A supervised patch-based image reconstruction technique: Application to brain MRI super-resolution

The image resolution and the tissue contrast of MR neuroimaging data are key factors for image analysis algorithms such that segmentation and registration. In this article, we focus on one of these two major components and we study image reconstruction techniques for image resolution enhancement. This issue is addressed in the general context of inverse problems. For such ill-posed problems, some forms of regularization plays a crucial role. We consider a supervised regularization technique that is driven by the similarities between the input image and a learning dataset. These similarities are defined at the voxel scale and are computed using a patch-based approach. We investigate the use of such a supervised method for single-image super-resolution. Experiments on the 20 anatomical models of BrainWeb show the robustness of the proposed approach with respect to other super-resolution techniques.