Biomedical Engineering Reference
In-Depth Information
Figure 6.
Area enclosed by algorithm-segmented outline versus corresponding area en-
closed by manually outlined area. Reprinted with permission from the AAPM.
w
int
, as it should be balanced with
w
int
.
As described in Section 2.2.5, the values used for these parameters were selected
empirically, and their effects on segmentation quality were not investigated or
optimized. Here, we describe a method for testing the effects of these parameters
using “virtual operators” [38]. The method described here can also be applied to
testing and optimization of other segmentation algorithms.
Chalana and Kim [39] have proposed a method for evaluating the single output
produced by automatic segmentation algorithms. The single algorithm-generated
boundary is compared to a “gold standard” formed by averaging together multiple
boundaries that are manually segmented by several observers, rather than com-
paring to a single boundary outlined by a single observer, as done in Section 2.3.
Forming a “gold standard” from multiple outlines reduces the effects of observer
bias and accounts for intra- and inter-observer variability. In our semiautomatic
algorithm, the output is dependent on the selection of the four control points and
is also subject to observer bias and intra- and inter-observer variability. Therefore,
multiple algorithm-generated boundaries, rather than a single boundary, should
be compared to multiple manually generated boundaries. However, applying our
important, but can be constrained to be 1
−
