcrocalcfcaons (MCs), whch appear n 30% -50% of mammographcally dagnosed cases wh ny brgh spos of dfferen morphology Because of s mporance n early breas cancer dagnoss, accurae deecon of MCs has become a key problem Recenly, a lo of dfferen approaches based on machne learnng have been appled o he deecon of MCs [4] Concernng mage segmenaon and specfcaon of regons of neres(roi), several mehods have been revewed n [5] Formulaed MCs deecon as a classfcaon problem, varous machne learnng mehodologes have been proposed for he characerzaon of MCs, such as, fuzzy rule-based sysems [6], suppor vecor machnes [7, 8], neural neworks [7, 9, 0], ec These mehods learn hypoheses from a large amoun of dagnosed samples, e, he daa colleced from a number of necessary medcal examnaons along wh he correspondng dagnoses made by medcal expers To make hese approaches work well, a large amoun of posve (wh suspcous regon) and negave (wh normal ssue) samples wh dagnoss are needed for ranng he learnng model Usually, hese negave samples can be easly colleced and he number s also very large, whch s always enough for ranng any learnng model However, he posve samples are ofen very small I s because ha makng a dagnoss for such a large amoun of cases one by one places a very heavy burden on medcal expers To solve he problem, one possble soluon s o learn a hypohess from he lmed number of he posve and negave samples In machne learnng, hs problem s called learnng wh unbalanced daa Recenly Jayadeva e al [] proposed a nonparallel plane classfers for bnary daa classfcaon, ermed as wn suppor vecor machne (TWSVM) I ams a generang wo nonparallel planes, whch use wo dfferen group of samples, such ha each plane s closer o one of he wo classes and s as far as possble from he oher TWSVM solves a par of quadrac programmng problems (QPPs) The wo QPPs have a smaller sze han he one larger QPP, whch makes TWSVM work faser han he sandard suppor vecor machne(svm) And also TWSVM has he poenal ably o deal wh he unbalanced daa ses by solvng he wo dfferen smaller sze problems Bu he same as SVM, TWSVM s unsable, because s sensve o he ranng samples To elmnae he underlyng weakness of TWSVM, we desgn a novel mechansm for MCs deecon -- Baggng and Boosng based wn suppor vecor machne (BB-TWSVM), whch /09/$ IEEE /09/$ IEEE 545

2 ncorporaes Baggng and Boosng wh TWSVM The key dea comes from he ensemble learnng or mulple classfer sysem [-4] Ensemble mehods, well known n he machne learnng communy, are ypcally composed of mulple mehods comprsng dfferen classfcaon sraeges or dfferen classfers wh a unfed objecve funcon The fnal predcons or decsons are chosen from he ensemble of mehods by a learnng rule, whch may be as smple as fndng he maxmum score from all he base learners, or as complex as opmzng a weghed scorng scheme from among he base mehods The consrucon of hs learnng rule s a key problem o he performance of an ensemble learnng mehod, as he performance of an ensemble mehod wh an neffecve learnng rule wll be he average of he performance of s componen algorhms To mprove he performance of he avalable MCs deecon algorhms, hs paper employs he ensemble learnng mehod wh Baggng and Boosng based wn suppor vecor machne (BB-TWSVM) as a fnal decson model o dsngush he MCs from he oher ROIs (regon of neress) BB-TWSVM s frs raned by he real mammograms from DDSM Then s used o deec oher ROIs Compared wh sngle learnng algorhm and oher exsng ensemble mehods, he proposed approach yelds superor