Image Registration Methods in MRI - Advance Image Processing in Magnetic Resonance Imaging - page 184

and local curvature extrema. In the segmentation-based methods, two
lines or surfaces sets (i.e., image features) are extracted from both
images and used as input for the alignment procedure. The voxel-based
methods operate directly on the image gray values. In principal axes
and moment-based methods, the image content is reduced to a repre-
sentative set of vectors, and the registration is performed using the
extracted vector set. The methods using the full image content attempt
to perform the registration maximizing the cross-correlation, the mutual
information of some other relationship between images. Voxel-based
registration does not generally require extensive preprocessing, such as
segmentation or feature extraction.

About the interaction, the registration methods can be divided into:

1. Automatic, when the user only supplies the algorithm with the image

data

2. Semiautomatic, when the user has to initialize the algorithm perform-

ing the segmentation or have to accept or reject suggested registrations

3. Interactive, when the user does the registration himself, helped by the

software

In automatic or semiautomatic registration algorithms, there are generally
three main aspects:

1. The search space is the class of potential transformations, such as rigid,

affine, and elastic, used to align the images. Three-dimensional (3-D)
rigid-body registration has six degrees of freedom: x, y, and z transla-
tion and rotation about x, y, and z axes. Affine transformations add
shearing and scaling. The most general class of transformation, elastic,
or nonlinear registration, has in theory infinite degrees of freedom.

2. The similarity metric is an indicator of how well the features or intensity

values of two images match. The sum of squared intensity difference [7],
generalized correlation coefficient [8], ratio image uniformity, and infor-
mation theoretic measures [2,3,9] are commonly used similarity measures.

3. The search strategy optimizes the similarity metric. Examples include

local or global searches, multiresolution approaches, and other optimi-
zation techniques.

In this chapter, we will first formulate the registration problem, focusing on
rigid 3-D registration. Although nonlinear registration is more realistic in principle
because tissues are deformable in some manner, rigid registration is often used in
unimodal registrations, which is a field of particular interest in MRI. Moreover,
when a large set of data is involved, as usually happens in MRI (e.g., fMRI, cardiac
imaging), nonlinear registration requires excessive computational power. In the
following text, similarity metrics are discussed with a focus on the mutual infor-
mation measure, which is most often used in the MRI registration field. We will