Published on

  • Be the first to comment

  • Be the first to like this

No Downloads
Total views
On SlideShare
From Embeds
Number of Embeds
Embeds 0
No embeds

No notes for slide


  1. 1. AN ONLINE MEASURING SYSTEM TO SUPPORT HEART VALVE SURGERY N.Rajanipriya, Deepa.M B.V.Raju Institute of Technology, B.V.Raju Institute of Technology, Narsapur,Medak(dist). Narsapur,Medak(dist). Email : Email: ABSTRACT of mitral valve for a suitable valve replacement. It In this paper we present an online system was also intended to provide a friendly solution to bedesigned to assist heart surgeries, in particular to used by surgeons in real time.provide support to make the decisions during implant Segmentation was done using activeheart valve operations. The 2-D echo image is contours and dimensions are measured from Globalobtained from the output console of the ultrasound Image Binarization.scanner. The heart valves in the image are We have also developed a Graphical Usersegregated by processing the image using active Interface using Visual Basic to record the suggestionscontours. The properties of that valve: the given by the doctor for artificial heart valvedimensional properties such as number of cusps, replacement.area of valve, radius of each cusp etc., and materialproperties such as the elasticity, viscosity and BACKGROUNDchemical interactions with in the body are obtained. The heart is the center of the cardiovascularThe computer assisted system consists of the system. Whereas the term ‘cardio’ refers to the heart,database which has the dimensional and the material the term vascular refers to blood vessels. The heartproperties of the artificial heart valves. When the propels blood through thousands of miles of bloodproperties of natural heart valve are given as an vessels, and it is magnificently designed for this task.input to the computer assisted system, it compares The interior of the heart is divided into fourthem with the properties of artificial heart valves in compartments called chambers that receive thethe database using the pattern recognition principles. circulating blood. The two superior chambers areThe matching of the properties in the system called the right atrium and left atrium. The twodetermines the selection of artificial heart valve with inferior chambers are the right ventricle and leftwhich the valve of the patient has to be replaced. ventricle. As each chamber of the heart contracts, itThis therefore forms the decision making system pushes a portion of blood into a ventricle or out of theassisting the doctor to identify the appropriate heart through an artery. To prevent back flow ofartificial heart valve for replacement. blood, the heart has valves. These structures are composed of dense connective tissue covered by INTRODUCTION endocardium. Valves open and close in response to pressure changes as the heart contracts and relaxes. In recent years mostly since the end of the Atrioventricular valves (AV) lie between the atrialast decade growing investment in the development and ventricles. The right AV valve between the rightand improvement of computer vision techniques has atrium and right ventricle is also called the tricuspidbeen a common policy of companies and countries. valve because it consists of three cusps (flaps). TheSuch techniques rely on the unstoppable advance of left AV between the left atrium and left ventricle hastechnology which provides cheaper and faster two cusps and is called the bicuspid (mitral) valve.hardware support on a regular basis. Both arteries that emerge from the heart valve that Computer vision aims to increase the speed prevents blood from flowing backward into the heart.and quality of some problem solutions, presenting as These are the Semi lunar (SL) valves. The pulmonaryits ultimate goal, the creation of reliable systems semi lunar valve lies in the opening where thecapable of taking the right automatic decisions pulmonary trunk leaves the right ventricle. The aorticsupported by image analysis thus achieving results valve is situated at the opening between the leftthat until then were only made possible by human ventricle and the aorta.intervention. Medicine one of the most ancient and Heart valve prostheses are intended for useimportant fields of science is surely no exception and as the replacement of diseased natural valves. Thethe main problem that led to our work was related in four natural valves (tricuspid, mitral, pulmonic andthis area in particular with heart surgery. aortic) become diseased, due, in part, to rheumatic The problem we tackled in this work was to fever, calcification, endocarditis, and congenitalautomatically segment and measure the dimensions defects. This results in the restriction of the forward
