Biomedical Engineering Reference
In-Depth Information
Table 1.
Quantitative Validation on 26 CT Images
DSC
D
surf ace
(mm)
D
centroid
(mm)
Average
0.8172
3.38
2.86
Standard deviation
0.0807
1.11
1.63
Minimum
0.6573
2.07
0.50
Maximum
0.9327
5.42
5.75
One should point out, however, that the centroid distance has only a very limited
capacity to quantify shape differences. Obviously, it cannot distinguish between
any two segmentations that share the same centroid.
Table 1 gives the average value of all three criteria computed for the entire
dataset in the leave-one-out strategy mentioned above. In addition, we displayed
the standard deviation, minimum, and maximum value of each criterion. Overall,
these values show that our segmentations typically agree well with the manual
ground truth.
5.2.3. Robustness to initialization
The level set method for image segmentation and also its implementation
with nonparametric statistical shape priors are
local
optimization methods. As a
consequence, experimental results will depend on the initialization. This aspect
is a common source of criticism, it is generally believed that
local
indicates that
segmentations can only be obtained if contours are initialized in the vicinity of the
desired segmentation. Yet, this is not the case for the region-based segmentation
schemes like the one developed in this work. The segmentation without shape
prior in Figure 4 shows a drastic difference between initial and final boundary:
clearly contours can propagate over large spatial distances from the initialization
to the “nearest” local minimum.
In order to quantify the robustness of our method to initialization, we trans-
lated the initialization by a certain distance in opposite directions and subsequently
computed the accuracy of the resulting segmentation process with nonparametric
shape prior. Table 2 shows that the accuracy is quite robust with respect to dis-
placements of the initialization up to 10mm in each direction.
5.2.4. Robustness to noise
The prostate CT images are in themselves rather challenging, since prostate
and surrounding tissue have fairly similar intensities (see Figure 1, right side). The
combination of statistically learned nonparametric models of both the intensity
