论文部分内容阅读
In the basic compressed sensing magnetic resonance imaging(CSMRI)model with fixed sparsifying transform,fine structure information(called “feature” hereafter in this paper for short)loss is usually observed and becomes especially serious at large acceleration factors.To alleviate this problem,dictionary learning and Bregman iteration techniques have exploited the patch-based sparsity prior to adaptively capture features and updated the acquired k-space data respectively for restoration of lost features.However,dictionary learning based methods usually suffer from high computational complexity due to the huge training matrix and Bregman iterative methods may converge to a noisy image due to the semi-convergence property.In this paper,we propose a novel CS framework by introducing a feature refinement module to the basic CSMRI model without increasing complexity.Unlike the basic CS model or the Bergman iterative method,which either throws away or adds back the residual image obtained from sparsity-promoting denoising,the proposed method extracts only the useful structure information using a feature descriptor,and refines the reconstruction by restoring the features.Numerical experiments on MR images demonstrate that the reconstructed image using the proposed iterative feature refinement scheme can preserve more features compared with other CS methods.