  2. 2. flow of blood known as stenosis or regurgitation flow snaxels, connected by line segments. The position ofof blood known as insufficiency. The advent in the snake is described by an M × 2 matrix X, where thescience and technology has led to new and improved ith row contains coordinates of the ith snaxeldesigns of these heart valves. Heart valve prostheses x0 y0may be classified into 2 types, mechanical valves and x1 y1tissue valves. Mechanical valves are classified into 3 X= .... .... ……………………… (1)categories, caged ball valves, tilting disk valves and xM − 1 yM − 1bileaflet valves. Tissue valves use materials ofbiological origin as the valving elements viz,allograft valves, aortic valve xenograft and The contour motion is a result of interplaypericardial xenigraft. between its mechanical properties (defined as its SEGMENTATION AND TRACKING IN inertia, internal dissipation, and elastic stiffness) and ECHOCARDIOGRAPHIC SEQUENCES: its potential energy (which is inversely proportional ACTIVE CONTOURS GUIDED BY OPTICAL to the image gradient). The equation of motion for FLOW ESTIMATES this model is Tracking algorithm that uses a Kalman filter MX(t)+CX(t)+KX(t)=F(t) …………….. (2)to track the object and estimate true motion has beenproposed. The observations for this Kalman filter are where F is the matrix of x and y components ofoptical flow estimates obtained by application of the image forces at each vertex, where M = μ.Im, where μblock-wise Horn and Schunk algorithm. With such is the mass assigned to each vertex, C = γ.I m, where γobservations, the tracker should be able to more is the constant damping density and K is an MxMsuccessfully handle sudden movements of the object stiffness matrix. Dots above the X represent theof interest. However, the use of optical flow as the derivatives in time.only input to the algorithm means that the current Two initial conditions are needed for solvingposition and shape of the object are estimated by the second-order differential equation. If the initialintegrating object motion. This results in position and velocity are known, (2) has a uniqueaccumulation of errors in the optical flow estimates solution. This suggests the way to incorporate theover time. This model does not provide a way of optical flow information into the active contourincorporating the current image information into the model. For the first frame of the image sequence, acurrent estimate of object shape and position. user defines the initial contour, and the initial velocity In the algorithm, the tracking problem is vector, is set to zero. Once the contour settles in thesolved by using optical flow estimates as initial first frame, the subsequent initializations are doneconditions when integrating the equation of motion. automatically. The initial position is given by theIn other words, optical flow is used to push the final position from the preceding frame, and thecontour toward the new position of the object. initial velocities of vertices are given by the opticalIn the simpler algorithm proposed by Ayache, an flow measure that tends to preserve the matching Optical flow estimates give the averageof high curvature points and to enforce a smooth field velocity of the object between two frames, not theof displacements vectors between contours in two initial velocity. They can, however, be used as initialconsecutive frames is minimized (such an approach velocities for two reasons. First, we are using a smallalso requires previous edge detection). In the method value for the damping factor, and second, since thewe propose, instead of matching pieces of contours final contour position is adjusted through the actionaround high curvature points, through the use of of image forces, we only need rough estimates ofoptical flow, the best matches for image regions initial velocity.around each vertex of the contour are found in the To estimate optical flow, we used anext frame. The benefits of the latter approach are multiscale version of Singh’s algorithm, which isnotable in the high deformation frames (our capable of detecting large displacements.algorithm required user intervention in only oneframe in three sequences showing the mitral valve.METHOD A. Integration of the Equation of Motion The contour is defined as a set of M vertices, 1) Single-Step Time Marching Scheme:
