Biomedical Engineering Reference
In-Depth Information
8.2.3.8
Non-Rigid Multimodal Registration
Although many efforts have been made to perform rigid multimodal registration,
as far as we know, there has been few research concerning non-rigid multimodal
registration. As a matter of fact, this is quite a challenging problem, since the
number of variables to be estimated can be very large (intensity mapping, and ge-
ometrical transformation, the two being dependent). Two different approaches
have been developed.
One option is to estimate the geometrical transformation with the original
intensities of the two images to be registered. In this category, Maintz
et al.
[89]
and Gaens
et al.
[57] proposed an algorithm that seek a non-rigid transformation
by maximization of mutual information. They use a “block-matching” minimiza-
tion scheme with a Gaussian filtering of the estimated deformation field to avoid
blocky effects. On local windows, the estimation does not take into account the
spatial context of the deformation field and only a translation is estimated. Fur-
thermore, these methods are only performed in 2D. Rueckert
et al.
[117] and
Kybic
et al.
[84] proposed an approach based on cubic B-splines and mutual
information. The spline deformation model intrinsically contains the regulariza-
tion and provides a smooth interpolation of the field. Displacement of the nodes
are computed such as to maximize the similarity measure (mutual information,
or normalized mutual information).
Another appealing option has been proposed by Guimond
et al.
[63]. This
method considers the multimodal registration problem as a monomodal registra-
tion problem, and therefore estimates alternatively an intensity correction and
a monomodal registration. The originality of the method resides in the decom-
position of the problem into two “easier” ones: a polynomial intensity mapping
and a monomodal registration problem based on the demon's algorithm [136].
8.2.4
Discussion
This section has presented a brief overview of non-rigid registration techniques.
Methods have been arbitrarily classified into two groups: geometric methods that
rely on the extraction and matching of geometrical features; and photometric
methods (or intensity-based) that rely on the luminance information directly.
Geometric methods are attractive because they rely on anatomical features.
The deformation is expected to be consistent in the vicinity of features that
are used. In addition, the complexity is significantly reduced compared to the
