Temporal automatic edge detection of echocardiographic images

M7.8 TEMPORAL AUTOMATIC EDGE DETECTION OF ECHOCARDIOGRAPHIC IMAGES L. To es, A. Gasull Dep . o Signal Theo y and Communica ions E.T.S.I. Telecomunicaci6n - UPC Apdo. 30.002 08080 BARCELONA, SPAIN ABSTRACT This pape p esen s a new ully au oma ic algo i hm ha ex ac s le en icula epica dial and endoca dial bo de s om 2D echoca diog ams. The algo i hm s a s wi h he end sys ole and end dias ole con ou s ob ained h ough a adial analysis o he image. A e some i e a i e app oxima ions, excellen quali y con ou s o s a ic images a e ound. A his poin , in o de o ake ad an age o he in o ma ion con ained in he empo al domain, he echnique is applied o he whole ca diac cycle hus ob aining a comple e se ies o ca diac con ou s ha a e used in a u he s age o md he le en icula chambe olume and some o he ela ed pa ame e s. The algo i hm has been applied o di e en ca diac iews, such a pa as e nal sho axis, apical and mi al, wi h he same deg ee o success. I. INTRODUCTION Du ing he las i een yea s echoca diog aphy has expe ienced widesp ead accep ance in he e alua ion o ca diac disease. This is in la ge pa due o i s nonin asi e na u e and i s eal ime capabili y o obse ing bo h ca diac s uc u e and mo ion. Two-dimensional echoca diog aphy is capable o p o iding an in eg a ed o e all iew o bo h ca diac s uc u e and mo ion. Since bo h global and egional le en icula unc ion a e majo a iables used o de e mine p ognosis in ca diac disease, he e is a g ea deal o in e es in abili ies o quan i a e unc ion indexes om a ious ypes o ca diac imaging echniques. Fo ins ance, he analysis o a comple e cycle o ca diac images, allows o ind le en icula chambe olume and le en icula ejec ion ac ion. Compu e s ha e been used mainly o assis in inding ca diac pa ame e s. In his ole, hey ha e been adap ed in a simple ashion, i.e., manual bo de s a e de ined using some ype o en y sys em, and subsequen a ea and pe ime e calcula ions a e made [l], [2]. Some a emp s o in oduce he compu e in he decision-making p ocess ha e been p esen ed wi h a mode a e amoun o success. Fo ins ance, Zhang and Geise [3] ha e p oposed an image p ocessing algo i hm o ex ac ing he endoca dial bo de om a se ies o 2-dimensional echoca diog ams (2D echo). Thei me hod is a e y e icien me hod o con ou ex ac ion o echog aphic images bu some ained ope a o has o manually d aw he end sys ole and end dias ole con ou s. Besides, he algo i hm ails o de ec con ou s when he ca i y is no noise- ee. In his con ex , he e is a need o e icien me hods o noise emo al and pos e io pa ame e ex ac ion. This wo k has been suppo ed by NTE, S.A. unde Eu opean Space Agency con ac 7320/87/NLPP A main objec i e o digi al p ocessing o echoca diog aphic images is o imp o e he signal- o-noise a io o he ideo images acqui ed om he ul asound equipmen . In addi ion o he whi e Gaussian noise associa ed wi h ideo signals, he e is a la ge amoun o acous ic speckle p oduced by ue signal ha a i es a he ansduce , bu is no associa ed wi h e um echo om he hea . So, a p ime goal is he emo al o speckle noise o enhance he image in o de o acili a e bo h isual inspec ion o he hea and compu e pa ame e ex ac ion. A e iew o classical me hods o le en icula con ou ex ac ion and p ep ocessing o echoca diog aphic images is p esen ed in [4]. Once he image has su e ed a smoo hing p ocess, a con ou ex ac ion algo i hm may be applied in o de o ex ac ca diac pa ame e s. We ha e de eloped a comple ely ully au oma ic algo i hm o ha pu pose. The algo i hm s a s wi h he end sys ole and end dias ole ob ained con ou s. The main di e ence wi h he Zhang and Geise me hod is ha hese i s con ou s a e ob ained au oma ically. In o de o ob ain he wo i s con ou s, he image in o ma ion is changed o pola coo dina es and a adial analysis o he image is ca ied ou . Th ough some i e a i e app oxima ions, excellen quali y con ou s a e ob ained. A his