56

CHAPTER 4

SEGMENTATION USING ACTIVE CONTOUR

MODEL APPROACH

4.1 INTRODUCTION

In this chapter an active contour model or snake model based on the

techniques of curve evolution and improved GVF methods are presented.
There has been a substantial research on segmentation of images with
deformable models Jain et al (1998). Snakes, or active contours, are used
extensively in computer vision and image processing applications, Bamford
and Lovell (1998 ) particularly to locate object boundaries. However, most
present snake models cannot provide better capture range McInerney et al
(1996) and evolution-stop mechanism. This chapter presents a new external
force for active contours, largely solving both problems. An extension of the
gradient vector flow snake (GVF snake) method is presented. First, the
adaptive balloon force Cohen et al (1989) is developed to increase the GVF
snake’s capture range and convergence speed. Then, a dynamic GVF force is
introduced to provide an efficient evolution-stop mechanism. In this way, the
snake from breaking through the correct surface and locking to other salient
feature points are prevented. The active contour models have been applied on
X-ray coronary angiogram images. The segmentation results demonstrate the
potential of improved GVF method in comparison with all previous active
contour methods. Performance analysis is performed and results are compared
with all active contour models.
 57

4.2 TRADITIONAL SNAKES

In 2D, a snake is a curve C(s) = (X(s), Y(s)) where S ε [0, 1]. The
curve moves through the image domain to minimize a specified energy
function. In traditional snakes, the energy is usually formed by internal forces
and external forces is described as

Esnake = Einternal + Eexternal (4.1)

Einternal tends to elastically hold the curve together (elasticity forces)

and to keep it from bending too much (bending forces). This energy is defined
where Cs and Css represent the first and second derivative, respectively. The
snake’s tension and rigidity can be controlled by the coefficients α and β.

Einternal = ½∫sα│Cs│2ds + ½∫s β│Css│2ds (4.2)

Einternal intends to pull or push the curve towards the edges.

Typically, the external forces consist of potential forces. This energy is
defined where Eimage represents the negative gradient of a potential function.
This energy is generally the image force where I denotes the image and X =
X(x,y) = [x,y]t

Eexternal = ∫s Eimage(C(s))ds (4.3)

Eimage (X) = −│▼I(X) │2 (4.4)

Using variational calculus and the Euler-Lagrange the differential

equation can be solved. Then, the solution of this force balance represents the
snake final position. The differences in the ways the energy function is
established will result in different snakes.

αCs − βCss − ▼Eimage = 0 (4.5)

 58

Although the traditional snakes have found many applications, they

are intrinsically weak in four main aspects.

1. They are very sensitive to parameters.

2. They have small capture range.

3. They have difficulties in progressing onto boundary

concavities.
4. The convergence of the algorithm is mostly dependent of the
initial position.

4.3 THE LEVEL-SET METHOD

When evolving a curve C in R2, the standard approach is to use the

level-set method or level-sets. The idea of level-sets was first discovered by
(Osher and Sethian 1988). It exploits the remarkable fact that treating a curve
C(s) as a zero level curve of a surface (x, y) simplifies various tasks when
evolving C, such as changes in its topological structure. This is suited well for
image processing partial differential equation (PDE) problems, since the
image itself can be considered as a surface u(x) with intensity values at
position x.

The following is a basic result in level-sets: Proposition 1 (Curve

evolution for level-sets) Let C(t0) the zero level curve of  (x, t0) at some time
t = t0. If x0 2 C(t0) and v(x0 ,t0) denotes the outer normal velocity of C(t0), then
it solves the PDE as

t = v(x, t) | ▼ | (4.6)

The widely used type of curve evolution is motion by mean

curvature, when v(x) = (x), the curvature of C at x is
 59

t = k(x)| | ▼ | = |▼|▼ · (▼ /▼ ) (4.7)

Denote b for the correct boundary between 1 and 2 and define an

edge function g as, g(0) = 1, decreasing and lim s-> g(s) = 0. An example of
such function is g(|▼u 0|) = 1/( 1 + |▼Jp*u0|2 ) where Jp*u02 is some
regularization of image u0 to a smoother function. The flow chart in
Figure 4.1 for level set method is as follows:

Input coronary angiogram frame

Image Pre-processing

Clicking circle center and radius

Snake Formation

1. Energy calculation
2. Edge stopping function

Display the Segmented image

Figure 4.1 Flow chart of level set method

 60

4.4 GRADIENT VECTOR FLOW METHOD

Xu et al (1998) proposed new GVF snake to achieve better object

segmentation. The basic idea of the snake is to extend influence range of
image force to a larger area by generating a GVF field. The GVF field is
computed from the image. GVF field is defined as a vector field V = V(x) that
minimizes the energy function

Q = ∫∫ µ▼ V + │▼f│2│V − ▼f │2 dx (4.8)

where f is the edge map which is derived by using an edge detector on the
original image convoluted with a Gaussian kernel, and µ is a regularization
parameter. Using variational calculus, the GVF field can be obtained by
solving the corresponding Euler-Lagrange equations.

The force balance equation of GVF snake can be expressed as

αCss − β Csss + γ V = 0 (4.9)

where γ is a proportional coefficient. GVF snake’s larger capture range and

concavity tracking ability are attributed to the diffusion operation. When
│▼f│ is small, the energy is dominated by the sum of the squares of the
partial derivatives of the vector field, resulting in a slowly varying yet large
coverage field. Whereas when │ ▼f │ is large, the second term dominates
the integral. Figure 4.2 depicts the flow chart of GVF method.
 61

Input coronary angiogram frame

Image pre-processing

Initialization of boundary conditions

Calculating the gradient of edge map

Initialize GVF to the gradient

Display the segmented image

Figure 4.2 Flow chart of GVF method

There are certain drawbacks in this method also. The capture range
of the active contour did not seem as large as expected. Active contour does
not converge to the ideal solution. GVF snake was sensitive to the shape
irregularities.
 62

4.5 IMPROVED GVF METHOD

Suhuai Luo et al (2006) proposed an improved GVF method which

adds new external forces, including an adaptive balloon force and a dynamic
GVF force. They also presented new scheme to integrate these external forces
in the snake mathematic formulation.

4.5.1 Adaptive balloon force

In the balloon model (Cohen et al 1989 and 1993) a pressure force

fp(s) is added to snake force as a second external force to push the curve
outward or inward. In this way, the curve is considered as a balloon that has
been inflated or deflated. The pressure force, where n(s) is the normal unit
vector to the curve at point C(s) in

Fp(s) = k.n(s) (4.10)

The balloon force is considered to increase the image potential

force capture range. This is a proper consideration given that the snake can be
set to start evolving inside the object. Unfortunately, balloon force introduces
unpredictability to the performance of the active contour and makes it more
sensitive to the value of its different parameters.

To overcome the unpredictability problem introduced by the

balloon force, this force is applied in an adaptive way. The main idea is to
give the balloon force bigger weight compared to the GVF force at the early
stage of the evolution, and to give the balloon force smaller weight at the later
stage. In this way, the speed of the convergence is increased, and the snake
can be correctly pushed toward the surface even if it starts far away with less
chance of being over-pushed.
 63

4.5.2 Dynamic GVF force

A dynamic GVF force is introduced to provide a unique evolution-

stop mechanism as well as all the characteristics owned by the original GVF
force. The evolution-stop mechanism is needed to prevent the snake from
breaking through the correct contour and locking to other feature points.
The breakage can happen in areas where two objects or organs are very close
to each other. The introduction of the dynamic GVF force is inspired by a
property of the GVF field. That is, when the GVF field passes a contour, its
direction will change.

It can be easily observed that the field vector changes direction at

the ellipse boundary. Therefore, a consistency degree is incorporated into the
new dynamic GVF force. The force varies according to the consistency. If the
evolution of the snake will cause the change of GVF force direction, it is said
that inconsistency has occurred and the snake is not allowed to evolve to the
new position.