  3. 3. The equation of the dynamic ε equilibrium of the contour (2) can be integrated using a single-step time- ∑ M − 2 dF I dX i marching scheme (SSTMS). In the ε = M =1 i 0 ……………………… (6) ∑ − ( dF I dX I ) i =0 0 SSTMS, the time t is divided into several intervals, or time steps ∆ t i . In . where dF = Mα + C X + kX − F is the force residue each time step, the position and the and dX is the position increment. If ε is less than velocity of the object are updated. 5%, it is assumed that the contour has reached the Calculations performed for each time equilibrium. interval are i) X.. + 2 = X.n ∆t θ2 n 3) Calculation of the Contour Position: The time interval between two successive frames is assumed to X.. + 2 = X. be one. This time interval is then divided into Nt n n intervals for which STMS integration is executed. To avoid attraction of the contour by noise and undesired  ∆ t2  objects that can be in its way while moving towardii) α n =  M ∆t θC1  θ2 K    2  the new position of the object of interest, we chose a nonuniform time division. The first interval ∆ t 0 is set× ( F − C X.. + 2 − k X .n +1) n at 0.9 and the remaining 0.1 is divided into Nt-1 equal intervals ∆ t n . During ∆ t 0 , image forces are set to ∆ t2 zero to prevent attraction of the contour by undesirediii) X n + 2 = X n + ∆t X.n + α n 2 valleys of the potential surface. In other words, we X.n + 2 = X.n + α n ∆t ………… (3) are making the potential surface flat for the first 90% of the time. During this time, elastic forces act on thewhere n is the time step number, X 0 = X(0) and contour, filtering inaccuracies in velocity estimates.X. 0 = X. ( 0 ) and θ1 and θ 2 are constants. The During the remaining 10% of the time, image forcesunconditional stability of the integration is are included, making the contour settle on the edgesguaranteed if θ1 ≥ 0.5 and θ 2 ≥ θ1 . Steps i)–iii) are of the object of interest.repeated for each of the time intervals. The problem The contour is resampled after each iteration tois nonlinear if the stiffness matrix K is not constant compensate for possible clustering of the snaxels inand depends on the position of the contour. In such a the local energy minima along the valley of thecase, the equation of motion takes a more general potential surface. After resampling, distancesform between snaxels are equal. Since both positions and .. . velocities of points from one iteration are used in theM X + C X + P( X ) = F ……………… (4) following one, velocities of new snaxels have to be recalculated as well. The position of a new snaxel is a In this case, the problem is linearized inside a time linear combination of positions of two old snaxelsinterval and, hence, its velocity is also a linear combination of _ velocities of the same two snaxels.P( X ) = P X n + K ∆X ……………………. (5)where K is the value of ( K ) at some average value _ B. Contour Stiffness and Internal Forcesof X inside the current interval. (4) has to be solved In the case of the discrete contour, internal forcesiteratively using steps ii) and iii) of the procedure depend on positions of vertices. Internal forces cangiven in (3). It will be shown that the stiffness matrix usually be decomposed into the product of thein our model depends on the position of the contour. stiffness matrix and the position matrix Fint=-K(X)XHowever, the internal forces can be decomposed as Where such decomposition is not possible , theP(X)=K(X)X,with stiffness matrix K recalculated at equation of motion takes on the form of (4).the beginning of each iteration. Depending on the shape of the object of interest,2) Equilibrium Criterion: The iteration process is different internal forces can be designed. The role ofterminated when the contour attains the equilibrium such forces is to enforce on the contour a shapeat each time step. In the proposed model, the ratio of feature that is characteristic of the object. Forcontour deformation energy in the current and the example, in the case of heart valve leaflets, assigningfirst iteration was chosen as an equilibrium measure the contour elastic properties that try to preserve
