Seyed Mostafa Mirhassani Department of electrical engineering, Saveh Branch, Islamic Azad University, Saveh, Iran

Transcription

1 nternatonal Journal of Coputer Applcatons ( ) Detecton of Narrowed Coronary Arteres n X-ray Angographc ages usng Contour processng of Segented Heart Vessels based on Hessan Vesselness Flter and Wavelet based age Fuson M.J.Rastegar Fate Departent of electrcal engneerng, Saveh Branch, slac Azad Unversty, Saveh, ran Seyed Mostafa Mrhassan Departent of electrcal engneerng, Saveh Branch, slac Azad Unversty, Saveh, ran Elha Ghase Departent of coputer engneerng, Saveh Branch, slac Azad Unversty, Saveh, ran ABSTRACT n ths paper a ethod for Detecton of narrowed coronary arteres n X-ray Angographc ages s addressed. For ths purpose frstly coronary arteres are segented base on soe probablty asks whch ndcates the ebershp functon of beng a vessel for each pxel n the age. n order to buld the ebershp functons whole of angographc fraes n the sequence are utlzed. To ths a, angographc sequence s fltered by the Hessan based vesselness flter to ncrease dscrnaton between the vascular structures and the background. Next, a D dscrete wavelet transfor s eployed for fuson of the angographc fraes and ther fltered versons. n the next step, the heart vessels are detected by applyng the fused age wth a couple of constant values as threshold coeffcents. The coeffcents are eployed to adjust the effect of threshold. As a consequence, two ages contanng the detected vessels are yelded. After that, n order to ask out the redundant partcles n the result orphologcal operatons are eployed. n order to fnd narrowed vessels ther contours are obtaned and the thckness of vessels s easured. A predefned threshold s eployed to fnd the abrupt decrease n the vessel thckness. Experents deonstrate the effcency of the proposed ethod n order to detect the narrowed heart vessels n X-ray angographc fraes. Keywords Vessel segentaton, X-ray angographc age, wavelet based age fuson, contour analyss, thckness easureent 1. NTRODUCTON n recent decades, edcal agng technques ncludng X- ray angography and coputed toography angography (CTA) have provded a rearkable aded the dagnoss of atheroscleross. n ths way, vessel analyss n edcal ages and vdeos s one of the ost sgnfcant research felds because t s approprate n dagnostc and nterventon plannng purposes. For varous vsualzaton applcatons such as ultplanar reforats or endovascular vews, extracton of vessel centerlne can be eployed to generate partcular vsualzaton nforaton. n addton, for soe of other applcatons ncludng quantfcaton, e.g. for deternng the denson of stents or stenoss gradng, vessel segentaton can play an portant role. n ost of ethods, proposed n lterature, for vessel analyss, ages enhanceent s the frst step to prove vascular structures. t helps vessel vsualzaton, e.g. n axu ntensty projectons or volue renderng technques. For vessel segentaton, n case of coronary arteres n angographc ages, dgtal postprocessng operatons are essental to solate vessels fro background. n any of technques the knowledge that vessels are tubular structures s eployed to descrbe vessels by ther central axes and wdth. Central vessel axs (CVA) segentaton s presented n [1, ], n whch for recognzng the type of local structure n age (e.g., tubular-lke), nforaton fro nd-order dervatves at ultple scales s extracted. To obtan ths nforaton, the an ode of varaton n Hessan atrx s nspected. Hessan-based vessel enhanceent flters (HBVF) have been proposed by Lorenz et al. [3], Sato et al. [4], and Frang et al. [5]. HBVF s a powerful ethod for enhancng vascular structures n varety of agng tasks ncludng DSA [3, 5], CTA [4], and 3D MRA [1, 4, and 5]. To prove vsualzaton and dagnostc purposes, n [6], vessel enhanceent s accoplshed usng vessel enhancng dffuson (VED). n ths ethod, vessel segentaton n X-ray angographc frae s realzed usng soe experts whch are n tande. Frstly Hessan based vessel