A Research Of Rapid Magnetic Resonance Imaging Based On Compressed Sensing

Magnetic resonance imaging(MRI) is a key imaging modality in modern medical iconography. As a noninvasive and nonionizing imaging technique, MRI enables excellent visualization of anatomical structures and physiological mechanisms in both spatial and temporal dimensions. Owing to these advantages, MRI benefits multiple clinical applications, such as the computer-aided diagnosis, vision-guided surgery and pathomechanism exploration. However, as the acquisition speed of data is limited by constraints including nuclear relaxation times and peripheral nerve stimulation, there always exists a trade-off between spatial and temporal resolution for present MRI scenario. Fortunately, it was empirically shown that for MRI sequences both the spatial and temporal dimensions present considerable information redundancy. This property makes the application of compressive sensing(CS) based MRI(CS-MRI) repeatedly successful. By reducing the amount of used sampling data through exploiting spatial and temporal correlations, CS-MRI provides an effective and efficient solution to acquire dynamic MR sequences of high spatiotemporal resolution.This paper focuses on CS based dynamic MRI(CS-DMRI) algorithm and obtains the following achievements:(1) The hybrid dynamic MRI based on generalized multi-structural total variation regularization(H-DMRI). Present CS-DMRI methods can be summarized to two categories: the offline and the online. Offline CS-DMRI models always present relatively high recovery accuracy, but they suffer from slower speed and complex system setup as well. The online methods can achieve high-speed restoration but demonstrate relatively worse image quality simultaneously. This study proposes a novel hybrid CS-DMRI restoration model(H-DMRI) to accelerate CS-DMRI with high fidelity from subsampled measurements. In the prediction process, a novel periodic subsampling scheme with variable sampling ratio in each periodic is introduced to promote reference reliability, then H-DMRI designs the periodic forward and backward reference predication scheme to enhance estimation quality, which can make the most of temporal correlations. To promote reconstruction accuracy, this study proposes a new joint spatial and temporal sparsity inducing norm called generalized multi-structural total variation(GMTV). GMTV aims at exploring the poly-directional texture similarity to make the most of information redundancy in spatio-temporal domain. Based on GMTV, H-DMRI establishes a novel residual and image joint sparsity constraint reconstruction model. This method not only emphasizes the spatial correlations of residual signal between the frames to be restored and the prediction frame, but also demands the structural continuity in image domain along both of the spatial and temporal directions. The experimental results show that H-DMRI outperforms previous online and offline methods for distinct details and fewer artifacts, and presents superiority over the existing offline methods in terms of reconstruction efficiency.(2) CS-DMRI based on low-rank and sparse components separation. The video-based research and applications always refer to the separation of background and foreground components. Considering dynamic MR images sequence possess same trait like the video, this study employs the robust principal component analysis framework to establish a multi-coil CS-DMRI model, the proposed method aims to solve restoration distortion along the temporal direction via low-rank plus sparse prior. The proposed method emphasizes the low rank of background along the temporal direction, and demands the sparsity of foreground in spatial and temporal infinite difference domains. Experimental results show that the proposed method enables desirable restoration, in terms of abundant and distinct details in spatial image and reliable preservation of temporal texture variation.