Biomedical Engineering Reference
In-Depth Information
We obtained somewhat larger prostate registration errors of about 3.0 mm when
volume pairs were obtained under very different conditions, e.g., legs flat and
legs raised, or with and without bladder or rectal filling. Rigid body registration
of the pelvis cannot follow prostate movements due to changes in the postures
of legs and deformation of the bladder and rectum [8]. In this section, we discuss
the ability of non-rigid registration to express this deformation.
Non rigid registration studies are reported for the brain [38, 39], for the
breast [40, 41, 41, 42, 42], for a variety of other organs [23, 43, 45], and for
excised tissue [46]. Far few reports described results of the pelvis and prostate.
Bharaha
et al.
reported a method using manually segmented prostate for rigid
body registration followed by finite element-based warping in the application
of prostate brachytherapy [47]. Voxel based methods, particularly those based
upon mutual information, are robust, require no segmentation that can be prone
to error, are highly accurate for brain registration [31], and are suitable for
abdominal registration where there can be deformation [20]. We are discussing
voxel-based non-rigid registration for the particular application in the pelvis and
prostate.
In this section, we perform experiments to compare non-rigid and rigid body
registration for the prostate and pelvis. By using high-resolution MR images giv-
ing distinctive anatomic detail, we test the ability of a non-rigid algorithm to cor-
rect anatomical variations throughout the pelvic region. We include conditions
with very significant changes in posture possible in interventional applications,
that is, we attempt to register image volumes from a diagnostic scan with legs
flat to those from a treatment acquisition with legs raised. We qualitatively and
quantitatively evaluated registration results using 17 volume pairs from three
volunteers.
3.3.2
Non-Rigid Registration Algorithm
Figure 3.8 outlines the non-rigid registration algorithm that includes three major
steps: control point selection, control point optimization, and thin plate spline
warping. Prior to non-rigid registration, we perform rigid body registration as
reported in Section 3.2. Again, the unchanging volume is the
reference
, and the
one to be warped is
floating
.
The manual selection of CP's is an important step. We used
RegViz
for vi-
sualizing and analyzing image volumes. Following rigid body registration, the
