Biomedical Engineering Reference
In-Depth Information
performed on nine RT3D cases. The correlation coefficient was 0.87 when com-
pared to MRI measurement. The main limitation of this study is the absence of
testing on normal physiological ventricular volume in the range [40-150 ml] for
which the behavior of the level set segmentation can be significantly different
as these volumes are much smaller.
2.4.3.5
Segmentation of RT3D Ultrasound with Implicit
Deformable Models Without Gradients
This research was published by Angelini, Holmes, Laine and Homma in [43].
Method:
This study focused on the same clinical problem as the previous
study for segmentation of echocardiographic RT3D ultrasound data. The pro-
posed method uses the homogeneity-based implicit deformable model proposed
by Chan and Vese in [40] as an extension of the Mumford-Shah segmentation
functional. Motivations for selection of this method include robustness with ar-
bitrary initialization of the object anywhere in the image, topology adaptation for
multiobject segmentation (for potential cosegmentation of ventricles and atria,
for example), self-adaptation of the deformation flow to inward and outward
flows. Minor modifications of the method were performed to adapt the design
to the specificity of the 3D ultrasound data. The homogeneity terms from Eq.
(2.28) were weighted by the mean intensity value as:
I
−
c
0
c
0
2
I
−
c
1
c
1
2
E
(
C
)
=
d
+
d
.
(2.43)
inside
C
outside
C
A similar approach was followed by Lin
et al.
[44] for segmentation of 3D echocar-
diographic data where they normalized the homogeneity term by the variance
of the data inside and outside the object segmented, after pre-processing with
multiscale Gaussian filtering.
Parameters were set to
υ
=
0 (no constant inflation force was used),
µ
=
1,
λ
1
=
0
.
25,
λ
2
=
0 (no homogeneity constraint on the outside of the ventricle
to reduce the effect of the noisy myocardium texture),
x
=
y
=
z
=
1,
t
=
min(
x
,
y
,
z)
/
|
υ
+
µ
+
λ
1
+
λ
2
|
(to respect CFL condition with explicit
numerical scheme). The system was let to deform over 20 iterations.
Experiments
: A clinical study was performed on 10 patients with pulmonary
hypertension to segment both right and left ventricular volumes. A 2D para-
metric deformable model and a 3D level set deformable model illustrated in
