Biomedical Engineering Reference
In-Depth Information
One of the difficulties with this approach is that it can converge to a local op-
timum instead of a global optimum. Using multiresolution in conjunction with
maximization of mutual information is very helpful when tackling this problem.
The work of Maes
et al.
[8], Studholme
et al.
[7] and Pluim
et al.
[9, 10], has
proved this. The idea of a multiresolution hierarchical approach is to register
the coarse (low resolution) image first and then to use the result as the starting
point for finer (high resolution) image registration, and so on.
In order to use NMI, an estimation of the intensity distribution is required.
There are a couple of methods used to estimate the intensity distribution of an
image. Colligon
et al.
[11] used joint entropy as the registration criterion. Viola
[12] obtained the estimation by Parzen, windowing an intensity distribution.
Camp
et al.
[13] proposed a binning method for registration using normalized
mutual information. The image intensities are assigned to a histogram bin by
a binning technique. The most commonly used binning method is equidistant
binning. With equidistant binning, once the bin number is given, the intensities
range assigned to each bin is also determined, after the overall image intensity
range is distributed evenly among all the bins. The weakness of the equidistant
binning method is that it totally ignores the anatomical information of the image.
From typical histograms of CT and MR images, as shown in Fig. 13.1 and Fig.
13.2, we can spot the same property: a giant peak around the intensities of the
background region. In our approach, we use region-growing to separate the
8 10
5
6 10
5
4 10
5
2 10
5
0
−
2 10
5
−
2000
−
1000
0
1000
2000
3000
Intensity
Figure 13.1:
A typical histogram for a CT image.
