Novel Qualitative and Quantitative Approach to Image Registration:
Discrete Optimization Meets continuous Deformations
Nikos Pragios (Ecole Centrale de Paris, France)
joint work with Ben Glocker, Technical University of Munich
In this talk we present a framework which is similarity-independent and
bridges the gap between continuous deformations and discrete
optimization. We present a rather generic formulation that assumes a
linear/non-linear continuous deformation field using a set of control
points, and an objective function defined on these points. In particular
free form deformations (FFD) are used to represent the space of
solutions. The optimal one corresponds to the lowest potential of an
objective function that aims to capitalize content similarities between
the source and the target. These are defined on the entire image domain
and are considered along with certain smoothness constraints on the
deformation field. Several similarity metrics are considered for the
case of inter and intra modality registration. The optimal solution of
this objective function is obtained through the use of discrete
optimization in an incremental fashion capable to account with important
displacements from the source to the target with low computational cost.
Last, we introduce the concept of multi-dimensional uncertainty
information of the registration process that provides a natural
qualitative interpretation of the results. Validation and comparisons
with the state-of-the-art methods demonstrate the great potentials of
our approach.