Learning Optical Flow for Fast MRI Reconstruction
Résumé
In this talk, we present a new model for the reconstruction of high quality medical images from few measurements. This novel model is built on traditional compressed sensing methods [Lustig, 2007] with addition of motion compensation by an optical flow approximation [Aviles-Rivero et al., 2018]. Our model shows that the optical approximation can be improved by a sparse representation in a dictionnary of reference samples. First, we introduce the classical compressed sensing methods along with motion compensated algorithms depicted in [Aviles-Rivero et al., 2018]. Then we show how the optical flow can be learnt from reference samples like in [Jia et al., 2011]. Finally, we embed in a single functional the image reconstruction process and optical flow approximation by the dictionary. High efficiency algorithms are given for the optimization of the functional. Numerical experiments are performed to demonstrate the effectivness of this new method which reaches state of the art among variational methods.
Origine | Fichiers produits par l'(les) auteur(s) |
---|---|
Licence |