enhanceent flter (HBVF) s appled to the angographc frae to be segented for enhanceent of vessel structures. Meanwhle, the flter s appled to other fraes n the angographc sequence. Afterward, angographc fraes n the sequence and ther vesselness fltered versons are fused usng -D wavelet transfor to ake an age whch s used as a threshold for detectng the vessels. Next, for detectng the vascular structures the fused age accopanyng wth a couple of thresholds s appled to the vesselness fltered frae to be segented. Fnally, n order to fnd the narrowed coronary arteres, vessel contours processng s eployed for the easureent of the vessel thckness. The reander of the paper proceeds as follows: n secton vesselness flter for enhancng the tubular structures n age s ntroduced. n Secton 3, -D dscrete wavelet transfor and ts applcaton n age fuson are revewed. Secton 4 focuses on the proposed vessel segentaton ethod, followed by vessel contour processng for detecton of narrowed vessels n secton 5, experental results n secton 6 and concluson n secton 7. 7

2 nternatonal Journal of Coputer Applcatons ( ). VESSELNESS FLTER Detecton of vascular structures s the an purpose of ths flter. To ths a, the egensyste of the Hessan atrx s analyzed by the flter to deterne the curvature drecton of age whch s necessty for detecton of vascular structures. Curvature drecton n each pont s the drecton accordng to egenvector of Hessan atrx n whch the second order of age nforaton s extreu. Eq(1) presents Hessan atrx whch deternes the second order nforaton of D age. xx xy (x, y) H yx yy Where H and denote Hessan atrx and orgnal age and xx, xy, yx and yy denote second dervatve of age nforaton, respectvely. To obtan and x y orgnal age can be fltered usng =[ 1 1]and = [ 1 1] T vectors. After coputng egenvalue and egenvector for Hessan atrx of each age pxel, followng condtons are possble: For brght vascular structures on a dark background f the orderng of egenvalues are 1, v ndcates the 1 drecton along the vessel when 1 0 and 1. Based on egenvalues of Hessan atrx, nuerous vesselness flter are proposed n lterature [3],[4],[5],[7]. n ths work, we use the defnton of vesselness functon that proposed n Frang et. al. work [5]. The forulaton of vesselness easureent, proposed n [5], s gven as follows: For D ages: f ( ) e 0 R B S c.(1 e ) R A 1 0 otherwse Where R A ake dstncton between plate and lne lke structures, RB nvestgates devaton fro a blob lke structure, and S dfferentates between vessels and background. The paraeters β, c adjust the nfluence of R A, RB and S. To deterne vesselness response, Hessan wth Gaussan dervatves at ultple scales s coputed and the hghest response s consdered as the output of the flter. Fg. 1: The structures of -D DWT wht 3 decoposton levels 3. DSCRETE WAVELET TRANSFORMS (DWT) The wavelet transfor s an effectve tool for ultresoluton age analyss. t decoposes age nto lower resoluton sub-spaces for analyss. To construct a wavelet decoposton of sgnal, the sequence of FR flters g and h can be eployed. To ths a, the dscrete scalng functon and the dscrete wavelet functon are defned as follow: /, j( k) h ( k j) / ( k j), j( k) g (4) Therefore, we obtan a wavelet orthonoral bass: DWT S { N, j, 1, j,, j,..., N, j} j Consequently, a dscrete sgnal x can be expressed usng these scalng functon and wavelet functon: N ( N ) ( j) N, j ( k) d( ) ( j), j ( k) 1 jz x( k) s 3.1 -D DWT For -D sgnals such as ages, -D wavelet transfors are eployed. Fgure 1 presents the structures of -D DWT wth 3 decoposton levels. After one level of decoposton, four frequency bands are obtaned ncludng Low-Low (LL), Low-Hgh (LH), Hgh-Low (HL) and Hgh-Hgh (HH). The next level decoposton s just appled to the LL band of the current decoposton stage, whch fors a recursve decoposton process. As a consequence, N-level decoposton wll fnally have 3N+1 dfferent frequency bands, ncludng 3N hgh frequency bands and just one LL frequency band. The -D DWT wll have a pyrad structure shown n fgure age fuson usng wavelet transfor age fuson s one of the ost portant applcatons of wavelet transfor. Fgure presents the block dagra of a generc wavelet-based age fuson schee. For age fuson usng wavelet, frstly wavelet transfor s perfored on each source ages to be fused, afterward based on a set of fuson rules; a fuson decson ap s generated. z (5) 8