poin , in o de o ake ad an age o he in o ma ion con ained in he empo al domain, he echnique is applied o he whole ca diac cycle hus ob aining a comple e se ies o ca diac con ou s ha a e used in a u he s age o ind he le en icula chambe olume and some o he ela ed pa ame e s. Tempo al p ocessing has been applied due o he ac ha ul asound images con ain a lo o co ela ion hus allowing us o use use ul echniques o dec ease he sea ch window and speeding up compu a ions. The algo i hm has been applied o di e en ca diac iews, such as pa as e nal sho axis, apical and mi al, wi h he same deg ee o success. 11. STATIC ALGORITHM As i has been said he s a ing images o he algo i hm a e he end sys ole and end dias ole con ou s. In o de o ind hese iis con ou s we ha e de eloped an au oma ic me hod which has been al eady epo ed [S].The i s s ep is o ind he coo dina e cen e in he in e io o he ca i y whe e wall con ou s a e being sea ched. Then he image is con e ed in o pola coo dina es by calcula ing I(kpo,l~o) om I(nxo,myo). Bilinea in e pola ion has been used. Once he image is in pola o m, he so called dis ance unc ion is ound, by de ining some special cha ac e is ic ( i s maximum, maximum alue, e c.) o each adius and d awing he esul ing unc ion. A di e en dis ance unc ion is e alua ed o each con ou . Fig. 1 shows he s a ing unc ion o he inne con ou (endoca dium) whe e 2149 CH2847-2/90/0000-2149 $1.00 0 1990 IEEE maximum alue o each adius has been used o de ine he dis ance unc ion. The goal o he algo i hm is o ind he bes possible dis ance unc ions o bo h he inne and ou e con ou s om hese i s s a ing dis ance unc ions. Fig. 1 also shows he inal con ou (dashed line). 'J m i7a S'I a Fig. 1. Ou e con ou dis ance unc ions. Con inuous line : Fi s Once he i s app oxima ion o he dis ance unc ion is ound, he algo i hm disc imina es which poin s o his unc ion belong o he con ou and which do no . To do his, successi e app oxima ions a e gene a ed un il he con ou , o a he i s co ec dis ance unc ion is gene a ed. The i s s ep is o de e mine, om he il e ed dis ance unc ion, a i s span o wha can be de e mined o be he ac ual con ou . This i s span will lead o ollow he con ou poin s. The p ocess o selec ing he bes span is done h ough ex apola ion o each span along he whole ange o he unc ion, and selec ing he span ha gi es he leas squa e e o . The ex apola ion is pe o med by a Fou ie se ies app oach, aking in o accoun ha he unc ion is con inuous, pe iodic and wi h many con inuous de i a i es. Th ough some i e a i e p ocess all co ec spans a e inco po a ed. Fo he sake o comple eness we summa ize he s a ic algo i hm in Fig. 2. 111. TEMPORAL ALGORITHM app oxima ion. Dashed line : Final esul . Once he end sys ole and end dias ole con ou s ha e been ob ained h ough he s a ic algo i hm, i is logical o apply he me hod o he whole se ies o empo al images. I knowledge o he ca diac s uc u e, mo ion and ela ed pa ame e s is inco po a ed, i will be possible o ind a as algo i hm o ind he whole se ies o empo al con ou s. The basic idea is o ind a me hod o con ou sea ching in such a way ha s a ing wi h an app oxima e knowledge o he posi ion o he eal con ou o e e y one o he images in a empo al ul asound sequence, i allows o ob ain he comple e se o eal con ou s. The ac o knowing he app oxima e posi ion o he con ou , allows o apply he eal con ou poin s iden i ica ion c i e ia o a es ic ed a ea a ound he p e ious es ima ion o he con ou posi ion. In his way he in e e ence due o o he ca diac s cuc u es such as papilla y muscles, al es, o e en ex emal con ou while he in emal con ou is being ob ained (o ice e sa), is a oided du ing he p ocess o de ec ing he poin s ha belong o he con ou wan ed o be ound. Sea ching in a educed egion allows ha he p ocess o con ou de ec ion be as e han in he case o s a ic images whe e all he poin s o he image need o be analized. Inpu IP ep ocessing I ICoo dina es cen e calcula ion] I