International Conference on Machine Intelligence, Tozeur – Tunisia, November 5-7, 2005 417

An Active Contour Model Based on Splines and

Separating Forces to Detect the Left Ventricle in
Scintigraphic Images

Najah Hraiech1, Daniel Weinland2, Kamel Hamrouni3

1
University of Mannheim, Institute for Computer Science V, Mannheim, Germany
[email protected]
2
INRIA: National Institute for Research in Computer Science and Control, Grenoble, France
[email protected]
3
ENIT: National Engineering School of Tunis, LSTS Laboratory, Tunis, Tunisia
[email protected]

Abstract—In this paper we present a novel method for the The paper is organized as follows: In the first part, we review
segmentation of scintigraphic images in order to detect the two the simplified diffusion snake as introduced in [1]. Then we
ventricles of the heart. The method is based on the active contour show how to integrate the separating forces. Finally we will
model using a region based approach and parametric give experimental results of our method in separating the two
representation of the contour. The novel aspect of the method is
ventricles of the heart, and compare our approach with one of
that we have integrated separating forces between the two
ventricles in the active contour model. The method has been the most widely used active contour approaches, the Chan and
applied on several scintigraphic images and compared to “active Vese model [2].
contour without edges” model. The results are satisfactory.
II. ACTIVE CONTOUR BASED ON SPLINES
The active contour model, or snake, is defined as an energy
I. INTRODUCTION minimizing functional, where the snake's energy depends on its
The need in the medical field of knowing how certain organs shape and location within the image. The goal of the active
of the body work, especially to diagnose and determine certain contour is to extract the boundaries of homogeneous regions
heart diseases, leads to scintigraphy. Scintigraphic imaging that constitute an image.
techniques can give images of any organ of the body. The organ The two most used active contours approaches are: Edge
to be investigated is injected by a radioactive solution. A special based segmentation approaches [3] and region based
camera measures the radiation emitted by a certain organ. One segmentation approaches [4, 5, 6]. Edge based active contour
application of these techniques is to capture the geometry of the approaches are based on gradient information to deform the
two heart ventricles. curve towards the edges of objects within an image. Region
Usually scintigraphic images are blurred with noise due to based active contours approaches include terms describing the
artifacts and noise from the gamma camera. As the intensity different regions by statistical spatial features, like mean and
contrast between the two ventricles and the myocardium is low, variance of color values. In our work we use a region based
the boundaries of the ventricles are often missing in certain active contour approach, because scintigraphic images are often
locations. Even for skilled personal the exact boundary blurred with noise, and edge based segmentation is known for
identification is difficult. Hence segmentation techniques can its sensitivity to noise.
help us to recognize the two ventricles.
A. Simplified diffusion snake
The aim of segmentation is to distinguish objects from
background to achieve homogeneous regions. It is widely used Our work is derived from the probably most widely studied
for medical image classification. region based Active Contour approach, which is that of
In this paper we will present a new approach to segment the Mumford and Shah [7]. Mumford and Shah proposed to
ventricles based on active contours. While existing active approximate a given input image f with a piecewise smooth
contour approaches can segment the heart in one “homogenous” function by minimizing the following functional:
region, most of them fail in determining the intermediate
E (u, C ) = (f − u )2 dx + λ2
boundary, which separates the two ventricles. 1 1 2
∇u dx + v C , (1)
For the evolution process of two active contours, centered on 2 Ω 2 Ω −C
the right and left ventricle, we demonstrate how by integrating
novel separating forces the segmentation can be efficiently where f is the input image, u is the segmented image, C the
improved leading to clearly separated ventricles. segmenting contour, and , v are regularization parameters.
 International Conference on Machine Intelligence, Tozeur – Tunisia, November 5-7, 2005 418

dC (s, t ) dE (u , C )
Constraining the segmented image to constant regions, u(x) =
ui, for x ∈ Ri , Ri ∈ Ω , the Munford-Shah functional (1) dt
=−
dC
[ ]
= e + (s, t ) − e − (s, t ) ⋅ n(s, t ) − vC ′′(s, t )
simplifies to the cartoon model as: ∀s. (6)
In order to obtain an evolution rule for the control points of
E (u , C ) = (f − ui ) dx + v Cs2 ds. (2)
1 2 1
the spline C we insert the spline curve (3) in equation (6) to
2 i
R
i 0
derive the following evolution equation for the control points:

dp i (t ) d 2 Bi (s )
As representation for the boundary function C we choose a
Bi (s ) = [e + (s, t ) − e − (s, t )]⋅ n(s, t ) + v p i (t )
N N
spline representation. .
i =1 dt i =1 ds 2
B. Spline representation (7)
The two common ways to represent the contour function are
parametric by spline curves, or implicit by the zero level set of a
D. Implementation
functional: Φ → ℜ , i.e. the intersection of Φ on the zero
Discretizing equation (7) with a set of nodes si along the
plane, see [8].
contour gives us a set of linear differential equations. The
In [2] and [9] T.F. Chan and L.A. Vese explain the
solution gives the temporal evolution for the coordinates of each
implementation of the Mumford-Shah functional using level
control point (xm, ym):
sets as a representation for the contours.
In our work, we choose a parametric representation based on
piecewise polynomial splines. This representation is especially
suitable for the integration of the separating forces, which are
dxm (t ) N −1
dt
= ( ) [( i
)
B mi es+ − es−i ⋅ n x + v(xi −1 − 2 xi + xi +1 ) , (8) ]
i =1
directly attached onto the control points of the splines.
For the experiments we used the closed, uniform, quadratic