performance when evaluaed usng recever operang characersc (ROC) curves I acheved average sensvy as hgh as 9409% wh abou 63% average false-posve rae n each esng phrase The res of he paper s organzed as follows The proposed BB-TWSVM algorhm s formulaed n Secon II Secon III repors he feaure exracon mehods The proposed MCs deecon algorhm based on BB-TWSVM s formulaed n Secon IV and he expermenal resuls are gven n Secon V Fnally, conclusons are drawn n Secon VI II BOOSTING AND BAGGING BASED TWSVM A Ensemble Learnng An ensemble consss of a se of ndvdually raned classfers, such as neural neworks, suppor vecor machnes, decson rees ec, whose predcons are combned when classfyng he npus [] Ensemble learnng combnes mulple learned models or classfers under he assumpon ha "wo (or more) heads are beer han one" Prevous research has shown ha an ensemble s ofen more accurae han any of he sngle classfers n he ensemble An ensemble combnes dfferen base learnng algorhms or he same learnng algorhms raned n he same or dfferen ways As we know, an ensemble s a group of classfers ha jonly solve a classfcaon problem Boh heorecal and emprcal fndngs suppor ha an ensemble wll ofen ouperform a sngle classfer In addon, he ensemble learnng framework provdes ools o ackle complcaed or large problems ha were once nfeasble for radonal algorhms In vew of he advanages of an ensemble agans a sngle classfer, many ensemble generaon algorhms have been presened durng he pas decade o sysemacally consruc collecve classfers ha as a whole can acheve beer performance han a sngle classfer Baggng and Boosng are wo mos famous algorhms o consruc an ensemble classfer B Twn Suppor Vecor Machne Twn suppor vecor machne (TWSVM) s a recenly developed and very effecve bnary classfcaon algorhm, whch obans nonparallel planes around whch he daa pons of he correspondng class ge clusered Each of he wo quadrac programmng problems n a TWSVM par has he formulaon of a ypcal SVM, excep ha no all paerns appear wh consrans of eher problem a he same me The TWSVM classfer s obaned by solvng he followng par of quadrac programmng problems: () () T () () T (TWSVM) mn ( Aw + e () () b ) ( Aw + eb ) + ce q, b,, and () w q s -( Bw + e b ) + q e, q 0 () () () () T () () T (TWSVM) mn ( Bw + e () () b ) ( Bw + eb ) + ceq w, b, q s ( Aw + e b ) + q e, q 0, () () where c, c > 0 are parameers and e and e are vecors of ones of approprae dmensons Ths approach ams o fnd wo opmal hyperplanes, one for each class, and classfers pons accordng o whch hyperplane a gven pon s closes o The frs erm n he objecve funcon of () or () s he sum of squared dsances from he hyperplane o pons of one class Therefore, mnmzng ends o keep he hyperplane close o pons of one class (e, class ) The consrans requre he hyperplane o be a a dsance of a leas from pons of he oher class (e, class -); a se of error varables s used o measure he error wherever he hyperplane s closer han hs mnmum dsance of The second erm of he objecve funcon mnmzes he sum of error varables, hus aempng o mnmze msclassfcaon due o pons belongng o class - As an example n Fg, one can fnd he dfference beween radonal SVM and TWSVM Smlarly as SVM, TWSVM can be easly exended o nonlnear kernel verson C Baggng Twn Suppor Vecor Machne Baggng s a boosrap ensemble mehod ha creaes ndvduals for s ensemble by ranng each classfer on a random redsrbuon of he ranng se I ncorporaes he benefs of boosrappng and aggregaon The fnal mulple classfers can be generaed by ranng on