Research On MRI Reconstruction Algorithm Based On Parallel Dual-domain Network

Due to its non-invasive and non-ionizing imaging modality,magnetic resonance imaging(MRI)has been widely used in physicals,biology,medicine,and so on,enjoying broad prospects in the diagnosis of lesions especially.However,there is a long scanning time due to the unique imaging mechanism of magnetic resonance,which is apt to produce motion artifacts and has a negative effect on the quality of reconstruction and the follow-up clinical diagnosis.Therefore,it is of great significance to reduce the time of MRI.With the rapid development of deep learning(DL)in the field of image reconstruction,the DL-based fast MRI image reconstruction has become the mainstream.Based on the existing studies,this paper proposes a fast MRI image reconstruction algorithm based on deep learning technology.The main contributions of this paper are summarized as follows:Firstly,we introduce the research background and significance of magnetic resonance image reconstruction,review the existing researches of magnetic resonance image reconstruction,and analyze the problems in the current reconstruction algorithms and the possible solutions.In addition,this paper also briefly introduces the relevant theory of magnetic resonance imaging technology and the basic concepts of deep learning related to this paper.Then,a MRI image reconstruction algorithm based on parallel dual-domain cascaded convolutional neural network is proposed.Different from the common single-domain data processing network,the algorithm proposed in this paper combines deep learning and compressed sensing to build a dual-stream network to simultaneously process the data in both frequency and spatial domains.After each convolution module of the network,the cross-fusion for the dualdomain data is carried out to fully explore the potential relationship between the dual-domain data.At the same time,in order to solve the problem of feature distortion due to the deepening of the cascade network,the frequency domain data consistency layer and the space domain data consistency layer are added in each stage to ensure that the data fidelity can be effectively implemented in the deeper network.Furthermore,the feature extractor adopts the residual encoder-decoder structure with batch normalization layer to avoid the problems of feature loss and gradient vanishing as the network gets deeper.The fusion module uses the attention mechanism to calculate the self-learning channel weights for the dual-domain data.Combined with the characteristics of complex value in magnetic resonance,the real part and imaginary part of the weighted dual-domain are respectively added and then concatenated to further improve the fusion efficiency and achieve the purpose of fusion of the dual-domain data into a single domain.The loss function includes the MSE loss in both frequency and spatial domains,and also integrates the perceptual loss to avoid the over-smooth of the reconstructed image.Finally,two public datasets(brain and knee)are employed to verify the performance of the proposed algorithm.Several different methods are compared to evaluate the reconstruction performance and robust verification and the results show that the proposed algorithm performs better in both quantitative metrics and visual effects under different sampling trajectories and sampling rates.In summary,this paper employs the parallel dual-stream network as the backbone and introduces the attention mechanism to make use of the potential relationship between the dualdomain data effectively while learning weights independently and improving the data fusion efficiency.In conclusion,A new magnetic resonance image reconstruction network algorithm is proposed and experimental results demonstrate the effectiveness.