  4. 4. length will result in failure of the active contour to the developed approach is quite flexible. A prioridetect sudden length changes. However, the leaflet knowledge about the shape characteristics of theshape feature that seems to be preserved throughout object of interest can be incorporated into thethe image sequence is leaflet thickness. This fact definition of internal forces, and the rest of thesuggests that in the case of valve leaflets, additional algorithm remains unchanged.forces that act to preserve contour thickness shouldbe added. To define such forces, for each point i on thecontour, the distance di from the matching point onthe opposite side is calculated (Fig. 1). Also, the 0 idistance d i calculated at the optimal position in thefirst frame is stored and considered to be optimal.The internal force that acts on the vertex i is thendefined as a force proportional to the relative change diin length d0 di − di di 0  1 1 j f int =k = k −  di 0 di  d0 di  di  i   1 1 = k  0 − ( X J − X i ) d   i di  Fig.1. Internal force acts to preserve the thickness of  1 1 the leaflet.The force at vertex i will be in the d  ( = k 0 −  x j − x i , y j − yi ) ……. (7) direction of di and its magnitude (positive or  i di  negative) will depend on the degree of change in the length of di.where k is the constant elastic coefficient. The formof (7) indicates that in the ith row of the stiffness C. Image Forcesmatrix, all elements are zero, except Kii and Kij To calculate image forces, we first need to define the potential surface. It is computed asIf we define H( i, j) = − G σ * ∇I 2  1 1  (i, j) ……. (8)   k i = k −  , then Kii=ki and Kij=-ki  d0 di   i  where I is the image intensity, Gσ is a two-Introduction of such a force prevents the opposite dimensional (2-D) Gaussian mask with standardsides of the contour from collapsing. After ∆ t 0 the deviation σ and * is a 2-D convolution. Thecontour is close to the leaflet, but has not reached it standard deviation of the Gaussian mask has to becompletely. At this time, image forces are introduced, chosen carefully. When the object of interest is veryand very often both sides of the contour can be small in at least one dimension (such as in heart valveattracted by the same side of the leaflet, i.e., be on the leaflets, which are very thin), too large value for σslopes of the same valley of the potential surface. may completely blur the edges of the object. We usedOnce opposite sides of the contour begin getting the value 1.2 for the mitral valve, and 1.8–1.9 forcloser, the elastic force pushes them apart and brings other image sequences of both a healthy and diseasedone side of the contour under the influence of the heart. These numbers were chosen experimentally.opposite edge of the leaflet. Components of the image force at each vertex are calculated as the negative of the partial derivatives ofFor the left ventricle and aortic root image sequences H in corresponding directionsused to test our algorithm, the elastic coefficient kwas set to zero. The velocity estimates seemed to be  ∂ [xi , y i ] H ; fi x =g − quite accurate and the problem of snaxel clustering  ∂x normwas solved by contour resampling in each iteration.The example of tracking the leaflet motion shows that
  5. 5.  ∂ [xi , y i ] H As the basis for this binarization decision f i =  g −  …….(9) y  ∂y  procedure was precisely the histogram, image gray norm levels are grouped into three different classes, accordingly to its brightness, the first one ranging from 0 to 49, the second one from 50 to 149 and thewhere index norm denotes that the variable is third from 150 to 255. Through histogram analysisnormalized to values from zero to one. The Sobel the system will then find out which of these classesoperator was used to calculate both H and F contain the largest set of points belonging to the.Parameter g is used to control the amplitude of image. Here, three distinct situations can arise. In theimage forces. first situation, when the largest set of pixels belongs to the first defined class [0 - 49], the image consists predominantly of dark tones, thus making it harder to find a good threshold value. For these cases, we developed a binarization algorithm that proved to be somehow very efficient. In this simple method, the threshold value is determined by averaging the histogram maximum gray level (level containing the most image points) with the medium histogram gray level (sum of every pixel’s level divided by the total number of image pixels). HIST(1) + HIST(2)........... + HIST( N ) MAX(HIST) + Threshold = N 2 Fig.2. 