4.5.3 Proposed Work

With these two novel inclusions, the proposed force balance

equation can be expressed as

αCss − β Csss + γ Vdyn + λfpa = 0 (4.11)

Vdyn is defined as a dynamic gradient vector flow force. Let X1 be

a point on the current snake and X2 its possible next position in the evolution
process. Cθ defines the consistency angle and is proportional to the angle
between the GVF vectors at X1 and X2. Tθ representing the cut-off angle
based in this experiment, Tθ = 20 found to be a good threshold.
 64

Vdyn = v if Cθ > Tθ (4.12)

Vdyn = (−αCss + β Csss + λfpa )/γ other wise (4.13)

The new dynamic gradient vector flow force will be the same as
conventional GVF if the snake point moves towards the contour. But when
the snake point tries to cross over an edge, the dynamic gradient vector flow
force will stop the point from moving. The threshold Tθ will decide when this
evolution-stop mechanism is triggered. Figure 4.3 shows the flow chart for
the improved GVF method.

Input coronary angiogram frame

1.Selection of control points

2. Drawing the contour

Energy calculation

Display the Segmented image

Automated Stopping Algorithm

Minimization of energy function

Figure 4.3 Flow Chart of Improved GVF Method

4.6 THE ACTIVE CONTOUR MODEL- ALGORITHM

The Active Contour system is especially designed for detecting

blood vessels on coronary angiogram images for medical applications
 65

(McInerney and Terzopoulos 1995, Xu et al 1999). The environment where

these blood vessel structures are found contains materials such as segmented
tissues and fluid debris. These become the unwanted background information
that hinders the performance of the active contour in tracking the object
outline accurately. In addition, angiogram images also contribute various
obstructions that reduce the performance of the active contour such as color,
noise, low contrast and poor quality images. Therefore, the Active Contour
System developed in this chapter aims at optimizing the performance of
object tracking by introducing various image processing techniques.
Figures 4.1 to Figure 4.3 illustrate the flow chart for various methods.

An active contour is essentially a curve made up of various

energies. The curve deforms dynamically to mould to the shape of a targeted
object. There are various methods for implementing an algorithm to achieve
object outlining. Traditionally, the active contour algorithm requires an initial
user approximation surrounding the interested object and then deforms this
curve by minimizing the energies contained to achieve the object outline
matching as shown in Figure 4.4 and Figure 4.5.

Figure 4.4 Contour Figure 4.5 Final contour

enclosing the object latching onto the Object outline
 66

The energies in the active contour can be divided into two

categories. Internal and External energy functions which have the relationship
are discussed. The Internal energy functions focus on the intrinsic properties
of the contour such as elasticity and curvature, while the external energy
functions are related to the image properties like contrast and brightness.

Esnake = Einternal + Eexternal (4.14)

These energy functions are then given various weightings to control

the different properties of the contour which are crucial for a good match of
object outline. These weightings have the effect of controlling the rate at
which the contour minimizes and most importantly stop the contour from
further shrinking once the outline of the object is tracked. This is known as
the final approximation of the contour.

Energies contained in the contour, are minimized due to simulated

forces acting upon it. When all the forces involved are at equilibrium, the
contour has essentially reached its minimal state and the outcome of the curve
is the final approximation of the object outline. A detailed discussion of the
energies and forces of the active contour is provided in the following sections.

4.6.1 Initial User Approximation

The Active Contour algorithm begins with requiring the user to

construct a rough contour enclosing the object to be tracked. This contour is
known as the initial user approximation and can be of any shape and size.
Instead of having the user draw the curve out explicitly, the active contour
algorithm simplifies the procedure by allowing the user to click points that
surround the object. These points are the basic formation of the active
contour. The red points in Figure 4.6 are user control input points. In most
 67

cases, joining the user-clicked points forms a rigid contour with sharp corners.
Essentially a smooth contour is preferred because having a smoother contour
ensures the forces acting on each individual point being evenly distributed.
This will prevent large movements of the contour which produces a higher
risk of the contour moving inside the boundary of the target object.

