Biomedical Engineering Reference
In-Depth Information
that the average volume error of the parallel re-slicing method is 5.4% of the vol-
ume, and it is 1.7% for the rotationally re-slicing method. The average editing
rate of the parallel re-slicing is about 20%, and 14% for the rotationally re-slicing
method. The variability in volume error and editing rate of the two methods is al-
most the same. This demonstrates that the rotational re-slicing method is superior
to the parallel re-slicing method, and that accurate and reproducible results can be
obtained.
4.1.4. Segmentation time
Both segmentation methods required a four-point initialization lasting about
3 seconds. After initialization, both algorithms were able to segment a complete
prostate within about 3 seconds. Contour editing, when required, added about 6 or
7 seconds per edited slice to the procedure. This time included the time required
to deform the boundary in the edited slice and to propagate the correction through
the subsequent slices in the segmentation series.
4.2. Slice-Based 3D Segmentation Using an AR Model Continuity
Constraint
4.2.1. Error accumulation and continuity constraint
The slice-based 3D segmentation described in Section 4.1 is efficient and can
produce accurate results, but it can fail when the evolving contour approaches
shadows at the prostate boundary or bright regions inside the prostate produced
by intra-prostatic calcifications or implanted brachytherapy seeds. As discussed
in Section 4.1, manual editing is helpful for solving the problem; however, the
following problems limit the utility of the slice-based segmentation methods.
1) Because the 3D boundary evolves from sequential 2D boundaries, it is
difficult to choose from which slice to start editing. In addition, manual
editing increases the variability of the final segmented prostate boundary.
2) Manual editing is performed by the user, who must decide when to start
editing. This adds time to the segmentation procedure and takes much
longer than a method that would not require manual editing.
Figure 15 shows an example of error accumulation in the segmentation of
a prostate using the rotational slice-based segmentation. In the 45th slice, the
refined prostate contour matches the manually segmented boundary, as shown in
Figure 15a. Then, the segmentation error increases and accumulates (Figure 13-
15b), and finally reaches its maximum deviation from the manually segmented
prostate boundary in the 56th slice (Figure 15c). This error accumulation can be
observed at the intersection between the 2D segmented contours and the coronal
plane passing through the approximate center of the prostate (with the rotational