mulple ses of samples ha are produced by boosrappng, e, random samplng wh replacemen on he ranng samples Aggregaon of he generaed classfers can hen be mplemened by majory vong or weghed averagng Expermenal and heorecal resuls have shown ha Baggng can mprove a good bu unsable classfer sgnfcanly Ths s exacly he problem of TWSVM However, drecly usng Baggng n TWSVM s no approprae snce we have only a very small number of samples To overcome hs problem, we develop a novel Baggng sraegy The boosrappng s execued only on he negave samples snce here are far more negave samples han he posve samples Ths way each generaed classfer () 546

3 wll be raned on a balanced number of posve and negave samples The Baggng TWSVM algorhm s descrbed n Table I In he algorhm, he aggregaon s mplemened by weghed averagng TABLE I ALGORITHM OF BAGGING TWIN SUPPORT VECTOR MACHINE Inpu: posve ranng se S +, negave ranng se S, weak classfer I (TWSVM), neger T (number of generaed classfers), and he es sample X ( =,, m) for = o T { S = boosrap samples from S +, where S = S + C = I( S, S ) + S w = C ( S, ) } * + C ( X ) = aggregaon{ C ( X, S, S ), T} Oupu: fnal classfer * C D Boosng Twn Suppor Vecor Machne Boosng s a machne learnng mea-algorhm for performng supervsed learnng Boosng s based on he queson posed by Kearns[5]: can a se of weak learners creae a sngle srong learner? A weak learner s defned o be a classfer whch s only slghly correlaed wh he rue classfcaon In conras, a srong learner s a classfer ha s arbrarly well-correlaed wh he rue classfcaon Whle Boosng s no algorhmcally consraned, mos Boosng algorhms conss of eravely learnng weak classfers wh respec o a dsrbuon and addng hem o a fnal srong classfer When hey are added, hey are ypcally weghed n some way ha s usually relaed o he weak learner's accuracy Afer a weak learner s added, he daa s reweghed: examples ha are msclassfed gan wegh and examples ha are classfed correcly lose wegh (some Boosng algorhms acually decrease he wegh of repeaedly msclassfed examples, eg, boos by majory and BrownBoos ) Thus, fuure weak learners focus more on he examples ha prevous weak learners msclassfed Expermenal and heorecal resuls have shown ha Boosng can mprove a good bu unsable classfer sgnfcanly Ths s exacly he problem of TWSVM However, drecly usng Boosng n TWSVM s no approprae snce we have only a very small number of samples To overcome hs problem, we develop a novel Baggng sraegy The boosrappng s execued only on he negave samples snce here are far more negave samples han he posve samples Ths way each generaed classfer wll be raned on a balanced number of posve and negave samples The Boosng TWSVM algorhm s descrbed n Table II E Baggng and Boosng based TWSVM In order o combne he advanages of he above wo algorhms, we formulae hem no one approach-- Baggng and Boosng based TWSVMTo In shor, he Baggng and Boosng based TWSVM algorhm performs a boosrap aggregaon on he boosed weak learner (TWSVM) The parameer selecon s fundamenal: T s he chosen number of he boosrap replcaes, M s he number of eraons of he Boosng algorhm The sub-sample consss of a fracon ρ of he ranng se Noce ha we do no allow Boosng o run for hundreds of eraons Ths s a fundamenal feaure of our algorhm and arses from he observaon ha Boosng s a self-smoohng algorhm We shall show ha runnng Boosng for many eraons, well pas he sample of achevng zero ranng se error rae, allows he algorhm o smooh self Snce he smoohng n hs algorhm s accomplshed drecly