Biomedical Engineering Reference
In-Depth Information
algorithm falls into one of three categories: algorithms that use a landmark-
based method, algorithms that use a surface-based method, or algorithms that
work directly on the image intensities (voxel-based). Maintz
et al
. [1] gave a
survey of registration methods.
For automated registration a quality measure of the registration is necessary
in order to find an optimal registration. Maximization of mutual information (MI)
of voxel intensities, the registration method independently proposed by Wells
et al.
[2] and Maes
et al.
[3], is one of the most popular registration methods for
three-dimensional multimodality medical image registration. This method mea-
sures the statistical dependence between the image intensities of corresponding
voxels in two images; this statistical dependence is maximal when the images
are totally aligned.
Intensity-based methods regard all voxels in the images as independent, thus,
anatomical information is not taken into consideration. Maurer
et al.
[4] and Gall
et al.
[5] exploited landmark-based methods. Audette
et al.
[6] gave an overview
on surface registration techniques. Landmark-based methods and surface-based
methods utilize features extracted from the images. The required preprocessing
is usually time-consuming and the accuracy of the registration is dependent on
the correctness of the landmark or surface extraction.
We have developed a two-stage method, which is both feature-based and
intensity-based. Three binning methods were utilized and the performance of
each is compared in this chapter. In the first stage, we segment the images
using region-growing. Then we perform one of the three binning methods on the
full foreground before the down-sampled images are registered. In the second
stage, the results from the first stage are taken as the starting point for the reg-
istration of the full original images. Experiments show that this new two-stage
method gives improved accuracy without loss of speed, compared to multires-
olution registration without bin preprocessing. Of the three binning techniques
used, the nonlinear binning method gave the best performance. Normalized mu-
tual information (NMI) is used as the similarity measure, and downhill simplex
method is taken as the optimization method due to its quickness in practice.
13.2.1
Image Registration Using Binning Technique
Registration based on maximization of mutual information uses an iterative ap-
proach in which an initial estimate of the transformation is gradually refined.
