Research On Parallel MRI Image Reconstruction Method Based On Complex Convolution Double-domain Cascaded U-net Network

Magnetic resonance imaging(MRI)technology has the advantages of no injury to human body,high soft tissue resolution,multi-parameter imaging and so on,and has become the main clinical examination technology.However,its biggest disadvantage is slow imaging speed.Parallel MRI and compressed sensing MRI are two classical methods to accelerate the speed of MRI imaging.Parallel imaging MRI technology uses multiple coils to collect data at the same time to shorten the scanning time,but due to the influence of noise,when the acceleration factor is large,the reconstruction image quality decreases significantly.Compressed sensing MRI technology breaks through the sampling theorem and further reduces the acquisition time,but the image reconstruction is carried out iteratively and the overall imaging is slow.Deep MRI image reconstruction is different from the traditional reconstruction algorithm.It mainly uses network to learn between a large number of undersampled magnetic resonance data and gold standard reconstruction images to obtain accurate mapping relations,so as to achieve fast scanning and obtain high quality reconstruction results quickly.We mainly study the methods of using convolutional neural networks with different structures for parallel MRI image reconstruction,propose a complex convolutional double-domain cascaded U-Net network,design four kinds of complex convolutional network structures,and compare and analyze the performance of the trained network models.Firstly,the principle of magnetic resonance imaging and the advantages and disadvantages of traditional reconstruction algorithms(parallel MRI and compressed sensing MRI)are studied.According to the different types of network input and output magnetic resonance data,four deep learning MRI models are analyzed.Then,the parallel MRI image reconstruction of U-Net network with real convolution is studied.According to the different types of learning MRI data,singledomain U-NET,double-domain cascaded U-NET,deep cascaded U-NET and other network structures were studied,and the normalization method,activation function and data consistency layer used in the network were selected,and the advantages and disadvantages of the three network structures were compared and analyzed.Then,the parallel MRI image reconstruction of complex convolutional U-Net network is studied.Compared with real convolutional networks,complex convolutional networks can better preserve the complex nature of magnetic resonance data.According to the different types of magnetic resonance data input to the network,the learning of spatial complex convolution U-NET network and frequency complex convolution UNET network is realized.The principle of the complex convolution used in the complex convolution U-NET network,the normalization method and activation function used in the network are analyzed.Based on the single complex convolutional U-NET network,we proposed the complex convolutional double-domain cascaded U-NET network,designed four kinds of complex convolutional network structures,and studied the input and output of the four kinds of network structures and the characteristics of network structures respectively.Finally,in the development environment of Spyder,Python3.7 is used to complete the coding and debugging.The data input from the network are preprocessed,including data screening,undercover mask design,acceleration multiple selection,etc.In the framework of Keras and Py Torch,the real convolution and complex convolution U-Net network was built respectively.In network training,network training parameters are set,and ADAM is selected as the optimizer.In the configuration of 2 16 GB GPU workstations,complete the training of different networks.The trained model was used for online reconstruction of MR images,and the results were compared and analyzed.First,the complex convolution U-NET network is obviously better than the real convolution U-NET network,which fully proves that the complex convolution can better learn the characteristics of magnetic resonance data than the real convolution.Second,in the complex convolutional network,the test results of the complex convolutional double-domain cascading U-NET network proposed by us are significantly better than those of the single-domain complex convolutional U-NET network,in which IK CW-NET has the best visual effect and quantitative evaluation.Third,comparing the IK CW-NET model with the deep cascading U-NET model(IKIK-NET)which has a better performance in the current fast magnetic resonance reconstruction research,it is concluded that our IK CW-NET model has a better performance in the reconstruction of details and textures and noise removal.The results show that the complex convolutional U-NET network is better than the real convolutional U-NET network for multi-coil parallel magnetic resonance image reconstruction.IK CW-NET network is a new method in the field of MRI image reconstruction.The application of this method to the development and development of MRI scanners can greatly improve the speed of MRI scanning,quickly obtain high quality reconstruction results,and help to improve the accuracy of clinical diagnosis.