by Baggng he large number of eraons are no needed and can be lmed o a fxed number TABLE II ALGORITHM OF BOOSTING TWIN SUPPORT VECTOR MACHINE Inpu: posve ranng se S +, negave ranng se S, weak classfer I (TWSVM), neger T (number of generaed classfers), he es sample X ( =,, N), dsrbuon D over he N examples Inalze he wegh vecor: w = D() for =,, N Do for =,,, T Se w p = N = w Call Weak classfer, provdng wh he dsrbuon p ; ge back a hypohess h : Χ [ 0,] 3 Calculae he error of h : ε N = = p h ( x ) y 4 Se β = ε /( ε) 5 Se he new weghs vecor o be + h ( x) y w wβ = Oupu he hypohess * f T (log/ ) h ( x) T log/ ( ) β C X β = = = 0 oherwse TABLE III ALGORITHM OF BOOSTING AMD BAGGING BASED TWIN SUPPORT VECTOR MACHINE Inpu: posve ranng se S +, negave ranng se S + ( S = S S, S = N ), weak classfer I (TWSVM), (Baggng) neger T (number of generaed classfers) and ρ > 0, (Boosng) parameer M Do for =,,, T Generae a replcae ranng se S ' from S + and S of sze ' Sk = ρ N by sub-samplng wh replacemen Generae a boosed classfer, C * ( X ), based on S ' usng M eraons of he base learner * Oupu he aggregae classfer C : III * C ( X) = arg max y Y * C : ( X) FEATURE EXTRACTION The research mehodology uses a novel Baggng and Boosng based wn suppor vecor machne o deec MCs wh a se of combned mage feaures from a number of exsng feaures As we know, feaure exracon s a very mporan par for he classfcaon problem To ge he bes feaure or combnaon of feaures and ge he hgh classfcaon rae for MCs deecon s one of man ams of he proposed research In our ask, a se of 64 feaures, shown n 547

4 Table IV, s calculaed for each suspcous area (ROI) from he exural, spaal and ransform domans n our research Before ranng he classfer, we use he feaure exracor dscussed n Table IV o exrac MCs feaures n he feaure doman The 64 feaures of each block are exraced as follows Inensy hsogram based exure feaure A frequenly used approach for exure analyss s based on sascal properes of he nensy hsogram One class of such measures s based on sascal momens The expresson for he nh momen abou he mean s gven by L ( ) n ( ) n z 0 m p z (3) where z s a random varable ndcang nensy, p( z) s he hsogram of he nensy levels n a regon, L s he number of possble nensy levels, and L m z p( z ) (4) 0 s he mean(average)nensy Table V lss he descrpors used n our expermens based on sascal momens and also on unformy and enropy TABLE IV COMBINED DIFFERENT KIND OF IMAGE FEATURES USED IN OUR EXPERIMENTS Feaure groups Type #No Mean Sandard devaon Hsogram based exure Smoohness 3 feaures Thrd momen 4 Unformy 5 Enropy 6 Mul-scale hsogram feaures Zernke Momens Feaures[6] Combned frs 4 momens feaures Tamura exure sgnaures Chebyshev ransform hsogram feaure [7] Radon Transform Feaures TABLE V Hsograms wh dfferen number of bns 3,5,7,9 Zernke Momens Combned momens feaure 7~30(oal 4) 3~66 (oal 36, D=0) 67~4 (oal 48) Tamura feaures 5~0 (oal 6) Chebyshev hsogram ~5 (oal 3) Radon feaures 53~64 (oal ) DESCRIPTORS OF THE TEXTURE BASED ON THE INTENSITY HISTOGRAM OF A REGION Momen Formula #No Mean L 0 m z p( z ) Sandard devaon () z Smoohness R /( ) 3 L 3 Thrd momen ( ) ( ) 3 z 0 m p z 4 Unformy Enropy L 0 L 0 U p ( z) 5 e p( z )log p( z ) 6 Mulscale nensy hsograms We compue sgnaures based on "mul-scale hsograms" dea Idea of mul-scale hsogram comes from he belef of a unque represenaon of an mage hrough nfne seres of hsograms wh sequenally ncreasng number of bns Here we used 4 hsograms wh number of bns beng 3,5,7,9 o calculae