Information Technology Reference
In-Depth Information
3.3
Improvement Phase
The segmentation phase gives quite well results but is designed only to find
just approximate localization of both ventricles. It does not take into considera-
tion any additional expert knowledge that should be considered (more detailed
description of that need will be discussed further). In this work a sample im-
provement is used which shoud imitate the behaviour of expert that manually
draws the contours. As it was mentioned earlier, 2D images are the result of
reconstruction, which aims at creating two chamber short axis view. Thus, the
pixel intensities of the obtained slices are averaged. That is why during image
examination, in order to find a proper contour, an expert must look not only
at the currently segmented slice bu also at the slices above and below. This is
imitated using the following energy forumulation for the right ventricle:
E
imp,ri
=
x
∈C
ou
∩L
le
D
(
x
)
O
imp,ri
+
(4)
n
(
C
ou
∩ L
le
)
and for left ventricle:
x
∈C
ou
∩L
ri
D
(
x
)
E
imp,le
=
O
imp,le
+
(5)
n
(
C
ou
∩ L
ri
)
Here
is a slightly scaled down image after thresholding with a strong distance
potential filter applied (Fig. 6a). The weights of the energy components (equal
to 1) are chosen experimentally.
The first component of the above equations ensures that all the pixels repre-
senting blood with injected contrast that lay on the proper side of the septum
line are inside the contour. The second component (identical for both ventricles
which need not to be a rule) is used to make the contour, in the given
D
points
equally distributed along the contour, to lay there where local differences be-
tween avarage pixel intensities (from the current slice and the slices above and
below) outside and inside the contour are the biggest:
K
K
1
K
imp,ri
=
imp,le
=
O
O
(1
− u
k
)
(6)
k
=1
where
u
k
=
t
k
if
t
k
≥
0
(7)
0if
t
k
<
0
and
t
k
=
x
∈T
i
k
x
∈T
ou
k
D
(
x
)
D
(
x
))
−
(8)
T
i
k
T
ou
k
n
(
)
n
(
)
Here
is a an image obtained by avaraging (after thresholding) the pixels of
the current slice and of the slices that lay directly below and above the current
slice (Fig. 5).
D