The solution to this problem comes in two parts:

1. Point Stuffing
2. Spline function

The initial step to achieve a smoother contour was to ensure it was

made up of an adequate amount of points. The idea of point stuffing was to
introduce more points into the contour in addition to the number of user input
points. This method puts no restriction on the minimum number of points that
the user must input while ensuring that the contour has enough points to form
a smooth curve. There is no set rule as to how many points must be stuffed
between two user input points to create a smooth contour. However, the
optimum point stuffing after much experimenting was calculated to be adding
one point every 7.5 pixels.

To further ensure that a smoother curve is created, a different

approach is taken to connect all the points together to form the contour. Spline
function using parabolic approximation is used.
 68

Figure 4.6 Spline Function using Parabolic Approximation Between

Contour Points

Figure 4.7 shows an initial approximation of the contour with point

stuffing the spline function included in the algorithm. Red dots represent user
enter points and asterisk in magenta are from point stuffing

Figure 4.7 Initial Contour with Point Stuffing

4.6.2 Internal Energy

The Internal energy of the active contour controls the shaping of the
curve that eventually models the area of interest on a given image.
 69

It also controls the spacing between the control points that form the curve.
The properties of the Internal energy can be further subdivided into two parts
with individual weightings to influence the motion of the contour:

Einternal = α * Eelastic + β * E bending (4.15)

The Eelastic is the elastic energy, which enforces the contraction of

the contour like an elastic band while Ebending introduces a resistance to
bending like a metal strip trying to a smoother curve.

4.6.2.1 Elastic Energy

The elastic energy models the elastic band characteristic of the

active contour. The reason for having an elastic band characteristic is to allow
the curve to shrink to capture the shape of the object. Since the final contour
is always smaller than the initial approximation, the shrinking is essentially
minimizing the elastic energy contained in the curve. Another way of
visualizing this energy minimization is to imagine that the contour behave
like a spring obeying Hook’s Law. Initially the elastic energy of the contour is
defined by the expression

N
E(elastic )  K1  ( xi  xi 1 ) 2  ( yi  yi1 ) 2  (4.16)
 
i 1

where N is the number of points used to form the contour and i is the index of
the current point the calculation is focused on. K1 is a constant defined by the
user, which has the effect similar to the constant corresponding to the strength
of the spring. The shrinking effect of the curve is modeled by simulating
forces acting on the contour points as pulling them together until all points on
 70

the curve are approximately equally distributed. This is effectively applying

Pythagoras rule on each point of the curve. The forces, derived from the
elastic energy equation, acting on a particular point in the curve are expressed

Fe xi    (x i 1  x i )  (x i 1  x i ) 

Fe yi   (y i 1  y i )  (y i 1  yi )  (4.17)

where α is the arbitrary constant related to K1 in the elastic energy expression.

Figure 4.8 shows the effect of the simulated forces acting on a

control point in the contour and the movement of the point due to these
forces.

Figure 4.8 Illustration of Elastic force acting on contour point

The forces in both x and y directions involved in the calculation

differ according to the distances between points. Points having larger distance
between them move at a faster pace than closely distributed points.
 71

4.6.2.2 Bending Energy

The contour adjusts itself to the shape of the interested object on

the image . So having a circular shaped contour as final approximation is
inadequate. In addition to the contracting ability, the contour must somehow
be able to have different curvatures at different points. Therefore, bending
energy is introduced to allow the contour to have curvature properties.
The bending energy models the contour with a behavior much like a thin
metal strip. The behavior of this energy is not too much to do with contracting
the size of the contour; rather it attempts to straighten the bending of the
curve. The contribution of this energy is to smooth out the sharp angles in the
contour. Similar to the elastic energy, the calculation of bending energy is
applied to every point on the contour. The bending energy equation is
expressed as

(4.18)

where K2 is the user defined constant and i is the entry to the index of current
point on the curve. This time the energy function on a particular point is
dependent on its 4 nearest neighbors. Therefore, the force derived on a
particular point is calculated as follows,

(4.19)

where β is the arbitrary constant related to K2 shown in the bending energy

expression. Unlike the elastic force, the bending force tends to smooth out the
sharp angles while not shrinking the contour. Therefore, having more
 72

neighborhood points involved in calculating the bending energy helps to

smooth out the forces acting on the current contour point, producing a
bending effect rather than a contraction. Figure 4.9 demonstrates the concept
of bending energy and the forces acting on contour points.

