Symmetrization of the Non-rigid Registration Problem Using Inversion-Invariant Energies: Application to Multiple Sclerosis

Abstract

Abstract. Without any prior knowledge, the non-rigid registration of two images is a symmetric problem, i.e. we expect to find inverse results if we exchange these images. This symmetry is nonetheless broken in most of intensity-based algorithms. In this paper, we explain the reasons why most non-rigid registration algorithms are asymmetric. We show that the asymmetry of quadratic regularization energies causes an oversmoothing of expending regions relatively to shrinking regions, hampering in particular registration-based detection of evolving processes. We therefore propose to use an inversion-invariant energy to symmetrize the registration problem. To minimize this energy, two methods are used, depending on whether we compute the inverse transformation or not. Finally, we illustrate the interest of the theory using both synthetic and real data, in particular to improve the detection and segmentation of evolving lesions in MR images of patients suffering from multiple sclerosis.

A Tensor-Based Morphometry Study of Genetic Influences on Brain Structure Using a New Fluid Registration Method

“…Images are driven to similarity using common landmarks, or measures over the whole image such as the squaredintensity difference (L 2 -norm), cross-correlation or more complex metrics that can be derived from information-theory, such as the Jensen-Rényi divergence [6]. The transformation is constrained through a regularizer to enforce desirable properties in the deformation, such as smoothness, invertibility and inverseconsistency [5].…”

Inverse Consistent Mapping in 3D Deformable Image Registration: Its Construction and Statistical Properties

“…This dependence can be termed inverse inconsistency as inconsistency arises if we switch the order of source and target. Secondly, as pointed out in [4], these inversely inconsistent approaches penalize the expansion of image regions more than the shrinkage of image regions. This imbalance in the penalty was also noticed and discussed in another paper [5] by the same group in which shrinking brain lesions were found to be easier to detect than expanding ones using inversely inconsistent methods.…”

“…In order to numerically compute (3), we need to solve for η 2 and ξ 1 using eqs. (3) and (4). To this end, we first utilize the inverse relationship given in eq.…”

Computing inverse optical flow

“…Moreover, it cannot process occlusions correctly. In the symmetric method presented in [8], the inverse optical flow is computed using a Newton scheme. This solution is similar to the previous iterative process, so it presents the same drawbacks, providing poor results at discontinuities and occlussions.…”