This talk presents a recently proposed variational strategy to learn spatially-varying metrics in the LDDMM diffeomorphic image registration framework. The metric model we use makes possible to smooth the deformations with different levels in different image regions. Smoothing can also be performed locally as well as globally. It can finally take into account the direction of the deformations.
Emphasis will be given to how spatially-varying metric parameters are learned from a set of reference images. The key motivation of our strategy is indeed to make computationally reasonable this learning step on 3D medical images.