Studies On Synchronized Compressed Sensing Reconstruction In Magnetic Resonance Imaging

Compressed Sensing (CS) has been used widely in dynamic MRI due to its ability to accelerate the scan process and improve time resolution. However, for success in traditional clinical MRI scans of static anatomy, CS is still waiting for breakthroughs in qualities of the reconstructed images and the speed of CS reconstruction. Synchronized scan and reconstruction provides a potential solution for these problems. In this work, we studied several issues related to synchronized reconstruction, including the accelerating of CS reconstruction and designing of new dynamic sampling schemes.Synchronized reconstruction, as its name suggests, conducts the CS reconstruction using the partially sampled k-space data during the process of image scan. The newly acquired data are continuously feed into the reconstruction process to yield better reconstruction results over time. Obviously, the reconstruction algorithm should be fast enough to be synchronized with the data acquisition. However, the commonly used solvers, such as nonlinear conjugate gradient (NCG) and Split Bregman method, are not fast enough. In this work, we make use of the powerful computing capability provided by Graphics Processing Unit (GPU) to accelerate the NCG solver by factors of up to 36. This increase in speed makes synchronous reconstruction possible.The sampling pattern has great influence on the quality of CS reconstructed images. The traditional variable-density random sampling pattern is based on a fixed probability density function (PDF) calculated from a prescribed sampling rate. However, synchronized reconstruction allows the user to interrupt the scan-reconstruction process based on the quality of the reconstructed images, so the sampling rate is not necessarily fixed. In this paper, we present a new dynamic sampling scheme based on variable density random method and then try to use the knowledge of k-space energy distribution in the process. It is proved that a two-phase variable density random sampling process if more suitable for use with the synchonized reconstruction.Since the synchronized reconstruction can provide information about the reconstructed image while scan is in process, we present a new adaptive dynamic sampling scheme. We used the reconstructed data to provide a guidance for the choice of phase encoding lines to be sampled next. Experiment results show that the proposed method not only has a good incoherence with the sparsifying transform, but also achieves better reconstruction results.