Fast imaging strategies play a key and important role in our community's goal of improving value of healthcare through MR. In recent years, there has been great advances in acceleration strategies. However, many of these strategies rely on iterative image reconstruction algorithms, which are computationally expensive, thus limiting their application to clinic. Reducing computation time of these iterative reconstruction algorithms would be an important step in translating the acceleration strategies to clinic. Non-Cartesian acquisitions are an integral part of fast imaging strategies. Thus, computation time associated with non-uniform FFT has a decent contribution to total reconstruction time. The authors aim to address this need of reducing the computation time of NUFFT by exploiting computation power NVIDIA based GPUs.
Currently, the gpuNUFFT source code must be compiled, and the authors have provided instructions for Linux and Windows. I was able to compile it on a Windows as well as a Linux machine. For installation on Windows, I ran into an issue associated with Microsoft Visual Studio 10.0. However, the authors were aware of this issue and provided me a solution which worked like a charm. The authors have posted these instructions in 'Known issues' section on the software package's wiki page.
In summary, gpuNUFFT appears to be a valuable initial step towards the goal of reducing image reconstruction time for computationally expensive reconstruction algorithms. Its ease of installation and integration to existing algorithms would promote its use in MR community and might act as a catalyst for the MR researchers to move towards multi-thread parallel computing based algorithms.