Figure 4.9 Illustration of Bending Energy concept

The reason for involving more than one point in the calculation of
one contour point’s motion is to smooth out the forces on that contour point
producing a bending characteristic rather than contraction. Having just the
bending force acting on the contour will produce a final approximation of a
curve resembling the shape of the initial approximation yet with smoother
corners.

4.6.3 External Energy

The external energy of active contour depends on factors related to

the image properties rather than the contour itself. Attributes such as noise
level, brightness and contrast impose constraints that affect the motion but not
the shape of the curve. Therefore, external energy exerts attraction on the
contour towards interesting features such as object boundaries in an image.
 73

4.6.3.1 Image Gradient Energy

The image gradient energy exerts forces to move the contour

depending on the different grey scale levels that define the image. Suppose
the contour is to move towards the brighter side of the image, the energy
definition would be as

(4.20)

where K3 is a constant and image ( xi , yi ) and defines the intensity value of

the pixel at the ith point of the contour. The farther the contour moves towards
the brighter regions, the more negative the function will be, thus, the more the
energy is minimized. The movement of each user contour point is dependent
on its neighborhood pixels’ intensity values. Figure 4.10 demonstrates the
concept of contour point movement under the influence of image gradient
force which is expressed as

(4.21)

where γ is related to the constant K3. Having γ as K3/2 results in equal

distribution of control values for the Image Gradient Force in both x and y
directions as shown in Figure 4.10.
 74

Figure 4.10 Illustration of the Image Gradient Energy concept

The overall effect the image gradient force has on the contour is to
move the contour without deforming the shape of the curve. The movement of
the contour towards brighter or darker regions of the image is dependent on
the definition of the image gradient energy.

Suppose the aim of the contour is to move to dark region, the

energy expression should be re-defined as

(4.22)
and the forces will be adjusted accordingly with γ set as a negative value.

4.6.4 The Balloon Force

The Active Contour System currently requires the user to construct

an initial contour enclosing the object of interest. This contour then shrinks to
fit the size of the targeted object boundary. This creates a restriction that the
initial contour must enclose the entire object, otherwise any part of the initial
contour is formed within the targeted object will lead to the failure in tracking
 75

the object outline properly. To prevent such a situation from happening, the
balloon force is introduced to the active contour algorithm. It allows the
contour to adapt to the expanding and shrinking properties of a balloon. The
direction (i.e. expanding or contracting) of the balloon force is dependent on
two aspects:

1. tangent vector of the initial contour

2. unit normal to the tangent vector

The direction of tangent vector is dependent on the direction the

contour points are first laid out to form the initial contour. A clockwise
plotting would result in a clockwise tangent vector and vice versa for an anti-
clockwise initial plotting contour. The expression of the tangent vector
calculation is:

(4.23)

where ti is the tangent vector at contour point entry i. The unit normal (ni) for
each individual contour point can be easily calculated by rotating the tangent
vector at that contour point by 90°. This results in an expression for the
balloon force as follows which is described

Fballon = λn i (4.24)

where λ represents the weighting of the balloon force similar to the , , and 
arbitrary constants for the elastic, bending and image gradient force, respectively.
The effect of having the balloon force integrated into the existing active contour
algorithm is shown in Figure 4.11 final approximation of object.
 76

Figure 4.11 The performance of the active contour under the influence
of the balloon force

Notice that the location of the final contour is now within the
boundary of the targeted object and not surrounding the outline of the object.
The final contour also shows the inadequacy of the curve in tracking corner
areas. This leads to the need for further investigation into the balloon force
before fully integrating it into the current active contour algorithm.

