Inter-Subject Non-Rigid Registration: An Overview with Classification and the Romeo Algorithm - Biomedical Image Analysis - page 273

problem. Methods that are presented here extract geometrical features from
images and compute a transformation that matches these features while inter-
polating smoothly the deformation throughout the image.

8.2.2.1 Points

Earliest methods rely on points. The most famous one, which is still a reference in
the field of neuroscience, is the Talairach stereotaxic space [130]. It has then been
extended to the Talairach proportional squaring system [132]. Both methods rely
on the identification of the anterior comissure AC and posterior comissure PC,
as well as five brain extrema which makes it possible to specify a partition of the
volume into 12 subvolumes. The transformation associated with the Talairach
proportional squaring system is a piecewise linear one that makes it possible
to embed the brain into a “box” centered at AC and whose anatomical axes are
known. This framework is known to be quite accurate in the central region but
less accurate for cortical areas.
Other authors have proposed methods based on anatomical points to register
brains of different subjects [16, 26, 50, 116]. However, the number of points that
can be reproducibly identified among a population of subjects is limited. It has
been evaluated as 36 [38] or 26 [50]. This number of points seems limited to
understand the intersubject variability; in addition to this, the extraction step
might be erroneous. To limit the dependency toward extraction, some authors
have proposed differential geometry operators to automate the process [135,
114, 115].

8.2.2.2 Curves

Gu eziec [66], Subsol [126] and Declerck [41] describe methods to register two
volumes thanks to curves: smoothing and curve matching in [66], application to
the registration of brains in [41], building of skull atlases in [126]. Crest lines,
introduced by Monga et al. [100], are defined as maximal curvature points and
can be automatically extracted using the marching lines algorithm [137].
Gueziec et al. [66] approximate curves using B-splines. This enables the
direct computation of features such as position, curvature and so on. Curves are
then registered using an iterative approach like the Kalman filter. Subsol [126]
and Declerck [41] have extended the ICP algorithm (Iterative closest points