Dynamic Magnetic Resonance Reconstruction In The K-t Domain Based On Generative Model

Magnetic resonance imaging is an important clinical diagnostic technique,which provides great help for medical workers in diagnosing diseases.With the continuous development of magnetic resonance imaging technology,dynamic imaging has become a new research hotspot.People expect to observe various changes and dynamic processes of human organs more accurately through technologies,such as functional magnetic resonance imaging and cardiac magnetic resonance imaging.However,magnetic resonance equipment has the problem of long scanning time,which greatly limits the dynamic imaging speed and reconstruction quality.Currently,there are two methods to improve imaging speed in clinical applications.One is to improve the hardware level of magnetic resonance imaging,but the development of such methods has reached its limit.The other is to achieve accelerated imaging by reconstructing undersampled data while ensuring reconstruction quality.The second method has become a research focus for people to achieve accelerated imaging.In recent years,the deep learning method has gradually extended to the field of medical imaging with its powerful learning ability,and has achieved amazing reconstruction results.However,at present,there are few studies on this kind of methods for dynamic magnetic resonance reconstruction.Therefore,the research focus of this paper is dynamic magnetic resonance imaging based on deep learning.The main research contents are as follows:(1)A universal generation model(UGM)for dynamic magnetic resonance reconstruction in the k-t domain,called k-t UGM,is proposed.This model uses a generative model based on stochastic differential equation(SDE)to reconstruct highly undersampled dynamic images.The model performs weighted operation on the input-space data to reduce the image support and make the training data sparser.Then,the noise distribution trained by SDE network is taken as a priori information.Finally,in the reconstruction process,the predictor-corrector PC in the generated model is used as the numerical SDE solver to convert random noise into sampling data.One of the advantages of this model is that it can achieve multi-frame data reconstruction in dynamic sequences by using single-frame weighted k-space data for network training,which has strong flexibility and versatility.(2)In order to fully exploit the spatiotemporal redundancy of dynamic MR data,a new regularization constraint is constructed by combining the deep generative prior with the traditional low-rank algorithm.The experimental results in different reconstruction scenarios show that the model can improve the reconstruction results of dynamic images qualitatively and quantitatively.In addition,the reconstruction results on the prospective data set(real-time OCMR)further prove the denoising and detail preserving ability of this method in real-time scenes.