Study Of Magnetic Resonance Imaging Technology Based On Convolution Neural Network With Skip Connections

Magnetic resonance imaging is widely used in clinical medical diagnosis because of its advantages such as high spatial resolution,non-ionizing radiation and non-Intrusive.However,magnetic resonance imaging requires a long scanning time in practical application,and it is easy to cause motion artifacts during scanning,which limits its application in many applications such as free breathing of the lung and hyperactive children.Although the classical MRI reconstruction algorithm can significantly reduce the acquisition time of MRI data,when the acceleration ratio is too high,the imaging quality will be significantly decreased.In recent years,deep learning has been widely used in semantic segmentation,medical image segmentation and other fields.Up to now,among the fast MRI methods based on deep learning,the method based on convolutional neural networks(CNN)is the most widely used.However,when the U-shaped CNN extracts image features from down-sampling,spatial information of images will be lost.The number of model parameters will be increased significantly with the deepen of the CNN.Therefore,it is of great significance to study how to suppress the loss of image spatial information caused by image down-sampling and how to use compact convolution to reduce the number of model parameters.Since the original MRI data is acquired in k-space as complex data,the image data reconstructed by Fourier transform is also complex data.Up to now,most MRI image reconstruction methods based on CNN are based on real data convolution operation,which does not fully consistent to the characteristics of complex MRI data.Therefore,it is of great significance to study fast MRI reconstruction methods based on complex CNN.In this paper,a fast MRI method based on complex CNN is studied.The method uses multilayer complex skip connections to suppress the loss of image space information caused by image down-sampling,uses complex compact convolution to reduce the number of model parameters,studies the MCUNet++ network of complex compact convolution with skip connections and applied it to multi-channel complex imaging.The specific research contents are as follows:(1)The CUNet++ network based on multi skip connections was applied to fast MRI image reconstruction.The network was mainly based on complex convolution and skip connections.Complex convolution can be consistent to the characteristics of complex MRI data.Skip connections were used to suppress the loss of image spatial information caused by downsampling.The experiment results showed that compared with real data convolution,complex convolution can reconstruct an image with better objective quantitative criteria;The use of multi-layer skip connection module greatly improved the performance of the network,and can fully suppressed the artifacts well with improved objective quantitative criteria values.(2)MCUNet network based on complex compact convolution was applied to fast MRI image reconstruction.The network replaced ordinary convolution with compact convolution module.As the depth of CNN model deepens,the number of model parameters will increase significantly,and compact convolution can reduce the number of model parameters.The experimental results showed that the use of compact convolution will not lead to the decrease of network performance,moreover,some of the objective quantitative criteria values were even improved.(3)The application of MCUNet++ based on complex compact CNN with skip connection in fast magnetic resonance imaging was studied.The network combines compact convolution with multi-layer skip connection.Multi-layer skip connection can effectively suppress the loss of image spatial information caused by U-shaped network down-sampling.Compact convolution were used to control the number of model parameters.The experimental results showed that MCUNet++ suppress image artifacts well caused by data down-sampling and the objective quantitative criteria values were higher than the current MRI image reconstruction method based on CNN.(4)The magnetic resonance imaging method based on multi-channel complex compact CNN was studied,and the PMCUNet++ network was proposed.On the basis of MCUNet++network,and single channel complex convolution operation is extended to multi-channel complex convolution operation.The experimental results showed that compared with PCUNet network,GRAPPA and CS reconstruction methods,PMCUNet++ network shows advantages in both regular and random under sampling modes.