Magnetic Resonance Imaging Reconstruction Based On Deep Network Prior

Magnetic resonance imaging is one of the most important imaging methods in clinical diagnosis and medical research.However,this method needs to acquire a large amount of K-space data using spectral imaging,so data acquisition takes a long time,which limits its application scope and efficiency.To address the issue,various reconstruction techniques have been developed to shorten the data acquisition time under the premise of guaranteeing the image quality.As deep learning has made major breakthroughs in the fields of computer vision and image processing,its application in magnetic resonance imaging reconstruction has gradually become a research hotspot.However,at present,several representative deep learning methods generally have the problems that the model is not stable and difficult to train,and the robustness is not enough to be used flexibly,and it requires long-term training of high-quality large data sets.Therefore,it is necessary to find a prior learning method with smaller error representation and stronger generalization ability.This dissertation focuses on mining the underlying representation and construction of deep netdissertation priors and then using them in the research of MRI reconstruction algorithms.The main contributions are as follows:(1)Aiming at the problems that the model is not stable enough,difficult to train,and difficult to use flexibly,a deep energy generation model is used as a carrier to study a self-confrontation prior construction method based on maximum likelihood estimation and an iterative algorithm based on Langevin dynamic implicit inference.The algorithm captures the statistical prior information of complex data distribution and expresses it uniformly with energy,and only needs to optimize the energy difference through the implicit iterative algorithm to achieve excellent image reconstruction performance.Therefore,the uncommon defect that the model needs to be trained multiple times under different samples is effectively avoided.A large number of experimental results show that the algorithm achieves relatively accurate reconstruction performance and the model can be applied to a wider range of scenarios.(2)Aiming at the problem that a large number of fully sampled image samples are required for model training,a high-dimensional correlation prior extraction method combining the idea of the law of large numbers and a multi-channel strategy is studied.Furthermore,a method is proposed to combine fully sampled with undersampled image data in high-dimensional space to train denoising autoencoding netdissertation,which effectively avoids the defect that clean sample data is difficult to obtain in real-world scenarios.At the same time,artificial noise of multiple levels is injected into the high-dimensional space,so that more samples span the entire space to improve the richness.A large number of experimental results show that this method can achieve more accurate reconstruction and better practicability in the absence of sufficient full-sampled image data.