GPU-Accelerated Self-Calibrating GRAPPA Operator Gridding for Rapid Reconstruction of Non-Cartesian MRI Data

A survey of GPU-based acceleration techniques in MRI reconstructions

“…If current PMRI methods are simply analyzed, they can be roughlyclassified into two types (50): one is the reconstruction procedure in image space which includes unfolding operation and inverse procedure, for example, SENSE (52); another is the reconstruction procedure in k-space, which has kernel calculation and recovery procedures of missing k-space data, for instance, SMASH (51) and GRAPPA (53). SENSE and SENSE derivative methods have been implemented in GPUs (29)(30)(31)(32)(33). For example, the GPUbased implementations of Cartesian SENSE and k-t SENSE have been presented by Hansen et al in Ref.…”

“…Their implementation claimed their reconstruction performances on 32 coils could achieve 42 ms acquisition time, 11.2 ms reconstruction time for under-sampled radial datasets, and their methods could be utilized into more challenging reconstruction scenarios which have larger numbers of acquisition coils, higher acceleration rates, or more GPUs than before. Furthermore, Inam et al proposed an acceleration method for Selfcalibrating GRAPPA operator gridding by using massively parallel architecture of GPUs (33). The LUTs were used to pre-calculate all possible combinations of gridding weight as well as avoid the race condition among the CUDA kernel threads.…”