( ) [(e ) ]
and periodic B-spline curve, for more detail see [10] and [11].
dym (t ) N −1
Using this spline approximation the contour C has the following = B mi
+
si
− es−i ⋅ n y + v( yi −1 − 2 yi + yi +1 ) . (9)
form: dt i =1

C : [0,1] → Ω, C (s ) = p n Bn (s ),
N
(3) The matrix B contains the spline basis functions evaluated at
n =1 the nodes si : Bi,j =Bi (sj ), ( si corresponds to the maximum of
Bi). Notice, that the matrix B is cyclic and tridiagonal.
where Bn are the uniform, periodic, quadratic B-spline basis The first term in the set of equations (8) and (9) maximizes
functions and pn = (xn, yn)T denote the control points. the homogeneity in the adjoining regions as measured by the
energy densities (5). The homogeneity in the adjoining regions
C. Curve Evolution forces the contour to recognize the boundaries of the object. The
We minimize the energies of the simplified diffusion snake second term smoothes the contour and enforces an equidistant
(2), using gradient descent with respect to both, the segmenting spacing between successive control points.
contour C and the segmented image u, what leads to the In our implementation we need at the beginning the
following Euler-Lagrange equation: initialization of two set of control points (each set corresponds
to one ventricle), then we interpolate them by the spline
dE
dC
[ ]
= e − (s ) − e + (s ) ⋅ n(s ) − vC ′′(s ) = 0 ∀s ∈ [0,1], (4)
contour. After that, we implement the evolution of this curve as
explained in equations (8) and (9).
Note, that with each new iteration, we have first to calculate
the new control points (the previous points + the evolution) and
where n is the outer normal vector on the contour for each then calculate new values for e+ and e- depending on the new
control point. The terms e+ and e- denote the energy density splines. These two steps are iterated until the difference between
outside and inside the contour C(s) respectively, see [12], the subsequence evolutions is certainly small. For equation (5)
defined by: we take the difference between inside mean and value of the
pixel lying on the inside pointing unit normal of the control; and
+
e − = ( f − u )2 . (5) outside respectively.
E. Discussion
We solve the minimization problem by gradient descent. The Applying the described actives contours to sample images we
result is summarized in equation (6): realize that the results depend on the initialization: when we
have a good initialization the active contour recognizes the two
ventricles. With random initialization the results were not
 International Conference on Machine Intelligence, Tozeur – Tunisia, November 5-7, 2005 419
acceptable as the two ventricles are often not separated, so that Fj is the force from all control points i onto the control
illustrated in figure 1 and figure 2. point j , i ∈ C1 and j ∈ C2 .
In the next section we will integrate separating forces in the The variance σ controls the region of influence of the forces.
curve evolution in order to prevent this effect.
Furthermore, p j − pi 2
is the Euclidean distance between
two controls points belonging to different curves.
Adding the forces (10) and (11) into equations (8) and (9)
leads to the following temporal evolution equation for the
coordinates of each control point (xm, ym):

dxm (t ) N −1
dt
= ( ) [(e
B mi
+
si
) ]
− e s−i ⋅ n x + v(xi −1 − 2 xi + xi +1 ) − h ⋅ Fi ⋅ n x , (12)
i =1
Figure 1. Result of segmentation using Active Contour based on Splines.

( ) [(e ) ]
False initialisation leads to poor final results: (left) original image, (middle)
dy m (t ) N −1
evolution of spline curves, (right) the two curves merge before the splines find = B mi
+
si
− es−i ⋅ n y + v( yi −1 − 2 yi + yi +1 ) − h ⋅ Fi ⋅ n y . (13)
the true boundaries. dt i =1

Regularization parameter h>0 determines the trade-off between

the segmenting and the separating forces.
IV. EXPERIENTAL RESULTS
To evaluate our work, we test our implementation with
scintigraphic images of different quality. In our implementation
we manually initialize the splines. Hereby the number of control
Figure 2. Result of segmentation using Active Contours based on Splines points is of importance (see below), while small variations in
using good initialization: (left) original image, (middle) evolution of spline
the initial shape did not affect the results. For the regularization
curves, (right) the segmented images.
parameters we use an empirical choice.
For most of the images, the tendency of the two splines to
III. INTEGRATION OF SEPARATING FORCES BETWEEN THE merge (figure 1) was successfully suppressed and the evolution
TWO VENTRICLES of the contours stabilizes around the true boundary of the
Usually active contours based on splines are a suitable model ventricles (figure 3).
to recognize boundaries within an image. Nevertheless, using
one active contour for every ventricle, experiments showed their
tendency to merge ignoring the separating boundary. Therefore,
we suggest introducing additional forces into the curve
evolution in order to keep the two splines separated from each
other.
These forces depend on the distance between the control
points on the two contours and are defined as:

