Biomedical Engineering Reference
In-Depth Information
incorporate surface coil correction algorithms. These are typically based on
low pass filtering of the data to suppress image structure and obtain an esti-
mate of the underlying spatial intensity variations, which can then be used
to normalize the intensity of the original images. The result is a much more
homogeneous appearance, which is likely to avoid failures of registration
algorithms but does not preserve intensity relationships between tissues.
Recently, more sophisticated algorithms for image intensity corrections that
appear to produce a more faithful final intensity distribution have been
introduced.
8,9
(See also Figures 14.2 and 14.4 in Chapter 14).
Another discrepancy between images that may need correcting is differ-
ences in size, skew, or scaling between images or as a result of drift in scanner
calibration with time. It is advisable to calibrate the scanners to be used for
collection of images that will be registered both to check that the image
dimensions and, in the case of CT, the gantry tilt, are correctly recorded in the
header, and also to ensure that any file format conversion process correctly
preserves this information. Tilting the gantry is common in CT acquisitions
to reduce dose to the eyes or to generate a coronal slice orientation, but, since
the direction that the bed moves is not also rotated, this results in the images
being skewed (see Chapter 10, Figure 10.2). Correcting skew involves a slice-
by-slice translation, which is normally carried out prior to registration. If this
skew is not corrected, substantial errors can result. Spatial distortion of
images may also need to be addressed. These factors are discussed in the
next chapter and may be corrected as part of the registration process or as a
preprocessing step.
4.5
Image Segmentation
Under some circumstances, it may be necessary or desirable to specifically
exclude some regions of the images from the registration process. The pro-
cess of dividing images into different regions is known as image segmenta-
tion. For example, in multimodality registration of PET and MR or CT images,
the Woods algorithm
10
requires elimination of the skull and scalp from the ana-
tomical MRI or CT images in order to produce reliable results. For rigid-body
registration of single-subject, single-modality data to subvoxel precision,
Hajnal et al.
11
found it helpful to exclude mobile tissue of the face and scalp
in order to achieve the most precise results in the brain. This is because,
although the brain is well approximated by a rigid body, other tissues may
change their shape or other spatial relationships, and so the final global posi-
tional match achieved is influenced by both (see Figure 4.4). Other authors
have found that, for the image data they considered, presegmentation was
unnecessary.
12,13
In recent work on the spine, Little et al. used a composite
model in which vertebrae were treated as rigid body surrounded by plasti-
cally deforming soft tissue (Figure 4.5).
14
This concept again requires image
segmentation (of the vertebrae) as a preparation step.
