Versatile Higher-Order Reconstruction Accelerated by a Graphics Processing Unit (GPU)

Michael Andreas Bieri¹, Christoph Barmet¹, Bertram Jakob Wilm¹, Klaas Paul Pruessmann¹

FFT-based MR reconstruction algorithms get increasingly complex when incorporating various encoding terms such as static off-resonance, dynamic higher-order fields, coil sensitivity maps, non-Cartesian k-space sampling, etc. Therefore in this work a versatile algebraic reconstruction is employed that allows for incorporating various encoding mechanisms with only minor increase in complexity. A drawback is the enormous computing time even for small images. A 100x100 image needs minutes to be reconstructed on a current high end CPU. By implementing the algorithm on a graphics processing unit (GPU), speed-up of about 600x was achieved compared to a CPU. All result are shown at the example of dynamic higher order fields combined with static off-resonance and sensitivity encoding.