Biomedical Engineering Reference
In-Depth Information
Table 1.
Average Distance- and Area-Based
Metrics for all 117 Images
Metrics
Average
Standard Deviation
Distance-based
MD (pixels)
-0.5
2.3
MAD (pixels)
4.4
1.8
MAXD (pixels)
19.5
7.8
Area-based
Sensitivity (%)
94.5
2.7
Accuracy (%)
90.1
3.2
Reprinted with permission from the AAPM.
2.2.3. Use of edge direction information during deformation
In TRUS images of the prostate, the interior of the prostate appears darker
than the exterior. We have chosen to only apply the radial component of the image
force at any vertex if the gray levels in the vicinity of a vertex vary from dark to
light in the direction of the outer radial vector,
r
i
; otherwise, no force is applied.
Figure 2b shows the final deformed DDC for the image in Figure 2a.
2.2.4. Editing
Intuitive editing mechanisms are easy to incorporate into the segmentation al-
gorithm, and permit the user to guide the algorithm if the segmentation is deemed
inadequate. Figure 3b shows a segmentation result in which the DDC converged
to incorrect features because of poor initialization in the localized regions, indi-
cated by the arrows in Figure 3a. The editing tools allow the user to move a few
vertices onto the prostate boundary, clamp the vertices into place and re-deform
the boundary. Only a few vertices (as indicated in Figure 3c) need to be moved
because re-deformation will initially cause neighboring vertices to move under the
influence of internal forces so that the DDC becomes smoother. As these neighbors
move to minimize local curvature, they may come under the influence of image
forces around the desired prostate edge and become attracted to it, as shown in
Figure 3d.
