Static Magnetic Resonance Imaging Based On Complex Convolutional Neural Networks

Prolonged acquisition time is one major obstacle of magnetic resonance imaging(MRI),so fast MRI is one of the most important research areas in MR filed.Parallel MRI based on multi-coil array and fast MRI based on compressed sensing(CS-MRI)are two major stages in this area.In recent years,neural network has become an effective approach of fast MRI with the development of big data and deep learning.However,most of these fast-MRI methods based on deep learning do not deal with the complexvalued MR data appropriately,and also lack of explanation.Moreover,the reconstruction results of these methods are blurry and fail to capture details in MR images.To solve these problems,we conduct this study of fast MRI based on convolutional neural network(CNN).(1)Firstly,we review the background and current methods of static fast-MRI,and elaborate the theory and main applications of neural networks.(2)This thesis develops a deep cascaded CNN for parallel imaging.Considering the complex property of MR data,we combine the arithmetic of complex numbers with the convolution operation,and design a complex convolutional layer,and then apply it to the deep cascaded CNN.In addition,a data-consistency layer follows each block of the network to guarantee the data fidelity and improve the accuracy of the final reconstruction.Experimental results show that the proposed network can obtain much better reconstruction under different sampling patterns and different sampling ratios compared to the classical parallel imaging algorithms.(3)Loss function plays an important role in the training of neural networks.Considering several metrics for image quality,this thesis designs different loss functions,and explores the impact of them on the neural network-based static fast-MRI.