Biomedical Engineering Reference
In-Depth Information
while the
F
-statistics were greater than the computed
f
-value. From the tabulated
p
-value, it is apparent that the null hypothesis
H
o
:
σ
3
=
σ
1
is unlikely to be true
(
p
-value = 0.0545), although we are not able to reject the hypothesis with 95%
confidence.
Prostate segmentation from 3D TRUS images is a critical step in planning
prostate intervention procedures such as brachytherapy. The slice-based 3Dprostate
segmentation method is fast, and therefore suitable for clinical application. How-
ever, this method requires manual editing of the boundary when the image con-
trast is poor or in the presence of intra-prostatic calcification artifact. To solve
this problem, we developed a modified slice-based prostate segmentation method
based on the continuity constraint implemented as a zero-order AR model. This
modified slice-based segmentation method propagates in the clockwise and anti-
clockwise directions, and the resulting segmented contours are combined. The
continuity constrained slice-based 3D prostate segmentation method can be used
in other applications when segmentation of ellipsoidal objects are required. Pos-
sible applications are solid tumors, the kidney, and heart chambers. Statistical
analysis of the results obtained using volume- and distance-based metrics veri-
fied that an optimal segmentation can be achieved by combining the clockwise
and anti-clockwise contours. This results with this approach are closer to manual
segmentation than with any other slice-based segmentation method, and with low
local variability compared to the other algorithms we studied in this chapter, and
with lower inter-observer variability for manual segmentation than that described
by Tong et al. [37].
4.3. Direct 3D Segmentation
4.3.1. Algorithm
An alternative to the slice-based DDC segmentation methods is to find the
surface of the prostate directly from a 3D TRUS image using a 3D deformable
model [45]. The direct 3D segmentation algorithm we developed is based on a
deformable model that is represented by a mesh of triangles connected at their
vertices [46,47]. The direct 3D segmentation algorithm is an extension of our 2D
algorithm presented in Section 2, and involves three major steps: (1) initialization
of the 3D mesh; (2) automatic deformation toward the prostate surface; and (3)
interactive editing if required.
The user selects six control points in the 3D TRUS image in order to initialize
themesh: four points in an approximatemid-gland transverse 2D slice, as in our 2D
segmentation algorithm, one point at the prostate apex, and one point at the base. A
sample 3D TRUS image with user-selected control points is shown in Figure 17a.
An ellipsoid is estimated from the control points, and has center (
x
0
,y
0
,z
0
) and
semi-major axes with lengths
a
,
b
, and
c
in the
x
,
y
, and
z
directions, respectively;
this assumes that the length, width, and height of the prostate are approximately