he mage feaure, so we ge a *4 row vecors Zernke momens We use derved momens based on alernave complex polynomal funcons, know as Zernke polynomals[6] They form a complee orhogonal se over he neror of he un crcle x y and are defned as Z ( p)/ f( x, yv ) (, ) dxdy, (5) pq x y V (, x y) V (,) R ()exp( jq), (6) pq pq pq ( pq )/ s ps ( )[( ps)!] Rpq( ), (7) s 0 p q p q s! s! s! where p s a nonnegave neger, q s an neger subjec o he consran p q even and q p, x y s he radus from (, xy ) o he mage cenrod, an ( y/ x) s he angle beween and x-axs The Zernke momen s Z pq order p wh repeon q For a dgal mage, he respecve Zernke momens are compued as Z pq ( p )/ f ( x, y) V(, ) dxdy, x y, (8) where runs over all he mage pxels Zernke momens are used as he feaure exracor where by he order s vared o acheve he opmal classfcaon performance Combned frs four momen feaures Sgnaures on he bass of frs four momens (also known as mean, sd, skewness, kuross) for daa generaed by vercal, horzonal, dagonal and alernave dagonal 'combs' Each column of he comb resuls n 4 scalars [mean, sd, skewness, kuross], we have as many of hose [] as 0 So, 0 go o a 3-bn hsogram, producng x48 vecors Tamura exure sgnaures Tamura e al [7] ook he approach of devsng exure feaures ha correspond o human vsual percepon Sx exural feaures: coarseness, conras, dreconaly, lnelkeness, regulary and roughness, are defned for he mage feaure for objec recognon Transform doman feaures We compues sgnaures (3 bns hsogram) from coeffcens of D Chebyshev ransform (0h order s used n our expermens) Also we used sgnaures based on he Radon ransform as a knd of mage feaures Radon ransform s he projecon of he mage nensy along a radal lne (a a specfed orenaon angle), oal 4 orenaons are aken Transformaon n/ vecors (for each roaon), go hrough 3- bn hsogram and convolve no x vecors IV BAGGING AND BOSSTING BASEDTWSVM FOR MCS DETECTION For a gven dgal mammography mage, we consder he MCs deecon process as he followng seps: Sep Preprocess he mammography mage by removng he arfacs, suppressng he nhomogeney of he background and enhancng he mcrocalcfcaons Sep A each pxel locaon n he mage, exrac a 548

5 A m small wndow o descrbe s surroundng mage m feaure Sep 3 Apply feaure exracon mehods o ge he combned feaure vecor x Sep 4 Use he raned BB-TWSVM classfer o make decson wheher x belongs o MCs class (+) or no ( ) A Mammogram Preprocessng Frs, a smple flm-arfac removal fler s appled o remove he flm-arfacs from mammography mage Afer he arfacs were removed, our ask s o suppress he mammogram background A hgh-pass fler s desgned o preprocess each mammogram before exracng samples Wh he Gaussan fler, f (, xy) exp( ( x y)/( )), we use a m m wndow sze Gaussan fler where m 4, expermenally n he sudy The oupu of hgh-pass fler s denoed by I (, x y) I (, x y) f(, x y) I (, x y), where s lnear convoluon B Inpu Paerns for MCs deecon Afer he mammogram preprocessng sage, we exrac combned mage feaures from A m m as a feaure vecor x Am m s a small wndow of m m pxels cenered a a locaon ha we concerned n a mammogram mage The choce of m should be large enough o nclude he MCs (n our expermen, we ake m =5) The ask of he TWSVM classfer s o decde wheher he npu wndow Am m a each locaon s a MCs paern ( y ) or no ( y ) C Tranng Daa Se Preparaon The procedure for exracng ranng daa from he ranng mammograms s gven as follows For each MCs locaon n a ranng mammogram se, a wndow of m m mage pxels cenered a s cener of mass s exraced; he