file1

Document Sample
file1 Powered By Docstoc
					    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 : priya_04_n@yahoo.co.in                            Email: deepa_madathil@yahoo.co.in

                     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 be
designed 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 active
heart valve operations. The 2-D echo image is             contours and dimensions are measured from Global
obtained from the output console of the ultrasound        Image Binarization.
scanner. The heart valves in the image are                         We have also developed a Graphical User
segregated by processing the image using active           Interface using Visual Basic to record the suggestions
contours. The properties of that valve: the               given by the doctor for artificial heart valve
dimensional properties such as number of cusps,           replacement.
area of valve, radius of each cusp etc., and material
properties such as the elasticity, viscosity and                             BACKGROUND
chemical interactions with in the body are obtained.                The heart is the center of the cardiovascular
The 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 heart
properties of the artificial heart valves. When the       propels blood through thousands of miles of blood
properties 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 four
them with the properties of artificial heart valves in    compartments called chambers that receive the
the database using the pattern recognition principles.    circulating blood. The two superior chambers are
The matching of the properties in the system              called the right atrium and left atrium. The two
determines the selection of artificial heart valve with   inferior chambers are the right ventricle and left
which the valve of the patient has to be replaced. This   ventricle. As each chamber of the heart contracts, it
therefore forms the decision making system assisting      pushes a portion of blood into a ventricle or out of the
the doctor to identify the appropriate artificial heart   heart through an artery. To prevent back flow of
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 atria
last decade growing investment in the development         and ventricles. The right AV valve between the right
and improvement of computer vision techniques has         atrium and right ventricle is also called the tricuspid
been a common policy of companies and countries.          valve because it consists of three cusps (flaps). The
Such techniques rely on the unstoppable advance of        left AV between the left atrium and left ventricle has
technology 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 pulmonary
its ultimate goal, the creation of reliable systems       semi lunar valve lies in the opening where the
capable of taking the right automatic decisions           pulmonary trunk leaves the right ventricle. The aortic
supported by image analysis thus achieving results        valve is situated at the opening between the left
that 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 use
important fields of science is surely no exception and    as the replacement of diseased natural valves. The
the main problem that led to our work was related in      four natural valves (tricuspid, mitral, pulmonic and
this 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 congenital
automatically segment and measure the dimensions          defects. This results in the restriction of the forward
flow of blood known as stenosis or regurgitation flow    snaxels, connected by line segments. The position of
of blood known as insufficiency. The advent in           the snake is described by an M  2 matrix X, where the
science and technology has led to new and improved       ith row contains coordinates of the ith snaxel
designs of these heart valves. Heart valve prostheses
                                                                    x0      y0
may be classified into 2 types, mechanical valves and
                                                                    x1      y1
tissue valves. Mechanical valves are classified into 3        X=                   ……………………… (1)
categories, caged ball valves, tilting disk valves and              ....   ....
bileaflet valves. Tissue valves use materials of                   xM  1 yM  1
biological origin as the valving elements viz,
allograft valves, aortic valve xenograft and                       The contour motion is a result of interplay
pericardial 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 been
proposed. The observations for this Kalman filter are     where F is the matrix of x and y components of
optical flow estimates obtained by application of the     image forces at each vertex, where M = μ.Im, where μ
block-wise Horn and Schunk algorithm. With such          is the mass assigned to each vertex, C = γ.Im, where γ
observations, the tracker should be able to more         is the constant damping density and K is an MxM
successfully handle sudden movements of the object       stiffness matrix. Dots above the X represent the
of 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 solving
position and shape of the object are estimated by        the second-order differential equation. If the initial
integrating object motion. This results in               position and velocity are known, (2) has a unique
accumulation of errors in the optical flow estimates      solution. This suggests the way to incorporate the
over time. This model does not provide a way of          optical flow information into the active contour
incorporating the current image information into the     model. For the first frame of the image sequence, a
current 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 the
solved by using optical flow estimates as initial         first frame, the subsequent initializations are done
conditions when integrating the equation of motion.      automatically. The initial position is given by the
In other words, optical flow is used to push the          final position from the preceding frame, and the
contour toward the new position of the object.           initial velocities of vertices are given by the optical
In the simpler algorithm proposed by Ayache, an          flow estimates.
energy measure that tends to preserve the matching                 Optical flow estimates give the average
of high curvature points and to enforce a smooth field    velocity of the object between two frames, not the
of displacements vectors between contours in two         initial velocity. They can, however, be used as initial
consecutive frames is minimized (such an approach        velocities for two reasons. First, we are using a small
also requires previous edge detection). In the method    value for the damping factor, and second, since the
we propose, instead of matching pieces of contours       final contour position is adjusted through the action
around high curvature points, through the use of         of image forces, we only need rough estimates of
optical 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 a
next frame. The benefits of the latter approach are       multiscale version of Singh‟s algorithm, which is
notable in the high deformation frames (our              capable of detecting large displacements.
algorithm required user intervention in only one
frame 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:
   The equation of the dynamic                             
   equilibrium of the contour (2) can be
   integrated using a single-step time-                                M
                                                                      i  0 2 dFI dXi
                                                                                            ……………………… (6)
   marching scheme (SSTMS). In the
   SSTMS, the time t is divided into
                                                                      M            
                                                                     i  0 1 dFI dXI 0

