Study Of Fast Magnetic Resonance Imaging Technology Based On Complex Convolutional Neural Network

Magnetic resonance imaging(MRI)is a very important technology in clinical medical imaging diagnosis,but it is limited in actual clinical application due to the long scanning time when collecting data.In order to accelerate the scanning speed,a method of under-sampling the k-space data can be used to reduce the amount of data collection.However,the reconstructed image obtained by directly performing Fourier transform on the k-space under-sampled data will produce serious artifacts.Therefore,some researchers have proposed MRI reconstruction methods based on algorithms such as compressed sensing and parallel imaging.However,the compressed sensing algorithm requires that the under-sampling mode of the magnetic resonance data is random undersampling and the under-sampling image artifacts are non-correlated,and the optimization algorithm during reconstruction requires multiple iterative calculations,which takes a long time.Parallel imaging algorithms need to rely on the sensitivity characteristics of each coil,and have very strict requirements on the structure of the coils.With the increase of the under-sampling rate,the data used for fitting is reduced,the weight coefficients obtained cannot reconstruct images with better quality.In recent years,researchers have proposed MRI method based on convolutional neural networks(CNN).The initial study of MRI using convolutional neural networks is mainly based on real-valued networks.However,the MRI method based on realvalued convolutional neural network has the disadvantage that it can only reconstruct the amplitude image,and all the phase information is lost.Generally,the original kspace magnetic resonance data collected is complex-valued data.Therefore,in order to retain the phase of MRI,two complex-valued convolution neural network methods are proposed to reconstruct complex-valued MRI data: The first is to use the real and imaginary parts of the complex magnetic resonance data as two-channel real-value to train the convolutional neural network;The second is to use the amplitude and phase information of the complex-valued magnetic resonance(MR)data as two-channel realvalue to train the convolutional neural network,or the amplitude and phase as two sets of data to train the corresponding convolutional neural network;However,the above two complex-valued magnetic resonance image reconstruction techniques will destroy the correlation between the amplitude and the phase,and the quality of the reconstructed phase image will be impaired.In clinical medical imaging diagnosis,not only the amplitude image can help doctors diagnose the patient’s condition,but also the phase can obtain a lot of important information,including blood flow velocity,blood flow,temperature,quantitative magnetization map,fat-water separation,chemical shift imaging and brain segmentation,so it is very meaningful to study complex-valued magnetic resonance image reconstruction.For the above reasons,two complex-valued convolutional neural networks(CR2UNet and CAR2UNet)are proposed in the paper to reconstruct complex-valued MRI.Starting from the mathematical theory of complexvalued convolutional neural networks,a reasonable complex-valued network is built to reconstruct complex-valued under-sampled magnetic resonance images.In this paper,we mainly study the fast magnetic resonance imaging method based on real-valued convolutional neural network(ADNet),complex-valued number convolutional neural network(CR2UNetand CAR2UNet).The specific research content is as follows:(1)Study of fast MRI method based on real ADNet convolutional neural network.For the first time,ADNet convolutional neural network is applied to fast magnetic resonance imaging.ADNet is a convolutional neural network that uses attention mechanism to denoise.The structure of this network is simpler than that of real-valued U-Net convolution neural network.According to the analysis of the reconstructed image,the total relative error function and the structural similarity error function,the ADNet convolutional neural network has a faster reconstruction speed than the realvalued U-Net network,and the reconstruction quality is slightly higher.(2)The reconstruction method of complex MRI based on CR2 UNet convolutional neural network is studied.First proposed CR2 UNet convolutional neural network and applied it to complex-valued MRI.The CR2 UNet network introduces a complex-valued recurrent residual module based on the CUNet convolutional neural network.This module mainly solves the situation that the gradient explosion or disappearance and the number of network model parameters can be controlled while the network depth is increased.The results show that the CR2 UNet convolutional neural network has a larger amplitude SSIM value and a smaller TRE value than the CUNet,indicating that the CR2 UNet has a better reconstruction effect.(3)The complex-valued MRI image reconstruction method based on CAR2 UNet convolutional neural network is studied.CAR2 UNet convolutional neural network is proposed for the first time and applied to complex-valued magnetic resonance imaging.The most interesting part of magnetic resonance images in clinical medical diagnosis is often the subject’s diseased part or abnormal part,the attention mechanism module focuses on irrelevant areas in the image during image reconstruction(irrelevant areas refer to possible lesions or abnormal areas).Based on the above reasons,we propose CAR2 UNet,which introduces the complex-valued attention module increases the sensitivity and reconstruction accuracy of the model.The experimental results show that the CAR2 UNet has a larger SSIM value and a smaller TRE value than the CR2 UNet convolutional neural network,indicating that the CAR2 UNet convolutional neural network has a better reconstruction effect.