4.6.5 Dynamic GVF force

A dynamic GVF force is introduced to provide a unique evolution-

stop mechanism as well as all the characteristics owned by the original GVF
force. The evolution-stop mechanism is needed to prevent the snake from
breaking through the correct contour and locking to other feature points.
The breakage can happen in areas where two objects or organs are very close
to each other. The introduction of the dynamic GVF force is inspired by a
property of the GVF field. That is, when the GVF field passes a contour, its
direction will change.

It can be easily observed that the field vector changes direction at the
ellipse boundary. Therefore, a consistency degree is incorporated into the new
dynamic GVF force. The force varies according to the consistency. If the evolution
 77

of the snake causes the change of GVF force direction, it is said that inconsistency
has occurred and the snake is not allowed to evolve to the new position.

4.6.6 Integration of all Energies

All energies contribute different characteristics to the active contour.

Having individual forces acting on the contour separately fails to produce the
desired effect of tracking objects in images. However, combining the effect of
these energies creates an active contour behaviour that is promising.

Although the contour behaves in a somewhat expected way, the

tracking is hardly acceptable. The characteristics of the contour are defined by
the combining of internal and external energies. However, the α, β and γ
weightings control the amount each energy characteristic effects the contour.
These control values must be fine tuned to improve the active contour’s
tracking of an object’s outlines. Apart from the active contour weightings that
control the behavior of the curve, the properties of an image are also
significant in controlling the action of the curve.

4.6.7 Control values of the Active Contour

The movement of the contour to track object boundaries is through

minimization of these energies contained in the curve. The motion of the
contour is due to the three forces simulated on the curve, namely the elastic,
bending and image gradient forces. Each of these forces corresponds to a user
defined arbitrary constant α, β and γ, respectively and these constants control
the strength of the forces acting on the curve. By assigning more weighting
towards a particular force, the contour is biased towards that force being
dominant throughout the movement of the contour. This creates undesirable
contour behavior as the shape of an object can have combinational features.
Having the elastic force being the dominant force on the contour will produce
an undesirable effect.
 78

Basically the contour would disregard the existence of the other

two forces and only have the elastic force characteristic. Even though there
might be the presence of other force’s characteristics, these only last for a
very short period of time or are very unnoticeable. This indicates an
equilibrium point among the arbitrary constants which are achieved to
produce an active contour that has a good behaviour in the three aspects of a
contour.

The values of the arbitrary constants found after various

experiments that produce a well behaved contour are
1. Alpha = 0.15
2. Beta = 0.005
3. Gamma = 0.0043

For level set, various parameters like 1 = 2 = 1, V=0 and the

approximation H2,, 2,  of Heavside and dirac delta functions (Tony et al
2001) are used in order to automatically detect interior contour and to ensure
the computation of a global minimum.

4.7 RESULTS AND DISCUSSION

Active contour models were applied to several normal and

abnormal images. For each input angiogram frame the improved GVF method
of segmentation is compared with all active contour models. Then the
iteration vs energy graph and comparison of all energy functions graphs for
all active contour models were plotted. The level set Method, GVF Method,
Dynamic GVF Method, Balloon Force Method and Improved GVF Methods
were applied to real normal and abnormal X-ray coronary angiogram images.
The segmented images for normal and abnormal are shown in Figures 4.12 to
Figure 4.15. The corresponding energy graphs are also shown for both normal
and abnormal images.
 79

Normal x-ray coronary angiogram – Image 1

(a) (b)

(c) (d)

(e) (f)

Figure 4.12 Segmented output for normal coronary angiogram - Image 1

(a) Input image (b)Level set Image (c) GVF method
(d) Dynamic GVF method (e) Balloon force method
(f) Improved GVF method
 80

Normal x-ray coronary angiogram – Image 2

(a) (b)

(c) (d)

(e) (f)

Figure 4.13 Segmented output for normal coronary angiogram – Image 2

(a) Input image (b) Level set image (c) GVF method
(d) Dynamic GVF method ( e) Balloon force method
(f) Improved GVF method
 81

