Research On Fast MRI Reconstruction Based On Model-driven

Magnetic resonance imaging is currently one of the most important imaging tech-niques in medicine.It not only can accurately obtain the physiological function,anatomi-cal structure,metabolic information and lesion information of tissues and organs,but also has the characteristics of non-invasive and non-ionizing radiation,which is a safe and reliable treatment technology.However,the full sampling time of MRI images is long,and the excessively long scanning time will bring discomfort to patients and may even delay the treatment of their conditions.Therefore,in clinical applications,a shorter scan time is needed to improve the patient experience.Currently,reducing sampled data in K-space is an important means to speed up MRI.To overcome the problem of image quality degradation caused by undersampling,a number of existing methods have been proposed,which can also effectively reconstruct images,but still suffer from large algorithm model parameters and poor ability to portray details.To address the problems of fast magnetic resonance imaging techniques,the following work are carried out in this paper.1.A fast MRI reconstruction method based on iterative optimization expansion of Framelet framework(Framelet-net)is proposed.The method divides the image into low-frequency and high-frequency information by means of Framelet framework and constrains them separately.After solving the problem,we expand the itera-tive process of the obtained analytic formula into a neural network and implement the network model.This method takes advantage of both the Framelet framework,which can help capture more edge detail information,and also reduces imaging er-rors by model-driven iterative unfolding.2.A cascaded network model that learns both regularization constraints and data con-sistency constraints(CRDC-net)is proposed.The model applies the Bregman up-date equation to update the undersampled images and obtain more complementary information.Meanwhile,for the two subproblems in the reconstruction equation, we input the combination of the variables to be updated into the neural network to obtain the reconstruction results in a learnable way.The model breaks the fixed structure in the iterative equation by relaxing the constraint of the data,and makes more full use of the learning ability of the neural network to obtain a better recon-struction quality.3.In order to further improve the reconstruction effect of CRDC-net,in the process of network design,we designed a frequency attention mechanism in the frequency domain using the characteristics of K-space data,so that the model can focus more on high-frequency information.In addition,we add a high-frequency refinement branch to further refine the reconstructed images in the final stage of the model and supervise this branch to obtain more accurate reconstruction results.To verify the validity of our model,we perform qualitative and quantitative com-parisons with other classical MRI reconstruction methods on several datasets and abla-tion analysis with the constituent parts and relevant parameters of our own method.It is demonstrated that our proposed model outperforms other methods and obtains better reconstruction results for different datasets and at different sampling rates.