3 nternatonal Journal of Coputer Applcatons ( ) Accordng to the fuson decson ap, the fused wavelet coeffcent ap can be constructed usng wavelet coeffcents of the source ages. Fnally to ake the fused age, nverse wavelet transfor s accoplshed. As the dagra ndcates, the fuson rules are playng a very portant role durng the fuson process. Pxel-based and wndow-based fuson rules are two frequently used ethods n lterature [8], [9]. n the forer ethod, to ake decson on one of the coeffcents of the fused age, the correspondng coeffcents n the source ages are consdered, whle n the latter, not only the correspondng coeffcents, but also ther close neghbors are consdered. n ths approach we use a pxel based fuson rule for fusng the angographc fraes. 4. VESSEL SEGMENTATON After applyng the vesselness flter to the angographc frae, the structures lke vessels becoe ore vsble. Note however that, n the background there are soe partcles such as rbs that take a great value due to applyng the vesselness flter. As a result of applyng the flter to such structures, the vessel segentaton algorth results n error. As a reedy, n ths approach we use nforaton obtaned fro dfferent angographc fraes to flter out the non-vessel structures. Fgure 3(a) presents a saple of angographc frae to be segented by the algorth and the result of applyng vesselness flter on the frae (Fgure 3 (b)). As t s shown n Fgure 3 the vessel pxels take low ntensty levels n the angographc frae whle ther ntensty levels n the fltered age are hgher than the background. Consequently, for detectng the vessels a threshold can be appled to the vesselness fltered age. As dscussed above, before applyng the threshold, soe of non-vessel structures n the background should be reoved. For ths purpose, accordng to the ntensty level of the vessel structures n the angographc frae and ts fltered verson, two ages n whch the level of each pxel presents the probablty that a pxel belongs to the vessels are coputed. Fg. : Wavelet based age fuson schee To obtan such ages followng forula s utlzed n whch each pxel between the whole fraes n an angographc ove s consdered. [ x, y ] n( [ x, y, t]), f [ x, y ] ax( [ x, y, t]) f :1,..., M t 1,..., L j :1,..., N (6) Where, M, N and L denote the angographc sequence to be segented and the nuber of rows, coluns and fraes of ts atrx, respectvely. n ost of dark pxels belong to the vessels; whle n brght pxels have a ore probablty f to belong to the vessels (See Fgure. 3(a, b)) to ake slar to f (n whch the dark pxels belong to the vessels) nv we use negatve verson of and nae t. For dscrnatng the vessels n the fltered age f and nv are used to ake a threshold for applyng to the fltered age. n ths approach we use an age as the threshold whch s obtaned by fusng f and nv usng wavelet theory. n the prevous secton we ade a revew about wavelet based age fuson. Here, for fusng two ages after decoposton step, we use pxel based fuson rule n whch the coeffcent of the fused age s obtaned by averagng the nv correspondng Haar wavelet coeffcent n f and. These coeffcents are coputed usng 5-level decoposton. We can suarze above dscussons as follow: f, nv Wavelet Based age Fuson FUS (7) Where FUS s the fused age. For akng the threshold for vessel detecton we use FUS as follow: V [ x, y, t] Th (8) FUS. Where V and Th denote the vessels and the coeffcent for adjustng the effect of threshold. Adjustng the threshold paraeter (TH) for detecton of vessels s a crtcal task because applyng a hgh threshold value can reove soe of vessel structures whle a low