Sparsity Level Constrained Compressed Sensing for CT Reconstruction

Abstract

It is a very hot topic to reconstruct images from as few projections as possible in the field of CT reconstruction. Due to the lack of measurements, the reconstruction problem is ill-posed. Thus streaking artifacts are unavoidable in images reconstructed by filtered backprojection algorithm. Recently, compressed sensing [1] takes sparsity as prior knowledge and reconstructs the images with high quality using only few projections. Based on this idea, we propose to further use the sparsity level as a constraint. In the experiments, we reconstructed Shepp-Logan phantom with only 30 views by our method, TVR [2] and stand ART [3] respectively. We also calculated the Euclidean norm of the reconstruction image and the ground truth for each method. The results show that reconstruction results of our method are more accurate than the results of total variation regularization (TVR) [2] and stand ART [3] method.