Biomedical Engineering Reference
In-Depth Information
3.
E
A
is an anatomy-driven constraint. It preserves the distance between the
myocardium border surfaces within an admissible range of values.
4.
E
S
is a shape-driven global consistency constraint. This knowledge-based
term performs a registration of the evolving contour to a prior shape model
(
A
0
,
A
1
) via rigid deformation. The prior shape models were defined in a
pixel-wise stochastic level set representation [56].
The authors reported some experiments on segmentation of the endocardium
from 2D cardiac MRI images. These experiments revealed that the anatomical
constraint played a minor role in controlling the deformation of the segmenting
surface and that the regularity term was overwritten by the shape prior term.
Final segmentation results showed a reliable performance of the method but
no quantitative validation was performed. It was pointed out by the authors
that the integration of the different modules was difficult and that future refine-
ments of the approach were considered such as the use of a single level set
function for segmentation of the myocardium, and tracking contours in time
by replacing the prior shape model with the segmentation from the previous
frame in the context of consecutive time frames segmentation over a cardiac
cycle.
In two related papers, Paragios [63, 70] proposed modified versions of the
method.
In [70] Paragios had proposed a version of the method where the regularity
term consisted of a boundary component derived from gradient vector flow [71]
to detect cardiac boundaries and curvature constraints on the segmented shape.
No shape-driven constraint was proposed in this early work.
In a posterior work [63], Paragios modified the method for segmentation of
the endocardial surface on ultrasound. The model was first modified to replace
the regularity term
E
R
by a boundary constraint
E
B
derived from a geodesic
active contour formulation [16]. The model was further modified to integrate
temporal tracking of the segmented contours between consecutive time frames.
A time-tracking constraint, in the form of a bounded error function using a robust
norm
ρ
was introduced as:
H
(
φ
t
t
+
1
t
)
ρ
(
I
t
−
I
t
+
1
(
T
))
d
E
T
(
φ
,φ
,
T
)
=
H
(
φ
(2.35)
t
+
1
)
ρ
(
I
t
(
T
−
1
)
−
I
t
+
1
)
d
+
