Biomedical Engineering Reference
In-Depth Information
Care must still be taken with these voxel similarity measure techniques.
Although they can be fully automatic and have an accuracy comparable to
bone-implanted markers,
24
they can also fail. One common cause of failure
can be that the images are poorly aligned at the start. For example, if the
patient is positioned very differently in the image volume for the two
scans, or if the images have different slice orientations (e.g., sagittal MR
and axial CT), the algorithms can fail unless a user provides a reasonably
good starting estimate. A further problem for some of these algorithms is
their sensitivity to the volume of overlap in the images. If one image has a
much larger field of view than the other, or if the overlap between images
at correct registration is a relatively small amount of the field of view of one
or both original images, the algorithms can fail, even with a very good start-
ing estimate. A solution to this problem is to use a normalized version of
mutual information, which is much less sensitive to image overlap,
22
as
described in Chapter 3.
10.2.2.3
Assessing Registration Accuracy
In a blind study of registration accuracy, it was shown that MR and CT
images can be registered with an accuracy of better than 1 mm but algo-
rithms can fail, resulting in errors of 1 cm or more.
24
It is, therefore, clearly
important for a method of quality assurance to be used to ensure that only
well-registered images are used for clinical decision making. The accuracy
required for most applications of MR-CT registration is about 2 mm, as nei-
ther surgery nor radiotherapy is likely to justify better accuracy. Visual
assessment of registered images can be used to check for errors, and it has
been shown that trained observers can effectively distinguish between regis-
tration errors below or above accuracy thresholds of 2 to 6 mm.
25
The sensi-
tivity and specificity of the observers to misregistration is additionally a
function of the distribution of errors produced by the registration algorithm.
The problem of ensuring good quality registration accuracy is discussed fur-
ther in Chapter 6.
10.2.3
Viewing the Combined Images
Once the images have been registered, the combined images can be viewed in a
variety of ways. These include displaying corresponding slices side by side with
a linked cursor identifying corresponding points in the slices, using color over-
lays, or segmenting bone from the CT scan and overlaying it on the MR image to
produce a combined image that has the soft tissue contrast of MR and also con-
tains bony detail. Combined display examples are shown in Color Figure 10.4.*
It is also possible to use volume visualization techniques to render an image
that combines features from the MR and CT images. An example of a ren-
dered image is shown in Color Figure 10.5.*
* Color Figures follow page 22.
