Biomedical Engineering Reference
In-Depth Information
was obtained at very large displacements where the overlap was reduced. This
occurs because MI is not only a function of how well the images match in the
overlap, but also by how much information is provided by the two images in the
overlap [37].
3.2.5.4
Computer Implementation
Accuracy is an important issue for automatic registration, but there are oth-
ers such as robustness, speed, and requirements for operator interaction. With
the multiresolution and restarting features, our modified MI algorithm is quite
robust. For a wide range of initial guesses, it worked well for all 22 volume
pairs reported here. Three of the volume pairs were from patients, and we are
confident that routinely acquired clinical images will have sufficient quality for
registration. Because good starting values are unimportant, operator interaction
is minimal. In one instance, cropping of the legs was important for registering
an image volume obtained in the treatment position with that in the diagnos-
tic position. It is not surprising that legs in a very different position have to be
cropped. Although this is easy to do manually, we can probably determine an
automated method if it is deemed desirable.
The mutual information similarity measure is quite robust. Even though our
images are very similar, we had less success with some other measures such
as the sum of the squared image difference. An advantage of MI is that it can
be used with images from different modalities, a feature that we are starting
to use.
3.3
Three-Dimensional Non-Rigid Body
Registration Algorithm
3.3.1
Why Non-Rigid Registration
In the previous section, we discussed rigid body registration of the prostate. For
volume pairs acquired over a short time span from a supine subject with legs flat
on the table, registration accuracy of both prostate centroids (typically
<
1 mm)
and bony landmarks (average 1.6 mm) was on the order of a voxel (
≈
1
.
4 mm).