area s denoed by A m m, wh respec o x afer subspace feaure exracon, and hen x s reaed as an npu paern for he posve sample ( y ) The negave samples are colleced ( y ) smlarly, excep ha her locaons are randomly seleced from he non MCs locaons n he ranng mammograms In he procedure, no wndow n he ranng se s allowed o overlap wh any oher ranng wndow V EXPERIMENTAL RESULTS A Combned Feaure Exracon The daa n he ranng (and valdaon), es, and valdaon ses were randomly seleced from he ranng se Each seleced sample was covered by a 5x5 wndow whose cener concded wh he cener of mass of he suspeced MCs The blocks ncluded 65 wh rue MCs and 3567 wh normal ssue Before ranng he classfer, we frs use he feaure exracon algorhm o exrac MCs feaure n combned feaure doman The 64-dmenson feaure vecor wll be calculaed for each mage block When we ge he mage feaure vecor, feaure normalzaon program should be user for normalzng he feaures o be real numbers n he range of [0, ] The normalzaon s accomplshed by he followng sep: (a) Change all he feaures o be posve by addng he magnude of he larges mnus value of hs feaure mes 0 ;(b) Dvde all he feaures by he maxmum value of he same feaure The normalzed feaures are used as he npus of he proposed ensemble learnng algorhm for ranng and classfcaon B BB-TWSVM Model Tranng and Selecon For base model selecon, he TWSVM classfer s frs raned by usng he 0-fold cross-valdaon procedure wh dfferen model and paramerc sengs In our ranng sage, we used generalzaon error, whch was defned as he oal number of ncorrecly classfed examples dvded by he oal number of samples classfed, as a merc o measure he raned classfer Generalzaon error was compued usng only he samples used durng ranng For he sake of convenen, he paramerc values of c and c n our expermen are se o be equal (e c c ) In Fgure (a), we summarze he resuls for raned TWSVM classfer wh a polynomal kernel The esmaed generalzaon error s ploed versus he regularzaon parameers c and c for kernel order p, p 3 and p 4 Smlarly, Fgure (b) summarzes he resuls wh RBF kernel Here, he generalzaon error s ploed for dfferen values of he wdh (, 7, 0, 5, 7, 0) Average Generalzaon error Average Generalzaon error Regularzaon parameer C (a) Regularzaon parameer C p= p=3 p=4 σ= σ=7 σ=0 σ=5 σ=7 σ=0 (b) Fgure Plo of average generalzaon error rae versus regularzaon parameer C ( c = c ) acheved by ranng TWSVM usng (a) a polynomal kernel wh orders, 3, and 4, and (b) a Gaussan RBF kernel wdh σ =, 7, 0, 5, 7, 0 by 0-fold valdaon mehod 549

6 TABLE VI EXPERIMENTAL RESULTS OF BB-TWSVM AND TWSVM CLASSIFIER Mehod Sensvy Specfcy Az TWSVM 900% 9037% BB-TWSVM 9409% 9379% 0956 From Fgure, can be shown ha, he BB-TWSVM algorhm has a hgher deecon accuracy rae compared o TWSVM wh he same confguraon By usng he same exraced feaure, compared wh TWSVM, BB-TWSVM has a beer deecon performance when we ran he classfer In parcular, he BB-TWSVM classfer acheved he averaged sensvy of approxmaely 9409% wh respec o 63% false posve rae and Az=0956 Wh he same ranng daa se and es daa se, he TWSVM classfer acheved a sensvy of 900%, 963% false posve rae and Az=09459 Sensvy TWSVM BB-TWSVM False posve rae Fgure Roc curves of MCS deecon usng BB-TWSVM and TWSVM classfer, where Az (BB-TWSVM)= 0967 and Az (TWSVM)= VI CONCLUSION In hs paper, we proposed a Baggng and Boosng based TWSVM (BB-TWSVM) echnque for deecon MCs n dgal mammograms In hs mehod, combned mage feaures are exraced from each mage block of posve and negave