The Study Of Real-time Online Dynamic Magnetic Resonance Imaging Technique

Magnetic Resonance Imaging (MRI) with advantages of high resolution of soft tissueand no radiation, after years of development, has become an important tool for clinicalexamination. But the long data collection time has become the bottleneck of its widely used.Therefore, how to generate high-quality images through the under-sampled data has becomea hot topic in the field of MRI.Donoho et al proposed a new signal reconstruction technique which is namedCompressed Sensing (CS) in2004. If the signal has a sparse representation in a transformdomain, using compressed sensing technology can make it recovery when the sampling rateis far below the Nyquist sampling theorem. We can reconstruct the images good enoughwhen we just collect a small amount of MR signal using CS reconstruct, it can save moretime. Dynamic MRI can be used its continuity information in the time direction to transformitself into a sparse representation, which provides the condition for the application ofcompressed sensing. This paper is mainly focused on the problems of real-time onlinedynamic MRI technique and its combining with video coding technique and accelerated byGPU finally:(1) Study the theory of Motion prediction used in real-time online dynamic MRI. Thecurrent frame can be predicted by Difference Calculation (DC) between previous frames(reconstructed). The difference between current frame (to be reconstructed) and the predictframe was sparse, we can use compressed sensing reconstruction to get the desired results.Experiment results show that the predict frame can improve the quality of reconstructiveresult.(2) In order to obtain sparser signal, an improved Motion Estimation (ME) and MotionCompensation (MC) method was proposed to predict the current frame from previousreconstructed frames with an extrapolation procedure. An overlapped block motioncompensation algorithm was used to suppress the block artifacts. The sparse residual signalwas used to reconstruct the current frame using an iterative soft thresholding algorithm. Theexperiment results show that, the ME/MC prediction can improve the quality ofreconstructed frames at slight additional computational cost. For the case that ME/MCcombined with previous DC method, a high quality image reconstruction can be achievedwith relatively small time consumption. (3) Real-time online MRI needs images reconstructed fastly, in this paper we use GPUto improve the speed of image reconstruction. In sparse MRI image reconstruction, the CSbased algorithm contains a large number of traditional inverse Fourier reconstruction whichis suitable for paralle. It can be speed up through GPU. The Parallel Computing Toolbox ofMATLAB is easy to use, and solving this problem effectively. The experimental results showthat, reconstruction based on the GPU can run the iteration fast which makes our systemapproach real-time more nearly.