Acceleration of l 1 -Regularized SPIRiT MRI Reconstruction by Fast DWT and GPU Computing

We propose a novel GPU implementation of the Discrete Wavelet Transform (DWT) to accelerate compressive sensing-based MRI reconstruction. Routinely, the implementation of DWT involves two kinds of operations: pixel-dependent computation and image-size-dependent computation. In existing implementations of compressed sensing-based MRI reconstruction, DWT is typically implemented by general-purpose GPU libraries such as CUDA SDK, which performs both pixeldependent computation and image-size-dependent computation at each execution. In this poster, we take advantage of the fact that the image size is fixed in MRI reconstruction, and thus we can perform the imagesize-dependent computation only once. Computational results show that the resulting implementation of fast DWT is more than two times faster than the existing DWT library, and fast DWT significantly accelerates SPIRiT based MRI reconstruction.