Image ans o ma ion o pola coo dina es I /P ep ocessing I IDis ance unc ion compu a ion I IMedianFil e ing 1 [Selec ion o con inuous spans I Fou ie se ies app oxima ion o o de 0 and 1 o each span Mean Squa e e o compu a ion be ween he dis ance unc ion and each app oxima ion Selec ion o he bes span by leas e o app oxima ion Icon ou acking I lpoin elimina ion I IFou ie se ies app oxima ion I In e pola ion by cha ac e is ic sea ching In e pola ion by a g ay-le el sea chinp. I Poin elimina ion 1 IFou ie se ies app oxima ion I I Final Con ou I Fig. 2. Au oma ic con ou ex ac ion algo i hm o s a ic images The i s s ep is o p ep ocess he image in o de o elimina e he e ec s o he noise p esen in he images du ing he con ou de ec ion. A e wa ds, he image is con e ed in o pola coo dina es. The bes p ep ocessing esul s in sequen ial images ha e been p o ided by a 5*5 empo al / spa ial median il e ing o h ee consecu i e images in pola coo dina es. 2150 In all he images o he sequence, he sea ch o he con ou is s a ed based on an es ima ed posi ion o he eal con ou . This posi ion is de e mined om he con ou s co esponding o he s a and end o he sequence - end dias ole and end sis ole- and assuming ha he walls mo ion is lineal du ing he ca diac cycle. The c i e ium o sea ching he con ou poin s is applied o poin s a ound o a speci ied cen e de e mined by he es ima ed posi ion. This p ocess is applied o e e y adii o he image ha has p e iously been con e ed in pola coo dina es. The size o he sea ching egion is a pa ame e ha depends on he image in such a way ha egions ha could lend o w ong poin s de ec ions - papilla y muscles o he con ou s e c. - a e a oided. The de ec ed poin s a e hen alida ed and om hese poin s a i s app oxima ed con ou is buil . This i s con ou will be he s a ing poin o a second sea ching p ocess. The inal con ou is ob ained om he esul o his second sea ch o e e y one o he images o he sequence. The con ou s o end dias ole and end sis ole a e he s a ing poin o he inal p oposed algo i hm o empo al images. These con ou s a e called: R'd i=l, ..., N Ris i=i, ..., N whe e N is he numbe o adiis o he image o be p ocessed; in his case N = 360. These alues ep esen he dis ance measu ed in pixels om he cen e o coo dina es o he image o he poin o he con ou o each adii i. F om hese alues, he es ima ed con ou s o he es o he images ( he whole empo al sequence) a e ob ained by means o a lineal in e pola ion in he ollowing way: .. R'd-R', RInd= R, - ~ x (N-nd) N -N, whe e ns = 1 , . . . , Ns. nd = Ns+l , . . . , N. N = Ns+Nd+l. Ns is he numbe o images be ween he ins an s o end dias ole and end sis ole, Nd is he numbe o images be ween he ins an s o end sis ole and end dias ole and N is he o al numbe o images in he conside ed sequence. The esul s o his in e pola ion gi e he es ima ed con ou s o e e y one o he images o he sequence. I he poin o he es ima ed con ou is aken as he sea ch cen e o each adii, he c i e ium o con ou sea ch is applied o a window de ined a ound his de ined poin . The window size mus be in oduced as inpu o he algo i hm. The window size mus be la ge enough in o de o de ec he poin s ha belong o he eal con ou pe each adii , bu small enough in such a way ha o he ca diac s uc u es do no a ec he con ou de ec ion. Fo he gi en images, good esul s ha e been ob ained using a window size o 10 pixels, i.e., 5 abo e and 5 below he es ima ed cen e . The applied c i e ia ha e been: 1) De ec ion o he absolu e maximum, 2) Maximum o he absolu e de i a i e, 3) Maximum o he posi i e de i a i e and 4) Fi s non isola ed maximum. The bes esul s in he ca i y con ou de ec ion ha e been ob ained using he c i e ia o de ec ion o maximum posi i e de i a i e. In he ex e nal con ou de ec ion (pe ica dium) he bes esul s ha e been ob ained using he c i e ia o he maximum o he absolu e alue o he de i a i e. The c i e ia o he maximum absolu e alue in a es ic ed egion p o ides a goad de ec ion in