Biomedical Engineering Reference
In-Depth Information
modalities or may be a physical view of the anatomy in space, a view which
we treat here as merely another “imaging” modality. An object in medical
applications is some portion of anatomy such as the brain, a limb, the chest,
the liver, etc. Each view will always include approximately the same anatom-
ical region. Typically the two views will be taken from the same patient, in
which case the problem is that of intrapatient registration, but interpatient
registration has application as well. Because most research on registration
has focused on intrapatient three-dimensional images and rigid-body regis-
tration, this chapter likewise concentrates the most attention on this problem.
To be useful in any application, registration must be embedded in some
larger system. The need for the embedding system arises because the trans-
formation that the registration step produces as output is a mathematical
mapping function, which is ill-suited for direct human use, particularly
when three dimensions are involved. In rigid-body registration, the transfor-
mation itself is of little clinical relevance. In some nonrigid registration appli-
cations, however, the mapping can provide information about change in
structures over time, or variability between individuals. In either case, the
registration transformation serves as input to some other component of a
larger system that leads ultimately to patient benefit. That next component
may be an image viewer that presents resliced volumes or surface renderings
to a physician during diagnosis or to a surgeon during image-guided treat-
ment. It may be part of a robotically controlled treatment system, such as a
bone drill in orthopedic surgery or a linear accelerator in radiotherapy, or it
may be a segmentation algorithm that requires two or more modalities to dis-
tinguish among various tissue types. In every case, the job of the registration
component is simply to deliver an aligning transformation. It is the job of the
embedding system to make that transformation clinically useful.
Regardless of the combination of views, the next component in the embed-
ding system, or the particular application, success or failure for the registra-
tion component of the system hinges on the quality of alignment of
homologous points provided by the transformation that it produces. The
alignment need not be perfect, but it must be adequate for the problem at
hand. In order to determine whether a given registration system is indeed
adequate for a given problem, or to determine whether it has performed ade-
quately for a given pair of views, it is necessary to measure the degree of
alignment. That measurement is the subject of this chapter.
6.2
Measures of Success
Our choice of the quality of alignment as the measure of success follows
directly from our definition of registration, which is the determination of
a transformation that aligns points in one view of an object with corre-
sponding points in another view of an object. We note that other aspects
