Image Processing Reference
In-Depth Information
index is extracted. The main advantage of this approach is about the
dramatic reduction of required processing time. In fact, the dimension
of extracted features is usually at least a magnitude order below the
dimension of images (i.e., lines vs. 2-D images, surfaces vs. 3-D
images). The main disadvantages are the arbitrary choice of the feature
to extract and the drawbacks in the correct extraction of the features.
2.
Similarity measure by direct comparison of images to be registered:
The main disadvantage of this approach is the required processing time.
In fact, all image data are involved in the analysis. The main advantage
is the independence from any user input. This kind of method is also
known as voxel-based (3-D) or pixel-based (2-D) methods.
Both approaches were extensively used for medical image registration. An
example of the first approach is the registration procedure for two MRI cardiac
images. We have to extract the same geometrical feature from both reference and
floating images. In the case of cardiac MR image registration, the left ventricle
contour is a natural choice and can be extracted from both images with an
automatic algorithm such as the one described in Reference 11. If the left ventricle
contours were correctly extracted from both images, the similarity between two
images can be defined as the difference between the two extracted contours,
introducing a definition of the distance between two closed curves. An optimi-
zation process can be used to minimize the previously defined distance in per-
forming the image registration. An example of the previously described procedure
is the iterative closest point (ICP) algorithm introduced by Besl and McKay [12];
it is a general-purpose method for the registration of two generic point sets
representations, including line segments sets, implicit curves, parametric curves,
and generic surfaces. At the end of the optimization process, we have the rotation
matrix and the translation vector that register the two curves with each other. The
convergence theorem guarantees the achievement of a local minimum. The roto-
translation matrix can be now applied to the floating image to perform the
registration. The example shows both the advantages and the disadvantages of
the feature-based approach: the registration operation involves 1-D data (i.e., the
extracted contour) and consequently is very fast and accurate. On the other hand,
the feature extraction (i.e., the localization of left ventricle contours) can be
difficult and error prone. In conclusion, the use of feature-based algorithms is
suggested only when fast and effective segmentation algorithms are available on
the images to be registered.
The term
voxel-based methods
implies the comparison of gray levels of the
images to be registered.
The simplest metric involves the use of difference or absolute difference
between images (mean square difference):
∑
MQ
(
A, B
)
=
(
a
−
b
)
2
(7.3)
i
j
ij S
,
∈
Search WWH ::
Custom Search