GPU accelerated high-dimensional compressed sensing MRI

Recently, we have developed a tensor-decomposition based compressed sensing (CS) method for dynamic magnetic resonance imaging (dMRI) [1]. The proposed CS-dMRI method exploits the sparsity of the multi-dimensional MRI signal using Higher-order singular value decomposition (HOSVD). Our preliminary study indicates that, compared with conventional approaches, the proposed CS method offers further acceleration in acquisition and also improves image quality. To further enhance the algorithm efficiency, in this work, we present a parallelized implementation of the HOSVD-based CS reconstructions using a graphics processing unit (GPU). The cine cardiac MRI study indicated the efficiency and accuracy of the GPU-accelerated high-dimensional CS-dMRI method.