Section: New Results

Fusion/Registration

Participants : Stavros Alchatzidis, Nicolas Honnorat, Fabrice Michel, Aristeidis Sotiras, Chaohui Wang, Alex Bronstein, Michael Bronstein, Christos Davatzikos, Ben Glocker, Nikos Komodakis, Yangming Ou, Dimitris Samaras, Regis Vaillant, Yun Zeng, Nikos Paragios.

• Intrinsic Dense 3D Surface Matching: in [38] a probabilistic tracking framework for registering two 3D shape that relies on accurate correspondences between all points across the two frames was proposed. The definition of the matching cost is done using the "uniformization" theory that is combined with regularization terms that enforce spatial and temporal motion consistencies, into a maximum a posteriori (MAP) problem which we approximate using a Markov Random Field (MRF).

• Optimal Linear Registration: in [26] we proposed a novel formulation to address linear registration of volumetric images (translation, rotation and scale) that guarantees the optimality of the obtained solution. This was achieved through the approximation of the volumetric data using a sparse representation and the expression of the registration criterion in the form of a difference of convex functions. Cutting plain algorithms in the high-dimensional space were used to provide the optimal solution of the registration problem.

• Quasi-real Time Registration: in [21] we proposed a novel message-passing based optimization method to for pair-wise Markov Random Fields models and their applications in medical imaging and computer vision. Such a method was integrated to the deformable registration paradigm introduced in [12] . Such an optimization framework was combined with efficient use of modern architectures (Graphics Processing Units) leading to a speed up of at least one order of magnitude with respect to [12] making quasi real-time deformable registration feasible.

• Metric Learning: in [31] we extend prior work on similarity sensitive hashing to address multimodal 3D registration. The method consists of combining invariant to translation/rotation/scale features defined at the Gabor space with a machine learning/boosting method that aims to projection corresponding visual patterns to binary vectors with minimal Hamming distance while maximizing the distance between no corresponding samples.

• Symmetric Deformable Fusion: in [9] a novel graph-based formulation combining image and geometric terms was proposed for deformable registration. The method aimed at constraining iconic registration using a set of landmark correspondences that are sparse, do not inherit redundancy and are symmetric. The central idea was to simultaneously deform the target and the source image using two symmetric flows such that the similarity criterion is reaching its lowest potential. This was achieved through the use of composite symmetric deformation fields. This formulation was expressed as a graph-based optimization problem leading to promising experimental results.

• Deformable registration of gene expression data: in [28] the combined iconic/geometric registration framework introduced in [9] was extended to deal with gene expression data. Similarity Sensitive Hashing was used to establish costs for landmark correspondences, and a graph-based formulations with unknowns the deformation vectors was adopted for the objective function. Such an idea was extended to deal with combined segmentation/registration approach through an atlas in [29] where subdivision surfaces were considered to represent the deformation grid.