Sparse Representation-Based Framework for Preprocessing Brain MRI

Abstract

This thesis addresses the use of sparse representations, specifically Dictionary Learning and Sparse Coding, for pre-processing brain MRI, so that the processed image retains the fine details of the original image, to improve the segmentation of brain structures, to assess whether there is any relationship between alterations in brain structures and the behavior of young offenders. Denoising an MRI while keeping fine details is a difficult task; however, the proposed method, based on sparse representations, NLM, and SVD can filter noise while prevents blurring, artifacts, and residual noise. Segmenting an MRI is a non-trivial task; because normally the limits between regions in these images may be neither clear nor well defined, due to the problems which affect MRI. However, this method, from both the label matrix of the segmented MRI and the original image, yields a new improved label matrix in which improves the limits among regions.