Biomedical Engineering Reference
In-Depth Information
registration. The multiresolution approach enabled the program to get close to
the final value quickly because of the reduced number of calculations. That is,
the time for reformatting at the lowest resolution (1/4) was 9.8 sec, which was
less than 1/59 times that at the highest resolution, a value nearly equal to the
1/64 expected from the change in the number of voxels. The number of restarts
was 5, 1, and 1 for resolutions at 1/4, 1/2, and the full number of voxels. Each
call to the simplex optimization resulted in 55 to 94 MI evaluations before the
tolerance (0.001) was reached. The simplex optimization method worked about
1.5-2.0 times faster than the Powell method in our implementation. The time for
registration using Simplex, typically 5 minutes on a Pentium IV, 1.8GHz CPU,
with 1GB of memory, could probably be greatly improved with optimized C code
rather than IDL.
3.2.5
Discussion
3.2.5.1
Registration Accuracy
Our results suggest that MI can be used to accurately register, with an error on
the order of a voxel, MR pelvic images obtained under similar conditions. Be-
cause it gives an independent, true 3D measurement, we like to use the method of
point bony landmarks to assess accuracy. However, as argued in section 3.2.4.3,
the true MI registration accuracy might be better than our ability to measure it
with point bony landmarks. That is, following point landmark registration, the
distance between registered, corresponding landmarks was on the order of that
following MI registration. Very possibly, MI is more accurate than point regis-
tration using bony landmarks. Additional, independent evidence of excellent MI
accuracy comes from the very low error value from the registration consistency
measurement (0.6
±
0.2 mm). Interestingly, this is obtained even though the
interpolation artifact present in MI similarity surfaces should reduce the likeli-
hood of subvoxel accuracy [30]. Our results for the pelvis with image volumes
obtained under the same conditions compare favorably with those for the brain,
where MI registers images very accurately giving errors as small as 0.7-0.8 mm
for CT-MR [31].
Visual and quantitative evaluation of prostate organ movement showed
good registration even when we acquired images under conditions that greatly
stressed the ability to register the images. The small prostate displacements in
