Biomedical Engineering Reference
In-Depth Information
be done manually since the appearance of overlapped curves, or of a
curve on top of an image, is not visually confusing. Matching can, of
course, also be done automatically. Curve-to-curve matching has
been used extensively in the 2D alignment of radiographs with DRRs.
Voxel-to-voxel matching
Surface-to-surface matching requires
delineation of the surfaces of the volumes of interest. This can be a
demanding and labor-intensive step. Moreover, only a limited part of
the information in the images is used in the matching process. A third
approach is to match the image values at each point - that is, at each
voxel in 3D or at each pixel (picture element) in 2D. For images of
the same modality, an autocorrelation approach can be taken. More
generally, the method of maximization of “mutual information”
(Viola and Wells, 1995) has met with great success. One problem
with voxel-to-voxel registration is that one tends to look at all the
information in the images whereas, sometimes, certain parts of the
image may be unreliable. For example, the mandible's location
relative to the skull may be different at the times that two images
were made, but it is irrelevant if one is only interested in the inter-
registration of features within the skull; in that case, the portions of
the images in which the mandible appears should be “thrown away.”
This is simple in theory, but very time-consuming in practice
whereas, with manual identification of points and/or surfaces, the
−
selection of relevant anatomy is easy and instinctive.
When matching 3D data sets, one has to remember that there may,
and generally will, be rotations and translations in the third dimension
such that one cannot pair-wise match 2D sections. One should use a
fully 3D approach.
Deformable image registration
The preceding methods can readily handle changes in scale between
the registered studies by simply including scaling factors as variables
to be determined in the fitting procedure. However, a much more
vexing and difficult problem is when one study is spatially deformed
relative to the other. Deformation may occur because the studies
were taken at different times (e.g., on different days, or at different
points in the respiratory cycle) or because one study is intrinsically
spatially distorted (as may be the case with MRI). At the time of
writing, the registration of deformed image sets is a matter of current
research and there is no well-accepted solution. Figure 3.18 shows a
comparison of rigid and deformable image registration. Figure 3.18a
shows a CT cross-section of a prostate cancer patient's anatomy
