MR Image Reconstruction Using PMRI and Compressed Sensing

Abstract

MR Image Reconstruction Using PMRI and Compressed Sensing Magnetic Resonance Imaging (MRI) is a safe and non-ionizing medical imaging modality that provides excellent soft tissue contrast. The main limitation in MRI is its long data acquisition time. Parallel MRI (pMRI) and Compressed Sensing (CS) provide a framework to lessen the MRI data acquisition time. Both pMRI and CS acquire less data in k-space and then minimize the aliasing artifacts by using dedicated reconstruction algorithms. PMRI was introduced in late 1990’s and some of its reconstruction algorithms are commercially available in the latest MRI scanners. CS was introduced in 2007 and it is mostly used in research. Recently, Graphical Processing Units (GPU) have been used to reduce the computation time in the parallel algorithms. GPU can also be used in MRI reconstruction algorithms to reduce the computation time of reconstruction algorithm (one of the main limitation in MRI). The objective of this thesis is to explore and develop new methods to improve the performance of pMRI and CS. There are four main contributions in this regard. (i) A Graphical Processing Unit (GPU) based design is developed in Compute Unified Device Architecture (CUDA) for fast implementation of Sensitivity Encoding (SENSE), a pMRI reconstruction algorithm. The results show that the proposed GPU implementation of SENSE significantly reduces its computation time while maintaining the reconstructed image quality as compared to the conventional SENSE implementation. (ii) Receiver coil sensitivity information plays an important role in SENSE reconstruction. Eigen-value method is used for sensitivity estimation in the proposed work for SENSE reconstruction on GPU and the results are compared with the conventional Pre-scan method. The results show that Eigen-value method considerably improves the SNR of the reconstructed images as compared to the results obtained using Pre-scan method. (iii) Conjugate Gradient SENSE (CG-SENSE), proposed by Pruessmann uses both the Non-Cartesian (NC) and Cartesian k-space for image reconstruction. NC trajectories are increasingly used in recent MR imaging because they are fast and provide efficient coverage of k-space. NC trajectories are computationally extensive and require special operations including gridding and de-gridding. A GPU based implementation of the gridding and de-gridding operations is proposed in this thesis that significantly reduces the computation time of CG-SENSE while maintaining the reconstructed image quality. (iv) A combined application of CG-SENSE (a pMRI reconstruction xi algorithm) and Projections onto Convex Sets (POCS) (a CS reconstruction algorithm) is proposed for reconstructing fully-sampled MR images from the under-sampled data. The results of the proposed method are compared with CG-SENSE and POCS reconstruction separately. The results show that the proposed method allows higher acceleration factors in MRI data acquisition and notably improves the reconstructed image quality.