   several intervals, or time steps  t i . In                                            .
                                                           where dF  M  C X  kX  F is the force residue
   each time step, the position and the                    and dX is the position increment. If  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 2
       n                                                   3) Calculation of the Contour Position: The time
                                                           interval between two successive frames is assumed to
             X..  2  X.n
              n
                                                           be one. This time interval is then divided into Nt
                                                           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 toward
ii)      M t C1
      n                   2 K 
                                 
                        2                                the new position of the object of interest, we chose a
                                                           nonuniform time division. The first interval  t0 is set
                      . 
 F  C X n  2  k X n 1
           ..
                                                           at 0.9 and the remaining 0.1 is divided into Nt-1
                                                           equal intervals  t n . During  t0 , image forces are
                                        t2                set to zero to prevent attraction of the contour by
iii) X n  2  X n  t X.n   n
                                        2                  undesired valleys of the potential surface. In other
           X.n  2    X.n   n t   ………… (3)             words, we are making the potential surface flat for
                                                           the first 90% of the time. During this time, elastic
where n is the time step number, X 0  X(0) and
                                                           forces act on the contour, filtering inaccuracies in
X.  X . 0  and 1 and  2 are constants. The
 0                                                         velocity estimates. During the remaining 10% of the
unconditional stability of the integration is              time, image forces are included, making the contour
guaranteed if 1  0.5 and  2  1 . Steps i)–iii) are    settle on the edges of the object of interest.
                                                              The contour is resampled after each iteration to
repeated for each of the time intervals. The problem
                                                           compensate for possible clustering of the snaxels in
is nonlinear if the stiffness matrix K is not constant
                                                           the local energy minima along the valley of the
and depends on the position of the contour. In such a
                                                           potential surface. After resampling, distances
case, the equation of motion takes a more general
                                                           between snaxels are equal. Since both positions and
form
                                                           velocities of points from one iteration are used in the
M 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 snaxels
interval                                                   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 Forces
                                                              In the case of the discrete contour, internal forces
of X inside the current interval. (4) has to be solved     depend on positions of vertices. Internal forces can
iteratively using steps ii) and iii) of the procedure      usually be decomposed into the product of the
given in (3). It will be shown that the stiffness matrix   stiffness matrix and the position matrix Fint=-K(X)X
in our model depends on the position of the contour.       Where such decomposition is not possible , the
However, the internal forces can be decomposed as          equation of motion takes on the form of (4).
P(X)=K(X)X,with stiffness matrix K recalculated at
the beginning of each iteration.                              Depending on the shape of the object of interest,
                                                           different internal forces can be designed. The role of
2) Equilibrium Criterion: The iteration process is         such forces is to enforce on the contour a shape
terminated when the contour attains the equilibrium        feature that is characteristic of the object. For
at each time step. In the proposed model, the ratio of     example, in the case of heart valve leaflets, assigning
contour deformation energy in the current and the          the contour elastic properties that try to preserve
first iteration was chosen as an equilibrium measure
length will result in failure of the active contour to       the developed approach is quite flexible. A priori
detect sudden length changes. However, the leaflet            knowledge about the shape characteristics of the
shape feature that seems to be preserved throughout          object of interest can be incorporated into the
the image sequence is leaflet thickness. This fact            definition of internal forces, and the rest of the
suggests that in the case of valve leaflets, additional       algorithm remains unchanged.
forces that act to preserve contour thickness should
be added.

