An Example-based Approach for High-Resolution Reconstruction of Developing Brain MRI

Imaging moving subjects remains an open issue for Magnetic Resonance Imaging (MRI). Several clinical imaging protocols make use of multiple orthogonal 2D multi-planar ac-quisitions with non-isotropic voxel size for brain studies. A registration-based method to compound multiple orthogonal sets of 2D fetal MRI slices into a single isotropic high resolution volume have been proposed in [2]. It is composed of the following steps: • rigid registration of 2D slices, • relative image contrast correction, • local approach for image reconstruction. In this work, we focus on the final image reconstruction step and we propose a new example-based super resolution approach based on the non-local means (NLM) framework [1] in order to improve reconstructed image quality. Super-Resolution SR modeling The most common SR approaches consist in modeling the physical problem and to invert it. These approaches use a generic observation model such as: yr = DBWrx + nr for 1 ≤ r ≤ n (1) where yr, r ∈ {1...n} denotes the low resolution (LR) images, x is the high resolution (HR) image, nr represents observation noise, D is the subsampling matrix, B a blur matrix,Wr is the geometric transformation of rth low resolution image. The HR image is computed by considering the following equation: x ̂ = argmi