Biomedical Engineering Reference
In-Depth Information
Fig. 6 The results of our algorithm are in (a); it is the ground truth. Models in (b-d) are results of
Bhattacharyya, RSS, and graph cuts, respectively. These algorithms were initialized with the
ground truth from (a). In the bottom row, surface curvature is displayed; these models were
computed on a mesh generated from the corresponding models (by column) in the first row. One
should notice that even initialized with ground truth, the algorithms in (b-d) move away from the
correct result; this means that all of these energies do not have the ground truth as a local minimum
for any region. Also notice (as more clearly shown by the surface curvature) that the largest errors
using these energies occur in regions of greatest importance for us: around the growth plate
separating the two pieces of bone
Model-based segmentation methods [
11
] and [
12
] are powerful but they require
a representative training set during the learning phase. Prior to the completion of
our population study, a training set was not available. For future work, we plan to
incorporate shape priors into our interactive segmentation framework.
3.3 Comparison to Interactive Segmentation Methods
The live-wire algorithm is an interactive segmentation method that “snaps” to
image edges [
13
]. In [
1
], the user provides control points by segmenting several
slices with the help of the live-wire algorithm, and variational interpolation
is performed on the resulting cloud of points. In [
14
], “Spotlight” guides the
user to places that require additional input; input is provided an a plane with the