2
N pi − p j
Fi = exp − 2
, (10)
j =1 σ

so that Fi is the force from all control points j onto the control
point i , i ∈ C1 and j ∈ C2 . Figure 3. Results using integration of Active Contours based on Splines and
forces between the regions of interest: (left) the initialization of spline,
And, vice versa: (middle) the evolution of spline curves, (right) the segmented images.

Figure 4 shows a comparison of our results with that derived

2
N p j − pi from the Chan and Vese approach [2]. Our method showed to
Fj = exp − 2
, (11) be superior in separating the two ventricles (e.g. figure 4 top
i =1 σ row), but has limitations depending on the number of control
points (figure 5).
The limitation for the control points is demonstrated as
follows: We start from a small contour and the contour grows to
 International Conference on Machine Intelligence, Tozeur – Tunisia, November 5-7, 2005 420
cover the boundaries, but has not a sufficient number of control implement a region-based active contour segmentation
points to describe the exact shape of the ventricles (figure 5). algorithm for scintigraphic images.
For further work we suggest to use a dynamical number of We conclude that the integration of separating forces into
control points, to circumvent the later problem. active contours model improves the segmentation of the two
ventricles, compared to common active contour approaches.
To show the performance of our results, we compared it with
the "Active Contour without edges" approach [2].
ACKNOWLEDGMENT
This work was partially done while visiting the CVGPR
group, University of Mannheim, Germany.
REFERENCES
[1] Daniel Cremers, Christoph Schnoerr, and Joachim Weickert. “Diffusion-
snakes: Combining Statistical Shape Knowledge and Image Information
in a Variational Framework”. IEEE Workshop on Variational and Level
Set Methods, Vancouver, CA, July. 13, 2001, pp. 137-144 IEEE.
[2] Tony F. Chan and Luminita A. Vese. “Active Contours Without Edges”.
IEEE Transactions on image processing, Vol.10, No. 2, February 2001.
[3] M. Kass, A. Witkin, and D. Terzopoulos. “Snakes: Active contours
models”. International Journal of Computer Vision, pages 321–331,
1988.
Figure 4. Results of segmentation of the image of the heart: (left) original [4] S. Zhu and A. Yuille, “Region competition: unifying snakes, region
images, (middle) segmentation with “Integration of separating forces to Active growing, and bayes/MDL for multiband image segmentation,” PAMI,
Contours based on Splines”, (right) segmentation of original images with vol. 18, pp. 884–900, september 1996.
“Active Contours Without Edges”. (top) the Chan and vese approach fails to [5] N. Paragios and R. Deriche, “Geodesic active regions: A new paradigm
separate the two ventrices. (bottom) variations in segmentation dependent on to deal with frame partition problems in computer vision,” JVCIR, vol.
contpour representation (spline vs. level-set) 13, pp. 249–268, 2002.
[6] Ronfard R. “Region-based strategies for active contour models“
International Journal of Computer Vision, Vol. 13, No. 2, pp. 229-251,
Oct. 1994.
[7] D. Mumford and J. Shah. “Boundary detection by minimizing
functionals”. In Proc. IEEE Conf. Computer Vision and Pattern
Recognition, 1985.
[8] J.A.Sethian, “Level Set Methods”, Cambridge University Press, 1996.
[9] Tony F. Chan and Luminita A. Vese. “A Level Set Algorithm for
Minimizing the Mumford-Shah functional in image processing”. In IEEE
Workshop on Variational and Level Set Methods (VLSM'01),
Vancouver, Canada, July 2001.
[10] Gerald Farin, “Curves and Surfaces for Computer-Aided Geometric
Design”. Academic Press, San Diego, 1997.
[11] R. H. Bartels, J. C. Beatty, and B. A. Barsky, “An introduction to Splines
Figure 5. Limitation of controlpoints: (left) the spline describing the boundary
for use in Computer Graphics and Geometric Modeling”, Morgan-
has not enough control points to cover the whole region, (right) true Kaufmann, Los Altos, Californie, 1987.
segmentation result derived with sufficient control points.
[12] D. Mumford and J. Shah. Optimal approximations by piecewise smooth
functions and associated variational problems. Comm. Pure Appl. Math.,
42:577-685, 1989.
V. CONCLUSION
In this paper, we proposed a new active contour method using
a uniform quadratic B-spline linked with extra forces to