Biomedical Engineering Reference
In-Depth Information
select 180 CPs. Based on our experience, we think that it is possible to create
an automatic or semiautomatic method for selection of appropriate CPs in the
pelvis. For example, one might use a grayscale threshold to detect the pelvic
outer boundaries and apply edge enhancement to extract feature of internal
structures. CPs would be placed on such structures automatically. We are inves-
tigating this and other methods for CP selection followed by automatic non-rigid
registration.
One way to adjust the movement of control points is to change the size of the
VOI. In the case of treatment-diagnosis volume pairs, a large VOI size of 64 on a
side worked better than smaller ones because displacements were large, because
larger VOIs tend to give a more robust optimization, and because no small local
deformations were required. However, a size of 16 on a side worked better for
the case of full-empty bladder volume pairs because small VOIs better capture
the small, local deformations. VOIs with a size of 64 on a side covered most
of the bladder and could not generate small local deformations. For volumes
with both large and small-scale deformations, we suggest using different VOI
sizes for different CPs.
With non-rigid registration, we have to be concerned about potential warping
errors affecting the application of interest. For the prostate, we used 3-5 CPs
near the prostate center because we desired to maintain the spatial integrity of
the organ and to preserve the tissue volume. We placed many CPs around the
pelvic surface to produce reasonable warping.
3.3.5.2
Evaluation of Non-Rigid Registration
Since there is no gold standard for non-rigid registration of anatomical images,
we used a variety of methods to evaluate registration quality. First, for routine
evaluation, a color overlay is simple, fast, and intuitive. To better visualize the
two data sets, we interactively adjust the transparency scale of each image.
Second, for illustration of subtle difference along an edge, we recommend a
sector display because it best shows small shifts. Third, for visual evaluation of
a specific organ such as the prostate, we like to superimpose manually marked
contours from one image onto another as shown in Figs. 3.12 and 3.13. This
clearly shows any displacement or deformation even in a printed figure. Fourth,
a more quantitative approach is obtained by calculating the displacement in
