gpuNUFFT – GPU based fast image reconstruction

In recent years, there has been great advances in accelerating MRI acquisitions. Sampling the data non-uniformly is often involved in this acceleration strategies.  However, computing FFT of non-uniformly sampled data is requires long computational times, which prevents translating the acceleration strategies to the clinic. gpuNUFFT can significantly speed up this process by exploiting the computation power of NVIDIA CUDA GPUs.

gpuNUFFT is a library mainly written in C++ with a Matlab mex interface. It can be integrated in any Matlab based image reconstruction program by adding few simple changes to the code. gpuNUFFT requires of Matlab, a CUDA capable graphics card and working installation of CUDA Toolkit. gpuNUFFT supports anisotropic 3D/2D grid sizes and arbitrary sector widths. Moreover, the implementation is able to process multi-channel/multi-coil k-space and image data automatically with the respective sensitivity maps.

gpuNUFFT promotes the use of non-Cartesian 2D/3D trajectories in the context of iterative image reconstruction algorithms by reducing the long computation time. It has been successfully employed with compressed sensing parallel imaging strategies speeding up the reconstruction up to a factor of 40.