Research On Parallel Processing Of Sparse MRI And Direct Fourier Transform Reconstruction Algorithm

Magnetic resonance imaging(MRI) have many advatages like ionizing radiation free, multi-parameters imaging,functional imaging and ability of producing tomograph in arbitrary plane.MRI has widely been used in clinic as one of the most important imaging method.Meawhile,despite of the advatages,low scaning speed is a bottleneck of MRI bringing much uncomfort to the patient caused by long time of waiting.additionally,it is hard to put MRI into use of advanced imaging that requires high speed like dymatic imaging.Speeding up the imaging has always been an important subject since MRI emerged. Non-Cartesian scanning methods like Spiral,Radiational,Propellor have are finely used in practice.A newly arised method Sparse Sampling MRI-called SparseMRI for short-is now a in-lab research subject as a hotspot.Data acquired with Non-Cartesian scanning methods is usually been resampled to uniform cartesian points by some interpolation way and then reconstruct the image by the classical fft method.Though many interpolation methods are ripe enough for recently use,the Conjugate Gradient using Direct Fourier Transform(DFT) without interpolation is considerd the most accurate. But the algorithmic complexity problem make the DFT method still in-lab and be used only for generate reference images in experiments for design of other algorithms but not for clinic. The newly arised SparseMRI is base on the theory of compressed sencing,the image reconstruction problem is to find a sparse or transform spase root from a highly illed equation set.This needs iteration,when using the Conjugate Gradient Descend algorithm,each step of the iteration have many operations on matrices such as space thrasform and gradient.Therefore it is also a time-consuming method.The way to solve time consuming problem is Parallel Computing.A large task could be split into multiple sub tasks,the sub tasks could be distributed to multiple processors that in a single computer or distributed in multiple computers.Classically,a parallel computer may be a multi-CPU computer or a computer cluster.These kind of super computers do not fit clinical use due to huge size and large quantity of power useage.The recently emerged GPGPU(General Purpose Graphic Processing Unit) reduced the size and power useage by integrating all hardware needed by grid computing onto a single circuit board.From another point of view GPGPU is a improvement of the conventional GPU.The conventional GPU could only process graphic tasks according to a graphic pipeline.Though every stage of the pipeline is highly parelelled,it is not easy to be exploited.GPGPU is a programmable parallel processor which surports high level language like C/C++ and Fortran.In this article,conjugate phase algorithm is firstly analysed for parallel decomposition.Then the algorithm is implemented on a multi-core CPU and a GPGPU,and 5 and 72 times of speed-up ratio are gaind respectively.The SparseMRI algortithm is a complex task with many sub procedures.In this article the sub procedures are analysed for parallel and implemented on GPGPU.The control of th entire iteration do not have much computation load and is made by the CPU of the Computer.Compared with serial implementation running only on the CPU,a speed-up ratio of 76 is gained.