Study Of Fast Magnetic Resonance Imaging Technology Based On U-shaped Neural Network

Magnetic Resonance Imaging(MRI)technology is very popular in clinical applications because of its high soft tissue resolution and no ionizing radiation damage to the human body.However,due to the slow imaging speed of MRI,the physiological motion of the subject often causes imaging artifacts,which low down the quality of real-time imaging.Therefore,how to fast the imaging speed of MRI is one of the hot topics in MRI study domain.Reducing the number of k-space data samples for MRI is an effective acceleration method,but a large reduction in data collection will result in a significant degradation in image quality.Researchers can use a variety of reconstruction algorithms,such as compressed censing techniques,to improve the reconstruction quality of under-sampling data,but may require longer reconstruction times which makes it difficult to meet the clinical needs.Magnetic resonance imaging based on deep learning technology with convolutional neural network has the advantages of off-line training and online fast imaging.With the support of high-performance computing hardware,the convolutional neural network can use a large number of undersampling magnetic resonance data and full-sampling magnetic resonance data for training.Using multiple back propagation calculations,the parameter-optimized convolutional neural network is obtained.The under-sampled image uses a trained network for fast,high quality reconstruction.Deep learning technology does not require the sparseness performance,and can quickly reconstruct under-sampled images using off-line trained networks.Therefore,it may have a better application in magnetic resonance image domain.This article mainly does the following work:(1)Study the rapid magnetic resonance imaging methods based on U-shape neural network.By using the Mean Squared Error as the loss function of backpropagation,using Stochastic Gradient Descent as the optimization algorithm,a U-net convolutional neural network was constructed.A large number of under-sampled and full-sampled magnetic resonance data are trained to obtain the optimized network parameters,then the under-sampled data is put into the trained network to obtain an output reconstructed image.The simulation results show that this method can effectively improve the reconstruction quality compared to the zero-fill reconstruction algorithm and to the compressed sensing algorithm,and can quickly reconstructed a lower under-sampled data.The reconstruction time is much shorter than the compressed sensing algorithm,and can satisfy the requirements of real-time online imaging.(2)Study the rapid magnetic resonance imaging based on the recurrent residual U-shape neural network.For the lack of high frequency details in U-net reconstruction,the recurrent residuals blocks were used as the up-sampling and down-sampling blocks to improve U-net network architecture.The residual module is used to reduce the difficulty for training deep networks,and the recurrent module is used to ensure the depth of the network and control the size of network parameters,which can solve the problem of gradient explosion and gradient disappearance in network back-propagation.Back-propagation uses the mean square error loss function and stochastic gradient descent to update parameters.Simulation experiments show that the recurrent residual U-net reconstruction results have better high-frequency detail performance compared to U-net networks.(3)Study the dynamic magnetic resonance imaging method based on 3D U-shape neural network.In dynamic magnetic resonance imaging,the difference between image frames is mainly caused by the movement of the tissue(such as heart contraction,etc.).The information between frames is redundant,and the total amount of feature extraction of the data can be increased by using the dimension of time.A 3D U-net network was constructed,by extracting the features of adjacent frames using the 3D convolutional layers.Adam(Adaptive Moment Estimation)optimization algorithm was used to improve the convergence speed of the network.The learning rate based on polynomial attenuation method was used to ensure that the network learning can decline steadily in the later training stages.Simulation experiments show that the 3D U-net network reconstruction method can be used for success training and reconstruction at a low sampling rate,and has better reconstruction quality for the motion interval in dynamic magnetic resonance image compare to 2D U-net network.