threshold value cannot flter out non-vessels. Fgure 3(c, d) presents the response of vesselness flter after applyng threshold. Effects of applyng hgh and low aounts of threshold on the fltered age by vesselness flter confor above dscusson. t s worth notng that although soe parts of vessels are reoved due to applyng a hgh aount of threshold, the drecton of reanng parts s obtanable; accordngly, whole of vessels can be recovered accordng to these drectons. n ths approach, to cope wth the proble of vessel segentaton, frstly a hgh level threshold s appled to the age to reove ost of nonvessels fro age. 9

4 nternatonal Journal of Coputer Applcatons ( ) Fg. 3: Orgnal age (a) and the response of the vesselness flter (b). Applyng a low (c) and a hgh (d) level threshold on the vesselness fltered age (For th = 0.7 (d) soe part of vessels are reoved, and for th = 0. (c) soe of rbs are shown as vessels n error). Subsequently, the reanng vessels are detected and ther drectons are obtaned for recoverng the reoved parts. Afterward, accordng to the obtaned drecton structure eleent for dlaton operator s constructed (as a bnary atrx) and t s eployed to expand the reanng vessels as a recoverng step (see Fgure 4 (d)). Ths yelds soe false postves n the detected vessels. To reduce the false postves n the yelded age, a low level threshold s appled on the vesselness fltered age. Fnally product of two ages s consdered as the result of the approach. Where ( x, y) ( x, y). ( x, y) r thl, D and thl D r denote vesselness fltered age after applyng low level threshold, the dlated age and the result age, respectvely. Accordng to eq.9 snce D are bnary ages, between thl and D. thl and r s result of applyng and operator As a fnal pont, accordng to the fact that soe of dagnostc applcatons only need an vessels, reovng fne partcles of age could yeld a ore clear result. For ths purpose, frstly labelng operaton s accoplshed n the result age. Fg. 4: Other exaple of X-ray angographc frae (a) ts vesselness fltered verson (b) the fused age to be used as threshold (c) recovered vessels after applyng the hgh level threshold ( D ) (d) the result age ( r ) before (e) and after (f) partcle reoval. Afterward, the sze of each label s consdered and the labels saller than a predefned sze are reoved. Adjustng the sze threshold for reovng the partcle s crtcal ssue n segentaton of vascular structure. t depends on the nu sze of vessels to be detected. Fgure 4 presents other exaple of X-ray angographc age and the result after perforng each part of the proposed algorth. 5. Detecton of narrowed coronary arteres n the angographc fraes As a ost portant a of ths study, n ths secton a ethod s proposed to detect the narrowed heart vessels n angographc fraes. Reducng the edcal stakes, arses fro the sght error, s the ost otvaton of scentsts beyond conductng such researches. n other words by ntegraton of decson support by eans of edcal age analyss nto clncal practce t s expected to reduce the huan error n dagnoss purposes. n ths secton we focus on detectng narrowed vessels n X-ray angographc fraes. The block dagra of the proposed algorth s presented n Fgure 5. n ths ethod vessel contours are eployed to easure the vessel thckness. n coputer vson, contours are known as the outlnes of age regons. To obtan the contours of vessels, boundary detecton s used n ths algorth n partcular, a orphologcal dlaton of segented vessels, followed by subtractng fro the vessels results n the boundary of the vessels. Subsequently, tracng of the yelded boundary usng the connected coponent analyss provdes the contours. 30