he lowe pa bu since he image con ains a lo o noise, i p o ides a bad de ec ion in he egion co esponding o he endoca dium, i.e., egion be ween bo h con ou s. The de ec ion using he maximum absolu e de i a i e p o ides good esul s in he i s s age o he sea ching p ocess. The esul o his ii sea ch is alida ed by de ec ing se s o poin s h ough some connec i i y c i e ium, like o ins ance maximum dis ance be ween de ec ed pixels belonging o he same se and minimum leng h o span (minimum numbe o de ec ed consecu i e poin s e i ying he connec i i y c i e ia in o de o conside ha se o poin s alid). The c i e ia o span de ec ion se es o elimina e poin s ha ha e been de ec ed in a posi ions o he eal con ou . The esul ing poin s o his de ec ion ha do no belong o any span a e elimina ed. A i s app oxima ed con ou is buil wi h he alida ed poin s. This con ou is cons uc ed by inding he Disc e e Fou ie T ans o m ( DFT ) o he alida ed poin s, and limi ing he esul ing spec um o some speci ied equency samples. Good esul s a e ob ained wi h 5 equency samples. I he in e se DCT is ound om hese samples, a smoo hed con ou is ob ained ha app oxima es e y much he de ec ed poin s. Since he de ec ion by de i a i e means o de ec he ansi ions be ween e y di e en g ay le els, he esul o he de ec ed smoo hed poin s always gi es a con ou ha in some egions o he image is smalle han he eal con ou . This is sol ed by doing a new sea ch in which he s a ing poin is p ecisely his i s app oxima ed con ou ins ead o he ini ial es ima ion ob ained by he lineal in e pola ion om he s a and end images o he sequence. The sea ch c i e ium o his second p ocess is he absolu e alue. This sea ch is no done a ound he app oxima e con ou - as in he iis sea ch - bu s a ing om he app oxima e posi ion o he con ou pe each adii and mo ing owa ds he ou e posi ion. Good esul s a e ob ained wi h sea ch windows o 4 pixels. The inal con ou is ob ained building a smoo hed con ou based on he DFT and in e se Dm wi h a limi ed numbe o equency samples o he de ec ed poin s in his second sea ch. The same me hod is applied in he de ec ion o he ca i y in e nal con ou (endoca dium). The decision c i e ia a e he only signi ica i e changes. In he i s de ec ion he applied c i e ium is ha o de ec ion by he maximum alue o he posi i e de i a i e. This a oids he s ong ansi ions o g ay le els in egions e y close o he con ou posi ion caused by papilla y muscles o al es, in sec ions co esponding o 2154 de e mined egions o he le en icle in pa as emal sho axis iews. Once he decision c i e ium has been applied, he algo i hm is exac ly he same as ha o he ex e nal con ou . Fig. 3 shows he inal ex emal and in e nal con ou s o a empo al sequence o 10 images be ween end sys ole and end dias ole images. 1V.CONCLUSIONS A ully au oma ed algo i hm ha ex ac s en icula endoca dial and epica dial bo de s o empo al pa as emal sho axis iews has been de eloped. The algo i hm may be applied o o he echoca diog aphic iews wi h he same amoun o success. Once he ex e nal and in e nal con ou s a e ound, en icula chambe olume and o he ela ed pa ame e s can be ound. REFERENCES [31 [41 S.Collins, D.J.Sko on "Ca diac Imaging and Image P ocessing" Mc.G aw-Hill, 1986 D.J. Sko on, e al. "Digi al signal and image p ocessing in echoca diog aphy" Ame ican Hea jou nal Vol. 110, NQ 6. pp.1266-1283, Dec. 1985 L. Zhang, E.A. Geise "An e ec i e algo i hm o ex ac ing se ial endoca dial bo de s om wo-dimensional echoca diog ams" IEEE T ans. Bio. Eng. BME-31, pp. L.To es e al. "Classical me hods o le enmcula con ou ex ac ion and p ep ocessing o echoca diog aphic images : A e iew." Ul asonics In ema ional 89, Mad id, 3-7, July 441-447, 1984 1989 [SI A.Gasull e al. "Au oma ic le en icula con ou ex ac ion o olume calcula ion om echoca diog aphic images" Ul asonics In ema ional89, Mad id, 3-7 July 1989 Fig.3. Ex e nal and In e nal con ou s o an end sys ole-end dias ole ca diac cycle 2152