Biomedical Engineering Reference
In-Depth Information
2.4.3.1
Registration of Multiple Images of the Same Patient
Acquired Using the Same Imaging Modality
There can be considerable clinical benefit in accurately aligning images of
the same subject acquired with the same modality at different times in order
to detect subtle changes in intensity or shape of a structure. This technique is
most widely used for aligning serial MR images of the brain, as discussed in
Chapter 7. Because the images are acquired using the same modality, an
approximately linear relationship will exist between the voxel intensities
in one image and voxel intensities in the other. In these cases the correla-
tion coefficient (CC)
18
is a good measure of alignment. The formula for the
correlation coefficient is presented in the next chapter, but it basically
involves multiplication of corresponding image intensities. One image is
moved with respect to the other until the largest value of the correlation coef-
ficient is found. Statistically speaking, this is where there is the strongest lin-
ear relationship between the intensities in one image and the intensities at
corresponding locations in the other. Instead of multiplying corresponding
intensities, we may subtract them, which leads to another measure, the sums
of squared intensity differences (SSD). In this case, alignment is adjusted
until the smallest SSD is found.
19
That which can be subtracted or multi-
plied can also be divided. If two images are very similar, their ratio will
be most uniform at registration. This is the basis of Woods
2
ratio image
uniformity (RIU) algorithm in which the variance of this ratio is calculated.
20
Alignment is adjusted until the smallest variance is found. In early publica-
tions this was referred to as the variance of intensity ratios (VIR) algorithm.
While the details of the formulae used are different, these algorithms are
conceptually very similar. Performance, too, is similar except when the
underlying assumptions are violated due to changes in overall image bright-
ness, shading, etc.
As the small differences in very similar images may have clinical signifi-
cance, care must be taken to ensure that the computation of the transforma-
tion neither removes nor masks this important information. Rescaling of
either size or intensity, for example, must not mask real changes in volume.
This danger can be avoided if all images are rescaled and intensities nor-
malized by reference back to an image of a standard object or calibration
phantom.
21
The most common technique for aligning these images is to find
a rigid-body transformation. Prior to carrying out the rigid-body registra-
tion, it is advisable to correct for any scaling or intensity errors in the images
(discussed in Chapter 5), and it may be necessary to carry out additional
preprocessing such as segmentation (discussed in Chapter 4). A promising
alternative approach is to register the images using a nonrigid transfor-
mation. In this case, the images after alignment should look virtually iden-
tical, and the calculated registration transformation provides quantitative
information on the parts of the images that have changed size between
images.
