Biomedical Engineering Reference
In-Depth Information
point-like anatomical features is more user dependent and will depend on
the points selected. The disadvantage of bone-affixed external markers is that
they are invasive. Also, the markers tend to be in the periphery of the MR
image field of view, where distortion is likely to be greatest (see Chapter 5 for
further discussion of MR distortion).
An alternative to using points in the image is to use surfaces.
3,14-16
These
algorithms are described in detail in Chapter 3 and have been used for MR-CT
registration for many years. For registration of MR and CT images of the head,
the easiest surface to define in both modalities is the skin surface. The skin,
however, tends to deform between scans as a result of differences in the shape
of the head rest and head restraint between modalities. The skin surface, there-
fore, is not a very accurate surface to use. A better alternative would be either
the inner or outer table of the skull. Both these surfaces are easy to identify in
CT images, but because cortical bone is not visible in MR scans, the position of
the skull surface must be inferred from adjacent structures. For T2-weighted
MR images, CSF is bright, so the inner table of the skull can easily be found as
the boundary of bright CSF and dark cortical bone. For T1-weighted MR
images, however, CSF is dark like cortical bone, so the inner table of the skull
is difficult to identify, especially in patients with atrophic brains. In this case, an
alternative surface to delineate is the outer table of the skull. In T1-weighted
images, the fat of the scalp is bright, so the boundary between scalp and cortical
bone can be identified quite well. The problem here is that, because of the high
fat content of scalp, the scalp can appear displaced in the MR image readout
direction relative to structures of interest in the brain, just as fat-filled markers
can be displaced. Images acquired with high readout gradient strength have
less fat-water shift, so errors introduced by this can be reduced.
When registering images using surfaces, it is desirable to have as much sur-
face visible in both images as possible. Preferably, for registration of the head,
the great majority of the skull should be visible in both modalities. The skull
has quite a lot of rotation symmetry, so without sufficient coverage, surface
matching algorithms can easily converge to an incorrect local minimum. It is
possible to increase registration accuracy in these situations by combining
the use of surfaces and points.
16
10.2.2.2
Registration Using Voxel Intensity Values
Since the mid-1990s, fully automatic algorithms have been available for regis-
tration of MR and CT images of the head by optimizing voxel similarity mea-
sures.
17-22
Van den Elsen proposed an algorithm based on correlation, in which
an intensity remapping algorithm was used to make the bone in CT dark, as it
is in MR.
17
An alternative approach is to correlate ridge images extracted from
both MR and CT images, rather than the images themselves.
18
More recently,
it has been shown that theoretical approaches such as mutual information can be
used for MR-CT registration, as well as for other registration applications.
19-22
These algorithms are all described in Chapter 3. A blind multicenter study
recently found that these measures based on voxel intensity values are more
accurate than surface-based methods for MR-CT registration.
23
