Biomedical Engineering Reference
In-Depth Information
millimeters from the 3D centroid of a segmented organ such as the prostate.
Finally, when images have comparable gray levels, a difference image can
provide a visual evaluation or a quantitative evaluation from image statistics.
A downside with MR difference images is that the inhomogeneity of the signal
response and interpolation can introduce artifacts in difference images. Since
MR image intensity can vary with different MR sequence parameters and the
signal response of MR coil, gray value statistic may have some limitations when
image acquisitions are not carefully repeated.
3.3.5.3
Algorithmic Robustness and Efficiency
The rigid body algorithm is robust for a global registration. Because of two
principal design features, the algorithm is quite robust and accurate for volume
pairs acquired in the same positions and with comparable conditions [1]. First,
using both CC and MI at different resolutions was an important feature that
increased robustness. CC gave fewer local minimums at low resolutions and
MI was more accurate at high resolution [1, 5]. Second, the restarting mecha-
nism was also quite important. Without restarting, we found that registrations
sometimes failed in cases of volumes with large mismatches and significant de-
formation. Even these cases resulted in a proper solution when restarting was
employed.
Based upon our initial experiments with interactive CP selection, we deter-
mined that many CPs were required for good matching throughout the pelvis.
As a result, we designed algorithm features to be computationally efficient for
TPS warping with hundreds of CPs. First, the optimization of small VOIs is very
fast. Second, we optimized each CP separately because the optimization of three
parameters (
x
,
y
, and
z
) is simple and fast. Conversely, as previously reported
by others [43, 44], the simultaneous optimization of many CPs leads to a much
more complicated error surface and local maximums. If one were to use 180 CPs
and optimize the 540 free parameters simultaneously, the optimization process
would become extraordinarily complex. Third, we applied the TPS transforma-
tion once to the final, optimal CPs, which saved considerable time. If TPS was
applied in each iteration, the registration time would be unacceptable for our
application. If we were to use optimized C code, the total time for rigid body
and non-rigid registration should reduce to within 5 minutes.
