The simulation results indicate clearly that the registrations computed by
improve substantially when using the extra cortical features. The
random deformations used to simulate different anatomies (section
2.1) only model one aspect of anatomical variability, namely
the second-order morphological difference between subjects (i.e., a
difference in position for a given structure after mapping into a
standardized space). The simulations do not account for different
topologies of gyral and sulcal patterns. For example, it cannot change the
region corresponding to Heschl's gyrus, to contain double gyri when the
phantom has only a single. Still, the value of these simulations is not
diminished since they can be used to determine a lower-bound on the
registration error -- corresponding to the optimal case when a 1-to-1
mapping exists between two subjects.
The experiments with real data indicate that the use of automatically
extracted and labelled sulci in conjunction with a chamfer distance
function can significantly improve cortical registration. One must keep in
mind that the
measure includes possible sulcal mis-identification and
sulcal extraction errors (from 
) in addition to
's
mis-registration error. We are working on quantifying the performance of
in order to improve the error measure. Further improvements to the
registration algorithm will require a slight change in fitting strategy to
account for different cortical topologies apparent in real MRI data. We
envision using a number of different targets simultaneously, where each
target will account for a particular type of sulcal pattern [35].
The current version of
now allows the use of a chamfer-distance
objective function to align sulci, however nothing in the implementation is
sulci-specific. Indeed, any geometric structure that can be voxelated can
be incorporated into the matching procedure to further refine the fit. We
are currently evaluating the incorporation of explicitly extracted cortical
surfaces [28,29] into the registration process.