Research On Magnetic Resonance Image Reconstruction Based On Deep Convolutional Neural Networks

Magnetic Resonance Imaging has been a significant technology in clinical diagnosis.In recent years,algorithms for magnetic resonance image reconstruction have attracted much attention from researchers.A large number of deep learning algorithms have been developed to accelerate MRI and improve the results of MR image reconstruction.This article does the following work:1.A novel multi-scale dilated residual convolution network for MR image reconstruction is proposed.The dilated convolution is employed to expand the receptive field of the convolutional kernel and reduce the parameters of network.Then,the global residual learnings are used to obtain the initial information which are lost in the process of extracting features,and the local residual learnings are used to improve the information flow.In addition,the concatenation layer is proposed to fuse multi-scale features which can accelerate the convergence of the network and improve the accuracy of the network.Additionally,in view of the possible noise in the process of magnetic resonance imaging,simulation experiments of the noisy environment are conducted.The MR images with different levels and types of noise(gaussian noise and rician noise)are input into the convolutional neural network for reconstruction.The proposed algorithm reconstructs better MR images with less noise.3.A novel MR image reconstruction algorithm of convolutional neural networks based on image and gradient domain is proposed to further improve the quality of MR image reconstruction.Firstly,the MR image is subjected to gradient decomposition in the X and Y directions.Then,the image domain and the gradient domain information are entered to deep convolutional neural networks to reconstruct the images respectively.Finally,the three reconstruction results are fused to reconstruct a new MR image.Compared with other deep learning algorithms,the proposed algorithm can improve the results of MR image reconstruction both in subjective evaluation and objective metrics.4.Aiming at low-resolution images resulted from artifacts during magnetic resonance imaging,deep convolutional neural networks are applied to super-resolution reconstruction of MR images.Firstly,the bicubic interpolation method is used to reduce the MR images resolution with different scales,and then the data sets were trained in a neural network to reconstruct and restore high-resolution MR images.