Study Of Fast MRI Technology Based On Convolution Neural Network

Magnetic resonance imaging(MRI)has become a routine medical examination method because of its advantages of higher soft tissue resolution and no ionizing radiation damage to the human body.However,because of the shortcomings of slow scanning speed in MRI applications,it is easy to produce motion artifacts and it is difficult to meet the requirements of real-time imaging.Therefore,how to accelerate the imaging speed of MRI is one of the hot spots in the field of MRI.In the past,the imaging time of MRI was usually accelerated from three aspects.One is to improve the performance of MRI hardware,but the physiological effects of the human body limit the magnetic field strength and magnetic field gradient switching rate of MRI hardware.The second is to use parallel imaging technology,but it is affected by the accurate measurement of the coil sensitivity distribution and self-calibration data rows,down-sampling rate,and the number of fitting blocks;the third is to reduce the amount of data collected in MRI k-space(frequency domain space),but the large amount of data acquisition will lead to a significant decline in image quality.Although we can improve the quality of under-sampled image reconstruction through multiple reconstruction algorithms,it often takes a long time to reconstruct and it is difficult to meet the clinical needs of real-time imaging.In recent years,the fast MRI technology based on the deep convolution neural network uses the loss function to calculate and minimize the error of the network by training a large number of full sampling and under-sampling image data.The network parameters are optimized after multiple forward propagation and back propagation.The under-sampling data can be reconstructed based on the optimized network.The deep learning method based on the convolutional neural network(CNN)has become a research hotspot in the field of MRI fast imaging because it has the advantages of off-line training and on-line imaging compared with the traditional methods,such as compressed sensing,which greatly reduces the time of MRI.This article mainly does the following:(1)The convolutional neural network theory is studied and a U-net convolutional neural network is constructed.In the network back-propagation,MSE is chosen as the loss function in the back propagation of u-net convolutional neural network,and the Adam algorithm is selected as the optimization algorithm,which avoids the problem of gradient disappearance in training and speeds up the convergence speed.Training network based on full sampling data and two kinds of under-sampling MRI brain data(random variable density under-sampling and uniform regular under-sampling),and then reconstruct the network after training the magnetic resonance brain under-sampling images.The simulation results show that high-quality under-sampling images can be obtained by U-net convolutional neural network reconstruction.The uniform and regular under-sampling image reconstruction quality is close to the random variable density under-sampling reconstruction quality.(2)The fast MRI algorithm based on ar2 u net convolutional neural network is studied.Based on the U-net convolutional neural network,a new AR2 U-net convolutional neural network is constructed after adding a recursive residual module and an attention gate module.Recursion can deal with information more effectively,control the size of network parameters,and avoid the problem of gradient disappearance during training.Residuals can solve the problem of network error caused by abnormal data in the network and reduce the difficulty of network training.Attention mechanism enhances the ability of network feature extraction.The AR2 U-net convolutional neural network was trained with 800 magnetic resonance brain image data,of which 720 were used for training,40 were used for verification,and 40 were used for test reconstruction.The experimental results show that the image reconstruction quality of AR2 U-net convolutional neural network is higher than that of U-net convolutional neural network.Under the same sampling method,the training time of U-net convolutional neural network is about 8 hours.The reconstruction time of each image is about 1 second,the training time of the AR2 U-net convolutional neural network is about 10 hours,and the reconstruction time is about 1.4 seconds.After the code optimization and hardware performance improvement,it can meet the needs of real-time online imaging.Experimental simulations show that the image reconstruction quality of AR2 U-net convolutional neural network is higher than that of U-net convolutional neural network.(3)Research on fast magnetic resonance imaging algorithm based on hybrid spatial convolutional neural network.A hybrid spatial convolutional neural network composed of a combination of k-net and i-net,which was constructed and applied to fast magnetic resonance brain imaging.Among them,k-net and i-net are based on convolutional neural network,k-net is for image k-space training,i-net is for image space training,k-net and i-net combine to form hybrid spatial convolutional neural network,k-net and i-net are alternately trained in sequence,iterating twice,and the final optimized parameters are used for image reconstruction of MRI under-sampled data.Experiments show that at 29% under-sampling rate,the reconstruction quality of the method based on the hybrid spatial convolutional neural network is significantly higher than that of the U-net convolution neural network,and slightly higher than the AR2 U-net convolutional neural network.The training time of the hybrid spatial convolutional neural network is about 20 hours,the reconstruction time of each image is about 1.7 seconds.After the code optimization hardware is upgraded,it can meet the needs of real-time online imaging.