Biomedical Engineering Reference
In-Depth Information
but in all cases less time was required to segment the image using the algorithm
than the manual approach [31].
Good initialization is required for the DDC to converge to the desired bound-
ary; however, to reduce the burden on the user in comparison to manual outlining,
the amount of input required for initialization should be kept as small as possible.
The four-point method works well for prostate shapes that are approximately el-
liptical. These are generally in the “easy” and “moderate” categories, representing
87% of our images. For the “difficult” cases, initialization is poorer but could be
improved by requiring the user to initialize the DDC with a
few
additional points.
2.4.2. Metrics
Table 1 lists the mean values and standard deviations for each of the metrics
described above: MD, MAD, MAXD,
C
s
, and,
C
a
. The means and standard devi-
ations were computed from the corresponding metric for each of the 117 images.
The mean value of MD is close to zero, that of MAD less than 5 pixels (0.63 mm),
and that of MAXD less than 20 pixels (2.5 mm), indicating good overall agreement
between the manual “gold standard” and semiautomatic segmentation methods.
However, the values of these metrics generally increase as the segmentation com-
plexity increases from “easy” to “moderate” to “difficult.” The mean sensitivity
and accuracy are greater than 90%.
Whereas the abovemetrics quantify the global performance of the algorithmas
compared tomanual outlining by an expert, the local metric
d
j
(
θ
i
) provides insight
as to locations, specified by the angle
θ
i
, where the manual “gold standard” and
algorithm outlines deviate or agree well with each other. The segmented prostate
boundaries are generally similar in areas where the contrast is high, such as the
lower boundary of the prostate defined approximately by 240
◦
θ
i
<
320
◦
.
The segmented prostate boundaries deviate from each other at
θ
i
= 190
◦
and
θ
i
= 350
◦
, where the ultrasound reflection is weak because the prostate boundary
is locally parallel to the ultrasound beam. Deviation also occurs in areas of shadow
caused by calcifications in the prostate.
Figure 6 is a plot of the area enclosed by the algorithm-generated outline
versus that enclosed by the corresponding manual outline for all 117 images. A
line passing through the coordinate origin was fitted to the data in a least-squares
sense, and has a slope of 0.987 (
R
= 0.993), suggesting that areas enclosed by the
manual and algorithm outlines agree over a large range of prostate cross-sectional
areas from 1.5 to 15.6 cm
2
.
≤
3.
TESTING AND OPTIMIZATION USING VIRTUAL OPERATORS
3.1. Introduction
Two parameters are critical to the performance of our algorithm for segment-
ing 2D TRUS images:
σ
in Eq. (3a) and
w
int
in Eq. (2). The value of
w
img
is also
