Biomedical Engineering Reference
In-Depth Information
coarser level. Nearest neighbor interpolation method is used in this level. In the
second level, the original images are registered and a trilinear partial volume
interpolation method [18] is employed. This process is done for each of the
three binning techniques.
13.2.2
Experiments and Discussion
13.2.2.1
Data Set
This study involved the data sets of seven patients, each consisting of CT and
a subset of six Magnetic-Resonance (MR) volumes (spin-echo T1, PD, T2 and
the rectified version of each of these three). The MR-T1-rectified image was
not available for patient six. Thus, 41 image pairs were available to be regis-
tered. These images were provided by
The Retrospective Registration Eval-
uation Project
database maintained by J. Michael Fitzpatrick from Vander-
bilt University, Nashville, TN, USA [26]. For each patient data set, all CT im-
ages were registered to the MR image using the MR image as the reference
image.
All image pairs were registered using the maximization of NMI. Registra-
tion transformation was limited to six-parameter rigid-body type (three trans-
lations and three rotations) transformations. Registrations were conducted on
a PC, having a 2.4 Ghz Intel Pentium 4 processor, and a 512MB DDR SDRAM
memory.
The experiments were performed using a two-stage multiresolution ap-
proach described above. When the binarization approach was used, the fore-
ground/foreground bin was given an additional weight when the joint histogram
was computed in the first level. The additional weight was heuristically de-
termined using the ratio of foreground to background voxels. The number of
histogram bins used to compute the normalized mutual information criterion
was set to 256 bins for both CT and MR.
The Downhill-simplex optimization method was used throughout the ex-
periments. Optimization for each pair of images started from the same initial
position with all three translation offsets set to zero millimeters and all three
rotation angles set to zero degrees. Convergence was declared when the fraction
difference between the lowest and the highest function value evaluated at the
vertices of the simplex was smaller than 10
−
5
.
