Research On MR Image Reconstruction Technology Based On Deep Learning

Magnetic Resonance Imaging(MRI)always suffers the disadvantage of long scan-ning time,which is more likely to bring motion artifacts to MRI images during scanning,and may also aggravate the patient's condition and increase the scanning cost.There-fore,rapid MR reconstruction has always been a research focus of MRI.Compressed Sensing(CS),a powerful mathematical method of reconstructing sparse signals,has proved to have an excellent prospect in the field of rapid MR reconstruction.However,its complicated theoretical knowledge makes it difficult for researchers to get innovative results and the online reconstruction process will also reduce the reconstruction speed to a certain extent.Therefore,a faster and more convenient reconstruction method is still needed urgently in the field of rapid MR reconstruction.Deep Learning(DL)is widely used in various research areas due to its powerful capability of estimating function.As one of the most popular model structures in DL,Convolutional Neural Network(CNN)can extract numerous image features hierarchies easily during image processing and thus has a good reputation between researchers.The DL-based reconstruction method takes the place of the CS-based reconstruction algorithm by converting the time-consuming online reconstruction into offline training of network.In addition,driven by the big data and the strong feature abstraction ability of CNN,the DL-based reconstruction method has better reconstruction performance and can reconstruct a clearer image structure.Based on the analysis,this paper proposes a fast MR reconstruction network based on light-weight CNN.The network comprehensively adopts cascade structure,DC layer and feature shuffle module to reconstruct MR images.Experiments show that the net-work has lower computational complexity and fewer network parameters,and the qual-ity of the reconstructed image is better than those of other reconstruction algorithms.In addition,this paper also verifies the generalization ability of the network by reconstruct-ing the unseen data under different undersampling rates and with pathological structures.The result shows that the network also has a good reconstruction performance on the unseen data.The self-attention mechanism has been proven to have excellent effects in the field of machine translation.Therefore,this paper also focuses on the combination of self-attention mechanism and CNN to reconstruct MR image reconstruction rapidly.This paper proposes a novel neural network model with U-Net structure based on the self-attention mechanism.The network introduces a self-attention mechanism in the residual connection of the third and fourth layers to improve the network's high-dimensional fea-tures abstraction ability.Experiments show that the network has a better reconstruction performance than the network without the self-attention mechanism,and can reconstruct clearer details and more complicated image structures.