Accelerating magnetic resonance imaging(MRI)speed is of great significance for clinical diagnosis.One important way to achieve it is the compressed sensing(CS)based algorithms(CS-MRI).Although this kind of method has achieved lots of promising performances,the traditional CS-MRI has three limitations.First,the transform operator of it may be too simple to catch the complicated structural information.Besides,the selection of the regularization parameter is empirical.What's more,the iterative optimization procedure takes a long time to be convergence.With the development of deep convolutional neural network,deep learning based methods for accelerated MRI has attracted a great of attentions.These approaches can be generally divided into two categories: data-driven and model-driven methods.For data-driven methods,their effectivenesses largely rely on big data,which is difficult to obtain due to patient privacy in MRI.Different from it,another strategy,the model-driven network,can obtain a promising performance with smaller dataset owing to the prior information of the mathmatical model.In this thesis,the main efforts are made on forming an enhanced iterative network for accelerated MRI.To achieve this goal,two main innovations are listed as follows:(1)Forming a basic iterative network for impulse noise removal(IIN).This work is implemented by unrolling the iterative optimization procedure of an impulse removing medel to a basic iterative network.(2)Forming an enhanced iterative feature refinement network for accelerated MRI(IFR-Net).In this work,all parameters can be learned and the CNN-based inversion blocks are integrated into the denoising module of the network to increase its capacity.Extensive experiments on both simulated and in vivo MR datasets have shown that the proposed network possesses a strong capability on the preservation of more structural information with fast reconstruction speed. |