5 nternatonal Journal of Coputer Applcatons ( ) As prevously entoned, the centerlne s utlzed as a reference for the easureent. Consequently, the two sdes of vessels contours (the red ponts n the Fgure 7) should be obtaned. For ths purpose, frstly the drecton of the vessel n the pont the reference pont (the blue pont n the Fgure 7) s coputed. Afterward a lne segent, perpendcular to the vessel centerlne, s generated. Concdence of the lne segent and the vessel contours (boundares) are consdered to obtan the thckness. After obtanng the vessel thckness n whole of the age, they are nvestgated for the observaton abrupt decrease whch can results n detecton of narrowed vessels. Fgure 8 presents a narrowed vessel and ts thckness plot, coputed by the proposed algorth. As Fgure 8 Shows, the edan of the thckness plot falls suddenly n the narrowng part of the vessel. n addton, n the soothed verson of the thckness plot, a local nu s observed. 6. EXPERMENTAL RESULTS n ths secton for bencharkng the proposed algorth t s appled to soe of X-ray angographc fraes for vessel segentaton. Fg 5: The block dagra of the proposed ethod Fgure 6 shows the angographc frae ncludng the vessels and ther contours. n order to easure the thckness of vessels, the dstance between each pont of the vessel contours, and ts correspondng pont n other sde of vessel should be coputed. For ths purpose, frstly the centerlne of the vessels s obtaned usng a thnnng algorth. The centerlne s used as a reference for the thckness easureent. Fgure 7 presents the dea eployed to obtan the thckness of the vessels. Fg 6: angographc frae ncludng the heart vessels and ther contours Fg 8: the detected narrowed vessel (up), the edan of the vessel thckness easured by the proposed algorth (ddle) and the soothed verson of the thckness Fg 7: Measurng the vessel thckness 6.1 Results of vessel segentaton algorth To evaluate the effcency of the algorth, 19 angographc fraes are used. Three paraeters are eployed as threshold ncludng hgh and low level threshold for detectng vessels and the partcle reoval threshold whch flters out fne partcle n the result age. The aounts of hgh and low level thresholds for vessel dscrnaton are 0.7 and 0., respectvely and the partcle reoval threshold s 3000 pxels. The block dagra of the approach s presented n Fgure 9. As dscussed n the prevous secton, the novelty of the proposed algorth s usng a fused age, obtaned fro the 31

6 nternatonal Journal of Coputer Applcatons ( ) fraes of angographc sequence, as a threshold for detectng the vessels n the vesselness fltered age. For fusng the angographc fraes and ther fltered verson, a pxel based nv rule s utlzed n whch n f and ages the coeffcents, provded after fve level decopostons, are cobned by obtanng ther average. As an assessent of algorth effcency, t s copared to vessel segentaton by a ethod [10] n whch the thresholds for vessel segentaton are only two constant values (not two Fg. 10: Coparson of two ethod accuraces for 16 angographc fraes ncludng the proposed ethod (blue lne) and the ethod whch used two constant values as threshold for detectng the vessels (red lne). 6. Results of narrowed vessel detecton and vessel thckness easureent Fgure 11 presents the edan and the soothed verson of the thckness of a typcal narrowed vessel. An abrupt change n the narrowed part of the edan of the thckness and a local nu n the soothed verson of the thckness plot are observable. The plots ndcate the effcency of the proposed ethod for to easure the vessels. Fg. 9: Block dagra of the proposed vessel detecton ethod ages). After applyng the algorth on the angographc fraes the result s copared wth a verson of frae n whch the vessels are anually labeled. To deterne the aount of paraeters ndcatng the accuracy of the ethod soe paraeters are coputed. These paraeters are false postve (FP), false negatve (FN) and true postve (TP). False postve presents nuber of non-vessel pxels whch are erroneously classfed as vessels. Whle false negatve eans the nuber of pxels whch are vessels but the algorth could not detect the, fnally true postves are the nuber of pxels whch are classfed correctly as vessels. Snce soe stakes are ade whle anual labelng of vessels, true postves are counted as followng way. For every pxel n the known vessels f there