samples, and BB-TWSVM s raned hrough supervsed learnng wh he 64 dmensonal feaure o es a every locaon n a mammogram wheher an MCs s presen or no The decson funcon of he raned BB-TWSVM s deermned n erms of ensemble classfer model ha are modeled from he posve and negave samples durng ranng sage Compared o TWSVM classfer, BB-TWSVM can solve he unsable problem of TWSVM whle mananng he deecon accuracy n a nosy envronmen In our expermens, ROC curves ndcaed ha he BB-TWSVM learnng approach yelded he beer performance compared wh TWSVM ACKNOWLEDGMENTS Ths work presened n our paper s suppored by he Naonal Naural Scence Foundaon of Chna (No ), he Naural Scence Basc Research Plan n Shaanx Provnce of Chna (No 007F 48), he Scenfc Research Plan Projecs of Shaanx Educaon Deparmen (No 09JK563 and No 09JK54) The auhors are graeful for he anonymous revewers who made consrucve commens REFERENCES [] Amercan Cancer Socey, Breas Cancer Facs & Fgures , Amercan Cancer Socey, Inc, Alana, 007 [] A Jemal, R Segel, E Ward e al, Cancer Sascs, 006, CA: A Cancer Journal for Clncans, vol 56, no, pp 06-30, 006 [3] J Dengler, S Behrens, and J Desaga, Segmenaon of mcrocalcfcaons n mammograms, IEEE Transacons on Medcal Imagng, vol, no 4, pp , 993 [4] L We, Y Yang, R Nshkawa e al, A sudy on several Machne-learnng mehods for classfcaon of Malgnan and bengn clusered mcrocalcfcaons, IEEE Transacons on Medcal Imagng, vol 4, no 3, pp , 005 [5] H Cheng, X Ca, X Chen e al, Compuer-aded deecon and classfcaon of mcrocalcfcaons n mammograms: a survey, Paern Recognon, vol 36, no, pp , 003 [6] R Mousa, Q Munb, and A Moussa, Breas cancer dagnoss sysem based on wavele analyss and fuzzy-neural, Exper Sysems wh Applcaons, vol 8, no 4, pp 73-73, 005 [7] A Papadopoulos, D I Foads, and A Lkas, Characerzaon of clusered mcrocalcfcaons n dgzed mammograms usng neural neworks and suppor vecor machnes, Arfcal Inellgence n Medcne, vol 34, no, pp 4-50, 005 [8] I El-Naqa, Y Yongy, M N Wernck e al, A suppor vecor machne approach for deecon of mcrocalcfcaons, IEEE Transacons on Medcal Imagng, vol, no, pp , 00 [9] S Halkos, T Boss, and M Rangouss, Auomac deecon of clusered mcrocalcfcaons n dgal mammograms usng mahemacal morphology and neural neworks, Sgnal Processng, vol 87, no 7, pp , 007 [0] L Bocch, G Coppn, J Nor e al, Deecon of sngle and clusered mcrocalcfcaons n mammograms usng fracals models and neural neworks, Medcal Engneerng & Physcs, vol 6, no 4, pp 303-3, 004 [] R Jayadeva, and S Chandra, Twn Suppor Vecor Machnes for Paern Classfcaon, IEEE Transacons on Paern Analyss and Machne Inellgence, vol 9, no 5, pp , 007 [] Z Zh-Hua, and Y Yang, Ensemblng local learners Through Mulmodal perurbaon, IEEE Transacons on Sysems, Man, and Cybernecs, Par B,vol 35, no 4, pp , 005 [3] G I Webb, and Z Zheng, Mulsraegy ensemble learnng: reducng error by combnng ensemble learnng echnques, IEEE Transacons on Knowledge and Daa Engneerng, vol 6, no 8, pp , 004 [4] N Ueda, Opmal lnear combnaon of neural neworks for mprovng classfcaon performance, IEEE Transacons on Paern Analyss and Machne Inellgence, vol, no, pp 07-5, 000 [5] R Schapre, and Y Freund, A decson heorec generalzaon of on-lne learnng and an applcaon o Boosng, J Compu Sys Sc, vol 55, pp 9-39, 997 [6] L Tsong-Wuu, and C Yun-Feng, "A comparave sudy of Zernke momens" n Proceedng of WI003, 003 pp [7] M S Tamura H, Yamawak T, Texure feaures corresponsng o vsual percepon, IEEE Transacons on Sysems, Man and Cybernecs,, vol, no 8, pp ,