Super-resolution reconstruction of cardiac MRI using coupled dictionary learning

High resolution 3D cardiac MRI is difficult to achieve due to the relative speed of motion occurring during acquisition. Instead, anisotropic 2D stack volumes are typical, and improving the resolution of these is strongly motivated by both visualisation and analysis. The lack of suitable reconstruction techniques that handle non-rigid motion means that cardiac image enhancement is still often attained by simple interpolation. In this paper, we explore the use of example-based super-resolution, to enable high fidelity patch-based reconstruction, using training data that does not need to be accurately aligned with the target data. By moving to a patch scale, we are able to exploit the data redundancy present in cardiac image sequences, without the need for registration. To do this, dictionaries of high-resolution and low-resolution patches are co-trained on high-resolution sequences, in order to enforce a common relationship between high- and low-resolution patch representations. These dictionaries are then used to reconstruct from a low-resolution view of the same anatomy. We demonstrate marked improvements of the reconstruction algorithm over standard interpolation.