Abnormal X-ray Coronary angiogram – Image 1

(a) (b)

(c) (d)

(e) (f)

Figure 4.14 Segmented output for abnormal coronary angiogram-Image 1

(a) Input image (b) Level set image (c) GVF method
(d) Dynamic GVF method (e) Balloon force method
(f) Improved GVF method
 82

Abnormal X-ray Coronary angiogram - Image 2

(a) (b)

(c) (d)

(e) (f)

Figure 4.15 Segmented output for abnormal coronary angiogram-Image 2

(a) Input image (b) Level set image (c) GVF method
(d) Dynamic GVF method (e) Balloon force method
(f) Improved GVF method
 83

The iteration vs energy graphs and comparison of all energy

functions graphs for all active contour models were also plotted.

(a) (b)

Figure 4.16 Iteration Vs energy for (a) Improved GVF method and
(b) Comparison of all energies for normal coronary angiogram
- Image 1

(a) (b)

Figure 4.17 Iteration Vs energy for (a) Improved GVF method and
(b) Comparison of all energies for normal coronary angiogram
– Image 2
 84

(a) (b)

Figure 4.18 Iteration Vs energy for (a) Improved GVF method and
(b) Comparison of all energies for abnormal coronary
angiogram – Image 1

(a) (b)

Figure 4.19 Iteration Vs energy for (a) Improved GVF method and
(b) Comparison of all energies for abnormal coronary
angiogram – Image 2
 85

Table 4.1 Results in Active contour model for 20 Frames

Area of Perimeter of
Cross Mean Square
Methods segmentation segmentation Time
Correlation Error
in mm in mm
Level set 0.64 ± 0.14 3.62 ± 0.83 0.0432 151.38 ± 13.82 2hrs
GVF 3.60 ± 1.51 13.86 ± 3.54 0.0538 137.69 ± 74.01 68sec
Dynamic
4.60 ± 1.26 33.59 ± 11.58 0.0571 139.05 ± 47.33 85 sec
GVF
Balloon
5.34 ± 1.03 49.40 ± 12.45 0.0672 39.42 ± 33.40 102 sec
force
Improved
6.60 ± 1.06 56.59 ± 11.58 0.0732 29.80 ± 58.35 102 sec
GVF

The inconsistency of mean square error in normal and abnormal

images is due to the fact that the algorithm detects the non-boundaries while
tracking the vessels. This is avoided by adding a balloon force to the dynamic
GVF method which leads to improved GVF. The improved GVF method
allows for flexible initialization of the snake or deformable surfaces and
encourages convergence to boundary concavities and to increase the capture
range and evolution- stop mechanism. Based on Figures 4.12 to Figure 4.15
and Table 4.1 two main points can be drawn as to the performance
comparison. One point is that in terms of subjective criteria all active
contour methods capture range is far enough to locate the blockage of blood
vessels and it easily became stuck on unwanted features. The proposed approach
succeeds in locating the blockage in all cases. The other point is that, according to
quantitative analysis, improved GVF method results are preferable to results from
all other active contour models.

Graphs in Figure 4.16 to Figure 4.19 show the different energy

content of the curve throughout the process of outlining an object. The
 86

continuation of the graph from the point of minimum energy is for justifying
the contour energy at minimum, i.e., stabilizing the contour. Based on
experiments on segmenting the blood vessels in real X- ray coronary
angiogram images, it has shown that the proposed method is robust to the
variation in initial position and efficient in preventing the snake from
breaking through correct contour and locking to other feature points. Area,
perimeter, cross correlation and mean square error are tabulated with respect
to manual segmentation. Based on the tables, segmented images and graphs it
is proved that improved GVF method gives better results compared to all
remaining active contour models.

Although the implementation of this approach is sensitive to its initial

position and is vulnerable to image noise, it provides a powerful interactive
tool for image segmentation. The parameter values shown in the results
depend on the initialization of the contour. But an automatic segmentation
method would be efficient and also helpful to doctors for diagnosis with less
computational time, which is discussed in the next chapter.