Model-Driven Fast Magnetic Resonance Imaging

Magnetic resonance imaging(MRI)technology,which can non-invasively provide the soft tissue structure and metabolic status of the scanned object,has been widely used in clinical diagnosis and scientific research.However,the long scanning time greatly limits its further application and development.Existing applications put forward more strict requirements for the acquisition speed and imaging resolution of MRI scanning How to perform high-quality and fast imaging is a major challenge for MRI.At present,reducing the number of samples in k-space is the main means to speed up magnetic resonance imaging.For dynamic magnetic resonance imaging and high-resolution static imaging,a high acceleration rate is very important,especially.In the former,the motion characteristics of the scanned object determine that only a limited number of samples can be obtained within the specified acquisition window,and images need to be reconstructed from these highly under-sampled data;in the latter,the imaging time must be controlled within an acceptable range,where a higher acceleration factor is also necessary.This article focuses on two application scenarios of fast imaging:cardiac cine imaging,which requires extremely high acquisition speed,and whole-brain vessel wall imaging,which requires extremely high spatial resolution.Two model-based magnetic resonance reconstruction algorithms were proposed,respectively.First,the author introduced the basic principles and fast imaging algorithms of magnetic resonance imaging.The author concisely explained the physical principles of magnetic resonance imaging,the correspondence between k-space and image domain,and introduced the technical details of the two acceleration methods:parallel imaging and compressed sensing.Then the new type of model-based deep learning reconstruc-tion algorithm was introduced.This part provides the theoretical background for the following research in this paper.Second,The author proposed a model-driven deep low-rank+sparse reconstruc-tion network(L+S-Net)in response to the high acceleration requirement of dynamic cardiac cine imaging.This method modelled the dynamic magnetic resonance image into a low-rank+sparse form and obtained an iterative solution through an optimiza-tion algorithm.Afterwards,it was unrolled into a deep neural network,which greatly reduced the imaging error and increased the acceleration factor.Third,the author further explored the combination of image reconstruction with image super-resolution,and proposed a model-driven deep super-resolution-reconstruction network(SRR-Net)for high-resolution whole-brain MRI.This method first established a magnetic resonance super-resolution-reconstruction model,and then an optimization algorithm was used to get the iterative solution of the inverse problem.The iterative steps were expanded into a super-resolution-reconstruction deep neural network after-wards.Then the idea of adversarial learning was introduced and a discriminator was used to improve the detail recovery ability of the network.This article focuses on improving the speed and quality of magnetic resonance imaging,and introduces the ideas of learned low-rank prior and super-resolution-involved reconstruction into dynamic magnetic resonance imaging and high-resolution magnetic resonance imaging,respectively,and achieved promising reconstruction results.