Accelerating magnetic resonance three-dimensional imaging with compressed sensing and GPU computing

Compressed sensing (CS) has inspired significant interest because of its potential to reduce data acquisition time. In application of CS to MR data acquisition, 3D Cartesian sampling is much more attractive than 2D Cartesian sampling, because random sampling can be implemented in two phase-encoding directions. In addition, 3D imaging is more time-consuming than 2D imaging, so scan time reduction has more impact. In this paper, we have proposed and demonstrated the 3D CS using the 3D FREBAS transform as sparsifying transform function. In was shown that 3D CS provides images higher peak signal-to-noise ratio (PSNR) images compared to 2D CS when the signal compression rate is the same. Since CS reconstruction is an iterative reconstruction technique, especially in 3D image reconstruction so it is more computationally intensive than traditional inverse Fourier reconstruction. Here we illustrate how graphics processing units (GPU) can be used to achieve significant increases in CS reconstructions of 3D MRI data sets. We have shown that GPU dramatically accelerate CS MRI reconstruction with 3D images. Experimental results show that CS reconstruction with 256 x 256 x 64 images based on GPU was executed in 53 s while CPU computing cost was 807 s. (author)