Research On Fast MRI Reconstruction Method Based On Deep Learning

Magnetic resonance imaging(MRI)has been widely used in clinical medicine because of its non-invasive and non-radiation advantages,especially the ability to directly image soft tissue,blood and other regions.However,because MRI technology requires a long scanning time,in order to shorten the sampling time,only part of the data is usually collected during the actual acquisition.Aliasing effects occur when the amount of data collected is not sufficient for Nyquist’s sampling law.The reconstruction task of MRI is to eliminate these artifacts,which has great value for medical application diagnosis.At present,the reconstruction technology has made great progress,but it is difficult to reconstruct the complete MR image details in the case of low sampling rate.In order to solve these problems,this paper explores from two perspectives of model framework improvement and how to combine traditional methods with deep learning methods.First,we propose a single-coil MRI reconstruction framework guided by high-frequency residual information,which uses existing method reconstruction results and residuals from real images to reinforce the model’s attention to these hard-to-learn information.And according to the characteristics of MRI raw data,we design a residual channel spatial attention module,in which the complex convolution operation and Fourier convolution operation are used to fully extract the inherent information in the MR data.The complex aliasing artifacts in the image are well removed under the combined action of the residual channel spatial attention module and the data consistency module,and the detailed texture of the reconstruction result is preserved.Second,we propose a fast reconstruction algorithm for multi-coil MRI based on gradient-fidelity priors.Although the existing model-driven methods can reconstruct relatively clear images after training on less datasets,they cannot preserve the details of the images well under the condition of low sampling rate.We introduce finer information by increasing the gradient prior fidelity term,which improves the model’s ability to reconstruct texture details.Moreover,we verify the effectiveness of the gradient prior fidelity term through extensive ablation experiments.Finally,we design a fast MRI reconstruction system based on deep learning.The system greatly improves the work efficiency of doctors and researchers,and can automatically reconstruct undersampled MR data,display,and the model can be fine-tuned based on the doctor’s feedback on the reconstruction results to improve the model’s ability to capture details.