Mutually coherent structural representation for image registration through joint manifold embedding and alignment

摘要

In this paper, we introduce mutually coherent structural representations (McSR) for image registration which learns a mapping of local structural descriptors to create a unique scalar representation, which is similar across modalities. The McSR is learnt using joint alignment and embedding of Laplacian eigenmaps using modality-combination specific dense structural descriptors. The resulting alternate image representation offers richer structurally-driven information for registration while being invariant to inter-modal differences in intensities. The proposed formulation has been evaluated for robustness and registration error on standard multimodal brain image datasets. It is observed to demonstrate superior systematic recovery and performance over comparative simultaneous registration methods.