Research On Magnetic Resonance Image Reconstruction Based On Convolutional Neural Network

Magnetic resonance imaging has the advantages of no ionizing radiation and high imaging contrast.It has a considerable proportion in various types of medical tests,but its clinical application is relatively limited.Because of its slow imaging speed,and the imaging quality is easily affected by the examiner's physiology.In order to reduce the scanning time and speed up the imaging speed,the acquisition of MR frequency domain data can be reduced.It is an important research content of magnetic resonance image reconstruction to reconstruct high quality MR images by collecting a small amount of data.However,the undersampling method violates the Nyquist sampling law.Direct reconstruction will produce images with artifacts.Traditional MR image reconstruction methods generally use regularization methods for reconstruction.The low rank and sparsity of MR images are used as constraints.Then the nonlinear algorithm is used to solve the optimization equation.The computational complexity is large and the real-time reconstruction needs to be improved.In recent years,convolutional neural networks are superior in natural image processing,but they need to be explored more in MR image processing.Therefore,the main research content of this paper is the reconstruction method of magnetic resonance images based on convolutional neural networks.Convolutional neural networks have many network structures in natural image processing,but compared with the millions of data,the number of magnetic resonance images is scarce,and many networks cannot be directly used for MR image reconstruction.This paper first studies the current magnetic applications.There are several types of convolutional networks for resonance image reconstruction,one is a network widely used in natural image processing,such as U-net,GAN,etc.,and the other is a convolutional network based on the characteristics of magnetic resonance images and traditional methods,such as ADMM-net.This paper mainly conducts magnetic resonance imaging research based on the second type of network.Different from natural image processing,the initial data of magnetic resonance image reconstruction processing is the undersampled image of the target image.The points of this part should remain unchanged after reconstruction.Therefore,in addition to the convolutional network,this paper adds a data consistency.In order to increase the fitting ability of the network,the convolutional network and the data consistency layer are used as a whole to perform several iterations with reference to the traditional iterative algorithm.Experiments show that the method has better reconstruction quality than the general convolutional network.In addition to the imaging characteristics of magnetic resonance images,we can also make use of the outstanding nature of magnetic resonance images-low rank.This paper studies the reasonable representation of this property and uses it as a constrain to improve the performance of the network.Finally,this paper studies the dynamic magnetic resonance reconstruction method based on convolutional network.The dynamic magnetic resonance image has a large correlation in the time domain.This paper proposes a data sharing layer as the pre-processing layer before the reconstruction of convolution network.