Magnetic Resonance Imaging Research Based On Compressed Sensing

In recently years, MRI(magnetic resonance imaging) based on compressed sensing has become a hotspot with the development of compressed sensing. The reconstructed image using MRI has a high resolution and can help doctors to diagnose better. However, MRI has a slower imaging speed which restricted its further development. With the premise of image resolution, improving imaging speed by reducing sampling data has become the core problem.However, the reconstructed image will occur artifact if the sampling frequency bellows Nyquist frequency. Compressed sensing provides a new idea for breaking the Nyquist frequency. Therefore, this thesis manly focuses on compressed sensing-based MRI methods after going into the theories of compressed sensing and MRI. In this thesis, the main research work is arranged as follows:(1) It will respectively result in the losses of image details or overall structures of MR images when the compressed sensing-based methods use the global sparse dictionaries or local sparse dictionaries separately. In order to solve this problem, this thesis has proposed a novel imaging model combining both local and global sparse constraints to capture details and overall structures of MR images. Firstly, the local sparse representations are trained from specific image. Secondly, traditional analytical dictionaries are used to promote global sparse performance of the MR images. Finally, the reconstruction is solved using a nonlinear conjugate gradient with known local and global sparse constraints. This procedure is repeated iteratively to improve the qualit y of reconstruction. Experimental results demonstrate a better image quality comparing with the DLMRI(dictionary learning MRI). Meanwhile, this thesis describes several important parameters of the imaging model.(2) With the consideration of the fact that the system noise in real-time dynamic MRI do not fit Gaussian, this thesis has proposed an model which reconstructs images using a hybrid optimization procedure that combine motion and sparse constraints. The proposed model uses the variance of the voxels to derive estimation for weighting matrix and is implemented in the compressed sensing framework. Experimental results demonstrate a better image quality comparing with the Kalman filter based method on simulated phantom and cardiac MRI cine datasets. Future work will focus on better motion estimation methods.