Biomedical Engineering Reference
In-Depth Information
method that uses the entire data. Despite these advantages, these methods ap-
pear limited in the context of inter-subject registration. As a matter of fact, the
number of features that can be reproductively identified among a population
of subjects is limited compared to the inter-subject variability. Furthermore, a
lot of information present in the data are not used by geometric methods while
photometric take advantage of all information available. This rapid comparison
may explain the popularity of photometric methods, which has been proved in
the particular context of rigid multimodal fusion [147].
Photometric methods differ by numerous aspects. Among them, two impor-
tant ones are the similarity measure and the regularization.
The choice of the similarity is crucial since this models the interaction be-
tween the data and the estimated variables. Roche
et al.
have shown [113] that
the choice of a similarity can be guided by the a priori knowledge that we have
about the data. Regularization is also crucial since it expresses the a priori
knowledge that we have about the deformation. The choice of a correct regular-
ization in the context of inter-subject normalization is difficult and still debated
since we do not know what should be the “ideal” deformation field between
two brains of two different subjects. Regularization often conserves the topol-
ogy of brain structures. While valid for internal structures such as ventricles,
the conservation of topology is a strong hypothesis when dealing with corti-
cal structures. Anatomists have indeed shown that cortical sulci have different
shapes and topology among individuals [103].
Recently there has been an increasing number of promising methods
[22, 29, 36, 68, 73, 79, 141] that combine the benefits of photometric and ge-
ometric approaches to register brains of different subjects. In these methods,
landmarks are used to drive the registration process so that the deformation
field is consistent with the matching of sparse anatomical structures.
8.3
Romeo: Robust Multigrid Elastic
Registration Based on Optical Flow
8.3.1
Introduction
We consider the registration problem as a motion estimation problem, which
has been studied by different authors [7, 9, 10, 11, 13, 31, 76, 82, 102, 120]. Our
