high performance computing on graphics processing units: hgpu.org

Magnetic Resonance Imaging (MRI) is a non-invasive and highly exible medical imaging modality that does not expose patients ionizing radiation. MR Image acquisitions can be designed by varying a large number of contrast-generation parameters, and many clinical diagnostic applications exist. However, imaging speed is a fundamental limitation to many potential applications. Traditionally, MRI data have been collected at Nyquist sampling rates to produce alias-free images. However, many recent scan acceleration techniques produce sub-Nyquist samplings. For example, Parallel Imaging is a well-established acceleration technique that receives the MR signal simultaneously from multiple receive channels. Compressed sensing leverages randomized undersampling and the compressibility (e.g. via Wavelet transforms or Total-Variation) of medical images to allow more aggressive undersampling. Reconstruction of clinically viable images from these highly accelerated acquisitions requires powerful, usually iterative algorithms. Non-Cartesian pulse sequences that perform non-equispaced sampling of k-space further increase computational intensity of reconstruction, as they preclude direct use of the Fast Fourier Transform (FFT). Most iterative algorithms can be understood by considering the MRI reconstruction as an inverse problem, where measurements of un-observable parameters are made via an observation function that models the acquisition process. Traditional direct reconstruction methods attempt to invert this observation function, whereas iterative methods require its repeated computation and computation of its adjoint. As a result, naive sequential implementations of iterative reconstructions produce unfeasibly long runtimes. Their computational intensity is a substantial barrier to their adoption in clinical MRI practice.