s a arked pxel as vessel, ts 3 3 neghborhood s also counted as a true postve. Fgure 10 presents the coparson of effcency of the proposed ethod and the ethod proposed n [10] whch uses two constant values as threshold. Effcency of the ethods s coputed usng followng forula: Accuracy Tp Tp Fn Fp (10) n the forer ethod, the hgh level threshold and the low level threshold are 0.7 and 0., respectvely and the threshold of partcle reoval step s equal to 3000 pxels, whle n the latter, the hgh and low level threshold are 0.8 and 0., respectvely. Due to usng the fused age, non-vessel partcles of age are effectvely fltered out because the fused age takes greater values n the non-vessel regons. Consequently, as Fgure 10 shows, the proposed ethod outperfors the other ethod accuracy. Fg 11: plots ncludng edan (up) and the soothed verson (down) of the thckness of a typcal narrowed vessel coputed by the proposed algorth However, the detected narrowed vessels depend on the results of the vessel segentaton stage. f the segentaton stage fals to detect a narrowed part of vessel, the easureent 3

7 nternatonal Journal of Coputer Applcatons ( ) stage cannot detect t. Moreover, soe of the detected narrowed vessels are not abnoral vessels. n other words, the sde s portant to observe an abnoral narrowng of vessel. As a result, nvestgaton of an operator s stll essental n order to ake the fnal decson of observng an abnoral vessel. 7. CONCLUSON n ths paper a ethod for segentaton of vessels n X-ray angographc fraes and detecton of narrowed vessels s proposed. n ths ethod frstly Hessan based vesselness flter s appled to the angographc fraes to enhance the vessels. Afterward, a couple of threshold ncludng hgh level and low level threshold, obtaned by fusng the whole of fraes n the angographc sequence, s appled to the vesselness fltered age for detectng the vessels. Due to applyng the hgh level threshold soe of thn vessels are reoved. To recover such vessels, orphologcal operatons s utlzed. Subsequently, the two yelded ages are cobned by obtanng ther product; fnally the partcle reoval s accoplshed to reduce the aount of false postve. To detect the narrowed vessels the vessel contours are analyzed and easured. After applyng the algorth on the X-ray angographc fraes, ost of vessels were effectvely detected. 8. REFERENCES [1] Wnk, O., et al D MRA coronary axs deternaton usng a nu cost approach, Magnetc Resonance n Medcne. 47 (6) [] Aylward, S.R., Bullt E. 00. ntalzaton, nose, sngulartes, and scale n heght rdge traversal for tubular object centerlne extracton, EEE Transactons on Medcal agng, 1 () [3] Lorenz, C., et al Mult-scale lne segentaton wth autoatc estaton of wdth, contrast and tangental drecton n D and 3D edcal ages, CRVMed-MRCAS Proceedngs, Sprnger Verlag, [4] Sato, Y., et al D ult-scale lne flter for segentaton and vsualzaton of curvlnear structures n edcal ages, Medcal age Analyss () [5] Frang, A.F., et al., Multscale vessel enhanceent flterng, MCCA Proceedngs, Sprnger Verlag,, [6] Y P.J. et al., 001. Vessel surface reconstructon wth a tubular deforable odel, EEE Transactons on Medcal agng, 0 (1) [7] Krssan, K., Malandan, G., Ayache, N., Vallant, R., Trousset, Y., 000. Model-based detecton of tubular structures n 3D ages. Coputer Vson and age Understandng, 80 (), [8] Lews, J. J., Robert, J., O Callaghan, S., Nkolov G., Bull, D. R., Canagarajah, N., 007. Pxel- and regonbased age fuson wth coplex wavelets, Journal of nforaton Fuson, Elsever, 8 (), Aprl 007, [9] Chareonta, O., Patanavjt, V., Jtapunkul, S., 006. A Robust Regon-Based Multscale age Fuson Schee for Ms-Regstraton Proble of Theral and Vsble ages,8th nternatonal Conference on Pattern Recognton (CPR'06) 3, 006, , [10] Mrhassan, S.M., Hossen, M.M., Behrad A., 009. proveent of Hessan Based Vessel Segentaton Usng two Stage Threshold and Morphologcal age Recoverng, nternatonal Conference on nnovatons n nforaton Technology, (T '09), (Dec. 009),