Biomedical Engineering Reference
In-Depth Information
segmentation algorithms, a lesser emphasis has been placed on appropriate eval-
uation of segmentation algorithms. Investigators continue to develop yet more
advanced and computationally complex segmentation algorithms but often fail to
evaluate them using uniform or standard performance measurement metrics. As a
result, finding the appropriate segmentation algorithm for a particular task as well
as choosing the optimal parameters internal to the segmentation algorithm is often
a problem. Thus, it is necessary to identify appropriate tools for evaluating and
comparing segmentation algorithms.
In this chapter we have reported on the evaluation of our approaches using
distance-, area-, and volume-based metrics to compare the performance of the
DDC-based semiautomatic prostate segmentation algorithms to manual outlining.
Since therapy planning often requires knowledge of the prostate volume or cross-
sectional area, area-, and volume-based metrics are important. These types of
metrics are easy to compare, as a single value is obtained for each segmentation
of each prostate. Accuracy and variability metrics can be used as discussed above
and statistical comparisons between segmentation methods can be performed. In
addition, area- and volume-based metrics can also be used to optimize the choice
of the DDC parameters. Examples of the use of these metrics are given in the
sections above.
Distance-basedmetrics provide both local and global methods for comparison.
Global distance-based metrics such as the MAD and MAXD averaged across the
ensemble of images indicate overall agreement between the two outliningmethods.
Since the values of these metrics may vary locally due to ultrasound imaging
artifacts, a local plot of the error supplements this global information by indicating
regions where the two methods agree or disagree the most. This approach is useful
to identify consistent segmentation errors where the prostate boundary may be
weak or missing altogether either because of the orientation of the prostate with
respect to the ultrasound beam or because of shadowing. Identification of these
regions will guide the developer where improvements are necessary.
In summary, we have described a DDC-based segmentation tool to be used
to segment the prostate from 3D US images. We have described a 2D method
to segment cross-sectional images of the prostate and have described methods to
extend this technique to 3D. In addition to the development of the algorithms,
we have used evaluation metrics and statistical tests to validate the segmentation
results and optimize the algorithm parameters. Based on our results, it is clear
that a DDC-based segmentation approach can provide fast, accurate, precise, and
robust 3D segmentations of the prostate from 3D US images and can be used for
intraoperative prostate brachytherapy procedures.