   To define such forces, for each point i on the
contour, the distance di from the matching point on
the opposite side is calculated (Fig. 1). Also, the                                           i
distance d i0 calculated at the optimal position in the
first frame is stored and considered to be optimal.
The internal force that acts on the vertex i is then
defined as a force proportional to the relative change                                  di
in length d0

                di    di di
                        0        1 1                                           j
  f int    k                 k   d i
                      0          d 0 di 
                     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 direction
        d
                          
     = k 0   x j  x i , y j  y i
               
                                           ……. (7)          of di and its magnitude (positive or negative) will
         i di                                              depend on the degree of change in the length of d i.

where k is the constant elastic coefficient. The form        C. Image Forces
of (7) indicates that in the ith row of the stiffness                 To calculate image forces, we first need to
matrix, all elements are zero, except Kii and Kij            define the potential surface. It is computed as

                                                               Hi, j   G * I (i, j)
                                                                                   2
If we define                                                                                 ……. (8)
                                                                                    
         1 1
 ki  k  0   , then Kii=ki and Kij=-ki
        d                                                  where I is the image intensity, G is a two-
         i di 
                                                             dimensional (2-D) Gaussian mask with standard
Introduction of such a force prevents the opposite
                                                             deviation  and * is a 2-D convolution. The
sides of the contour from collapsing. After       t 0 the   standard deviation of the Gaussian mask has to be
contour is close to the leaflet, but has not reached it       chosen carefully. When the object of interest is very
completely. At this time, image forces are introduced,       small in at least one dimension (such as in heart valve
and very often both sides of the contour can be              leaflets, which are very thin), too large value for 
attracted by the same side of the leaflet, i.e., be on the    may completely blur the edges of the object. We used
slopes of the same valley of the potential surface.          the value 1.2 for the mitral valve, and 1.8–1.9 for
Once opposite sides of the contour begin getting             other image sequences of both a healthy and diseased
closer, the elastic force pushes them apart and brings       heart. These numbers were chosen experimentally.
one side of the contour under the influence of the            Components of the image force at each vertex are
opposite edge of the leaflet.                                 calculated as the negative of the partial derivatives of
                                                             H in corresponding directions
For the left ventricle and aortic root image sequences
used to test our algorithm, the elastic coefficient k                     H xi ,           yi   
                                                                fi
                                                                 x    g                               ;
was set to zero. The velocity estimates seemed to be                         x                    norm
                                                                          H xi , yi 
quite accurate and the problem of snaxel clustering                    
was solved by contour resampling in each iteration.           fi
                                                               y    g                     …….(9)
                                                                            y         norm
The example of tracking the leaflet motion shows that
where index norm denotes that the variable is            third from 150 to 255. Through histogram analysis
normalized to values from zero to one. The Sobel         the system will then find out which of these classes
operator 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 the
image 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
             MEASURING SYSTEM                            that for very dark pictures, where other algorithms
                                                         return a low threshold value that results in poor
Automatic Global Binarization                            quality binarizations and subsequent loss of some
                                                         calibration target points (due to noise), this method
          As mentioned before, a reliable image          settles a higher threshold value located between the
binarization method was developed in order to            maximum and the average histogram levels.
provide good quality binarized image (in real - time),             On the other hand, if the largest set of pixels
which represent the most important requirement of        belongs to the second defined class [50 - 149], then
the template matching algorithm used in the system       the image is mostly composed by intermediate gray
to detect the calibration target set of points.          levels and therefore the best method to apply is the
          To achieve this goal, some known               well known Otsu method (which works especially
binarization methods were tested. However, as            well if there are no major image brightness
expected, no method was found to be general enough       variations).
– the pictures analyzed, mostly heart tissues, had                 Finally we reach the last case, where the
indeed very specific characteristics (in terms of        largest set of pixels belongs to the third defined class,
brightness, contrast, and color balance) that changed    ranging from 150 to 255. Here, the images are in fact
greatly from picture to picture, leading to a            too light and both methods described above reveal
miscellany of good and bad results for each method       their lack of efficiency regarding to finding a
applied.                                                 threshold value that maximizes binarization quality.
          To cope with this problem, the solution        In this case it is used an algorithm introduced in
found was to introduce a decision procedure based on     O‟Gorman, which is also a global approach, but uses
the image histogram, capable of choosing the most        a measure of local information, namely connectivity.
appropriate binarization method according to image       This method maximizes connectivity of the resulting
properties. It should be noticed that histograms,        thresholded regions and works exceptionally well if
although very simple concepts, encapsulate               more than 50% of image gray levels belong to the
information that turns out to be very helpful in the     third defined class.
study of certain cases.                                            At this point, after choosing the binarization
          As the basis for this binarization decision    method according to the rules presented above and
procedure was precisely the histogram, image gray        performing that operation, some cleaning was found
levels are grouped into three different classes,         necessary in order to remove isolated pixels. This
accordingly to its brightness, the first one ranging     was achieved through a morphological filter.
from 0 to 49, the second one from 50 to 149 and the
                      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
          We have presented an active contour model
for tracking objects in image sequences. The model
successfully handles large frame to frame
displacements of the object of interest. The presented

				
DOCUMENT INFO
Shared By:
Categories:
Stats:
views:4
posted:3/3/2011
language:English
pages:6