Biomedical Engineering Reference
In-Depth Information
0.05
0.045
10000
10000
100
0.04
ignored
100
1
0.035
ignored
1
0.03
0.025
0.02
0.015
0.01
0.005
0
−
15
−
10
−
5
0
5
10
15
X Rotation (degrees)
Figure 10.19: The reversed cross-entropy profiles of an MR/Em image pair in
the vicinity of a manual registration when the x-rotation angle changes indepen-
dently. The numbers around the profiles indicate the
α
value used to compute
a small probability value which is used to replace zero pdf. Ignored means that
those pdfs are ignored when computing the profile.
density estimation with a Gaussian kernel to add smoothness and stability to
the similarity functions, as in Wells
et al.
[14]. Nevertheless, the simple, heuris-
tic approach we used here proves to be effective. The main motivation for this
heuristic approach is to make the similarity measure profiles smooth and to
have a maximum (peak) or minimum (valley) around the unknown truth with a
reasonble capture range. Both of these conditions are necessary for good regis-
tration measures. Moreover, it is important that the modified priori or estimated
joint pdf do not deviate from a real pdf. That is, the sum of the modified pdfs shall
still be close to 1 and the difference from the original values shall not be signifi-
cant. To further improve the performance of these similarity measures, one can
employ different interpolation methods. It was reported recently that different
