Adaptive Compressed Sensing MRI

INTRODUCTION Most of the prior work in compressed sensing MRI has been based on pre-defined transformations to sparsify image representations, e.g. discrete cosine transforms (DCT), wavelets, and finite differences [1]. Even though an analytical transformation usually features a fast implementation, its performance as a sparsifying transform is limited by the underlying basis functions, which tend to be over-simplistic for real images. Recent work on sparse signal representation based on learning the basis functions from the signal itself has demonstrated improved performance over analytical functions. For example, the K-SVD method [2] uses image patches to get very sparse representations. This method was originally proposed for image denoising. In this work, we propose to adapt the sparsifying transform during the reconstruction process based on the K-SVD method in order to increase sparsity of image content and thus enable higher accelerations and/or improve image quality.