A Study Of Parallel Magnetic Resonance Imaging Reconstruction Based On Multiple Weights

Magnetic resonance imaging has the advantages of non-invasive,multi-tissue contrast and produces very clear,multi-dimensional,high-resolution images,which makes it one of the indispensable techniques in medical diagnosis and research.However,compared with other medical imaging techniques such as X-ray and CT,MRI usually requires longer scanning time,which limits the popular application of MRI.In order to reduce MRI scan time while maintaining imaging quality,many effective solutions that can be used to accelerate MRI acquisition have been explored.Parallel imaging is widely used in magnetic resonance imaging as an accelerated technique.However,conventional linear reconstruction methods in parallel imaging usually suffer from noise amplification.Recently,Scan-specific Robust Artificialneural-networks for K-space Interpolation(RAKI)have shown better noise immunity than other linear methods.However,RAKI performs poorly at high acceleration and requires a large number of Auto Calibration Signal(ACS)as training samples.To address these problems,this study proposes a multi-weight matrix-based method that can be used for MRI reconstruction under high acceleration and effectively reduces the number of ACS lines.The main work of this paper is summarized as follows.(1)To address the problems of poor performance of RAKI under high acceleration and the need for a large number of self-calibrated signals as training samples.Based on the principle of multi-weight matrix,a parallel reconstruction algorithm of magnetic resonance based on matrix weighting technique is proposed.Firstly,the designed weight matrix can map the K-space data and scale the magnitude of K-space data to a certain range,which is conducive to network learning and makes network training easier to converge;secondly,the algorithm uses a high-dimensional strategy and the idea of multi-weight matrix to generate different feature images through multi-weight matrix to make the samples more diverse and rich,which can cover the data space more comprehensively and better cope with noise and anomalies.After a lot of experiments,it is proved that the algorithm can achieve more accurate reconstruction under high acceleration rate.(2)To address the problem that many current deep learning-based methods require a large amount of fully sampled training data,yet it is difficult to obtain a large amount of medical data due to privacy and variability.Inspired by the robust artificial neural network with K-space interpolation(RAKI),this study proposes a small network model that requires only the scan data of this patient and no other additional training data.Thus there is less dependence on the training data.The MWRAKI is formed by applying weighting to the undersampled data using multiple weighting matrices and then feeding the processed data into the proposed small network model.the effect of noise can be effectively reduced and data constraints can be increased by processing the measurement data with multiple weighting matrices.In addition,we fuse the strategies of multiple weighting matrices into rRAKI(residual RAKI)and form MW-rRAKI.The proposed method was experimentally compared with GRAPPA,RAKI and rRAKI under different acceleration factors and ACS conditions,and the proposed method has good reconstruction performance and can reconstruct high-quality images.In particular,the PSNR of MW-RAKI and MW-rRAKI was improved by about 3 d B and 4 d B compared with RAKI and rRAKI,respectively,when only 12.5% of the Kspace data were available at high acceleration rates.