2-D echo images of Mitral valve From the formula above, it can be easily understood that for very dark pictures, where other algorithms MEASURING SYSTEM return a low threshold value that results in poor quality binarizations and subsequent loss of someAutomatic Global Binarization calibration target points (due to noise), this method settles a higher threshold value located between the As mentioned before, a reliable image maximum and the average histogram levels.binarization method was developed in order to On the other hand, if the largest set of pixelsprovide good quality binarized image (in real - time), belongs to the second defined class [50 - 149], thenwhich represent the most important requirement of the image is mostly composed by intermediate graythe template matching algorithm used in the system levels and therefore the best method to apply is theto detect the calibration target set of points. well known Otsu method (which works especially To achieve this goal, some known well if there are no major image brightnessbinarization methods were tested. However, as variations).expected, no method was found to be general enough Finally we reach the last case, where the– the pictures analyzed, mostly heart tissues, had largest set of pixels belongs to the third defined class,indeed very specific characteristics (in terms of ranging from 150 to 255. Here, the images are in factbrightness, contrast, and color balance) that changed too light and both methods described above revealgreatly from picture to picture, leading to a their lack of efficiency regarding to finding amiscellany of good and bad results for each method threshold value that maximizes binarization quality.applied. In this case it is used an algorithm introduced in To cope with this problem, the solution O’Gorman, which is also a global approach, but usesfound was to introduce a decision procedure based on a measure of local information, namely connectivity.the image histogram, capable of choosing the most This method maximizes connectivity of the resultingappropriate binarization method according to image thresholded regions and works exceptionally well ifproperties. It should be noticed that histograms, more than 50% of image gray levels belong to thealthough very simple concepts, encapsulate third defined class.information that turns out to be very helpful in the At this point, after choosing the binarizationstudy of certain cases. method according to the rules presented above and
  6. 6. performing that operation, some cleaning was found We have presented an active contour modelnecessary in order to remove isolated pixels. This for tracking objects in image sequences. The modelwas achieved through a morphological filter. successfully handles large frame to frame displacements of the object of interest. The presented RESULTS algorithm requires the user’s interaction only in the first frame of the sequence. Optical flow information is used to estimate the initial position of contour in the subsequent frames. We have also presented a software package to perform online measuring of mitral valve structures to assist heart surgery. The developed application was designed to capture and process mitral valve pictures. Fig. a Fig. b FUTURE WORK We intend to implement the same concept of active contours and generation of database on the Aortic valve. REFERENCES [1]I.Miki´c, S.Krucinski, and J.D.Thomas, Fig. c Fig. d “Segmentation and tracking of mitral valve leaflets in echocardiographic sequences: Active contours guided by optical flow estimates,” SPIE Med. Imag., vol. 2710, pp. 311–320, 1996. [2]T.McInerney and D. Terzopoulos, “Deformable models in medical image analysis: A survey,” Med. Image Anal., vol. 1, no. 2, pp. 91–108, 1996. [3]M.Kass, A.Witkin, and D.Terzopoulos, “Snakes: Active contour models,” in Proc. 1st Int. Conf. Computer Vision, 1987, pp. 259–268. [4]L.D.Cohen, “On active contours models and Fig. e balloons,” CVGIP: Image Understanding, vol. 53, no.2,pp.211–218,1991. [5]L.D.Cohen and I.Cohen,“A finite element method applied to new active contour models and 3-D reconstruction from cross sections,” in Proc. 3rd Int. Conf. Computer Vision,1990, pp. 587–591. [6]L. D. Cohen and I. Cohen, “Finite-element methods for active contour models and balloons for 2-D and 3-D images,” IEEE Trans. Pattern Anal. Machine Intell. vol. 15, Nov. 1993. [7]A.Amini,T.E.Weymouth, and R.C.Jain, “Using dynamic programming for solving variational problems in vision,” IEEE Trans. Pattern Anal. Fig. f Machine Intell. vol. 12, pp. 855–867, Sept,1990. [8]D. Geiger, A. Gupta, L. A. Costa, and J. Vlontzos, “Dynamic programming for detecting,Fig. a. Initial Contour. tracking, and matching deformable contours,” b. Contour moved to one step. IEEE Trans. Pattern Anal. Machine Intell. vol. 17, c. Contour moved to two steps. pp. 294–302, Mar. 1995. d. Contour moved to three steps. e. Tracking of mitral valve by active contour. f. Patient Case History Form. CONCLUSION