Detecting mutually-salient landmark pairs with MRF regularization

Abstract

In this paper, we present a framework for extracting mutually-salient landmark pairs for registration. Traditional methods detect landmarks one-by-one and separately in two images. Therefore, the detected landmarks might inherit low discriminability and are not necessarily good for matching. In contrast, our method detects landmarks pair-by-pair across images, and those pairs are required to be mutually-salient, i.e., uniquely corresponding to each other. The second merit of our framework is that, instead of finding individually optimal correspondence, which is a local approach and could cause self-intersection of the resultant deformation, our framework adopts a Markov-random-field (MRF)-based spatial arrangement to select the globally optimal landmark pairs. In this way, the geometric consistency of the correspondences is maintained and the resultant deformations are relatively smooth and topology-preserving. Promising experimental validation through a radiologist's evaluation of the established correspondences is presented. © 2010 IEEE. Personal use of this material is permitted. However, permission to reprint/republish this material for advertising or promotional purposes or for creating new collective works for resale or redistribution to servers or lists, or to reuse any copyrighted component of this work in other works must be obtained from the IEEE. Author(s)Yangming ABSTRACTIn this paper, we present a framework for extracting mutuallysalient landmark pairs for registration. Traditional methods detect landmarks one-by-one and separately in two images. Therefore, the detected landmarks might inherit low discriminability and are not necessarily good for matching. In contrast, our method detects landmarks pair-by-pair across images, and those pairs are required to be mutually-salient, i.e., uniquely corresponding to each other. The second merit of our framework is that, instead of finding individually optimal correspondence, which is a local approach and could cause self-intersection of the resultant deformation, our framework adopts a Markov-random-field (MRF)-based spatial arrangement to select the globally optimal landmark pairs. In this way, the geometric consistency of the correspondences is maintained and the resultant deformations are relatively smooth and topology-preserving. Promising experimental validation through a radiologist's evaluation of the established correspondences is presented.

Graph-based Deformable Image Registration

“…In [61,91], for example, the authors used Laplacian operations to search for blob-like structures, and used the centers of the blobs as landmarks. In [58], the authors used regional centers or edges at various scales and orientations as landmarks, which were of strong response to Gabor filters. While a detailed survey of landmark detection is outside the scope of this section, we want to emphasize that the described graph-based formulation can seamlessly integrate landmark information coming from any algorithm or expert.…”

“…Another approach is to exploit attribute-based descriptors and mutual saliency [58] and define the potential as:…”

“…8 The Dense deformation fields generated by (a) M1 -no MRF regularization and (b) M2 -with MRF regularization. This figure is re-printed from [58] where p 2 V geo and q 2 V ico , l p 2 L geo and l q 2 L ico , and .p; q/ 2 E hyb . Such a pairwise term couples the displacements given by the two registration processes and imposes consistency.…”

Discrete Visual Perception

“…Contextual representation of images using either local representations [14] or graphs [15] is a well studied problem in the field. In [16] we introduced a new image representation encoding the general layout of groups of quantized local invariant descriptors as well as their relative frequency.…”

“…Continuous methods have certain strengths but also exhibit a number of limitations in particular during the inference process like for example their strong dependency from the initial conditions. This was an issue that was addressed in [42], [43], [15], [44]. The central idea was to represent prior knowledge through a point distribution representation mapped to a pair-wise probabilistic graphical model.…”