| Magnetic Resonance Imaging(MRI)has the advantages of high resolution of soft tissue imaging,no ionizing radiation,and arbitrary orientation imaging.It is widely used in clinical disease examinations.However,the slow speed of MRI imaging limits its further application.Parallel and compressed sensing MRI imaging technologies use multi-coil acquisition and compressed sensing theory to accelerate magnetic resonance imaging.However,these two technologies,due to the limitations of hardware and their own theory,have poor imaging quality under high-factor acceleration.At present,deep learning is also beginning to be used in rapid MRI reconstruction.Compared with traditional reconstructio n methods,it can accurately learn the mapping relationship from under-sampled MRI scan data to gold standard images,and quickly obtain higher-quality reconstructed images.However,deep learning MRI reconstruction methods mostly use real convolutional ne ural networks,and the processing of complex MRI data is not reasonable enough.In response to this problem,we propose to combine the dual-domain cascade network with complex convolution,design a complex-convolution dual-domain cascade network,and use it for single-coil compressed sensing MRI reconstruction.The main work includes:First,the principle of magnetic resonance and several traditional imaging methods are studied.The principle of MRI imaging is analyzed from the aspects of NMR phenomenon,spatial encoding positioning mechanism,k-space,etc.In-depth study of the principles of MRI imaging technologies such as full sampling sequence,parallel,compressed sensing,etc.Then,the existing deep learning MRI reconstruction methods are studied.The deep learning MRI reconstruction methods based on image domain,k-space domain,and dual-domain cascade network are analyzed and compared,and the principles and functions of network structures such as data consistency layer and dual-domain cascade are explored.Next,the combination of dual-domain cascade network and complex convolution is studied.By improving the complex convolution,the number of convolution kernels,and the loss function of the hybrid dual-domain cascaded network HRCNN,a complex convolution dual-domain cascaded network HCCNN is built.Compared with the original HRCNN network,the HCCNN network uses a complex convolutional neural network,the number of convolution kernels in all convolutional layers is halved,and the average absolute value error loss function MAE is used to guide the network to conduct back propagation.Finally,on the Anaconda3 platform,the Python language is used to write and debug the experimental code.Data screening was performed on the Calgary single-coil full-acquisition data set,and the under-sampling mask was used to simulate the accelerated scanning of the data,and the original data set was obtained.It is divided into training set,validation set,and test set according to 5:1:1.Under the pytorch framework,various network models are built,network hyperparameters are set,network structure is optimized,the network is trained and tested,and a highly accurate network model is obtained.A comparative analysis of the reconstruction results of various networks is carrie d out.First,Compared with the HRCNN network,the HCCNN network we proposed can still obtain better reconstruction quality when the number of parameters is halved.Second,compared with other deep learning MRI reconstruction networks,the HCCNN reconstruction network also obtains the best reconstruction quality,and the amo unt of network parameters is the smallest.Third,the HCCNN reconstruction network also has a very high reconstruction speed,which can reach 88 frames per second.Four th,compared with the mean square error loss function MSE,the average absolute val ue error loss function MAE is used as the loss function of the complex convolution d ual-domain cascade network,which can obtain a better reconstructed image and recover better brain details.The double-domain cascaded network HCCNN proposed by us can effectively extract the double-domain and complex features of MRI data,improve the reconstruction quality and reduce the amount of model parameters.And the reconstruction speed is fast,which can meet the clinical requirements for MRI real-time imaging.This research work has strong engineering reference significance,and can provide ideas for the integration of deep learning MRI reconstruction methods on MRI scanners,truly realize the integration of fast scan and fast reconstruction MRI imaging,and broaden the clinical application of MRI range. |
