Research Of MR Parallel Imaging Based On Deep Learning

Magnetic resonance imaging,as a non-damage,multi-modal clinical medical disease detection technique,can provide high-resolution,arbitrary dimension soft tissue structure images for clinical diagnosis.At the same time,functional magnetic resonance imaging,as an important tool in brain science research,has also been rapidly developed.Because of the influence of its own imaging mechanism,the imaging time of magnetic resonance imaging system is relatively long,and sometimes it can not meet the special needs of clinical medicine.In order to improve the imaging speed of magnetic resonance imaging system and further meet the clinical needs,parallel magnetic resonance imaging has been proposed and deeply studied.Parallel magnetic resonance imaging first reduces the acquisition of K space data in the direction of phase coding by parallel coils acquisition technology,and then obtains high quality magnetic resonance images by parallel magnetic resonance reconstruction algorithm.Parallel magnetic resonance reconstruction algorithms include GRAPPA and SENSE.Through studying the SENSE algorithm,this paper proposes a hybrid magnetic resonance parallel reconstruction algorithm using deep learning to improve SENSE algorithm,which can further improve the reconstruction effect of traditional SENSE.In addition,this paper proposes a multi-channel magnetic resonance parallel reconstruction algorithm based on deep learning by studying the general framework of traditional parallel reconstruction algorithm.The main research findings are as follows:1.This paper presents a novel hybrid magnetic resonance parallel reconstruction algorithm to improve the SENSE reconstruction effect by deep learning algorithm.SENSE reconstruction framework mainly utilizes the sensitivity distribution of parallel coils,i.e.,coil sensitivity maps,to remove aliasing artifacts due to under-sampling of K space.Obtaining accurate coil sensitive maps is crucial for SENSE reconstruction.Consequently,we mainly use the powerful nonlinear mapping ability of deep learning in image generation to synthesize high quality and accurate coil sensitive maps,thus further improving the reconstruction effectiveness of traditional SENSE algorithm.Another advantage of this method is that it provides an idea to improve the generalization ability of fast imaging methods directly based on deep learning.In this paper,the above two advantages of this method are proved by detailed comparative experiments.2.In this paper,a multiple-channel parallel magnetic resonance reconstruction algorithm based on deep learning is proposed.For improving the imaging quality and speed of magnetic resonance equipment,multiple-coil acquisition technology has been developed rapidly in MRI field.Multiple-coil acquisition technology is to use array phased array coils to collect echo signal simultaneously.Then,multiple-channel magnetic resonance images are obtained from echo signal by Fourier transform,and the difference between each channel image reflects the relative spatial position information of each coil.At present,the inputs of the most reconstruction algorithms based on end-to-end deep learning are single-channel images.Single-channel images obtained by square sum algorithm(Sum of Squares,SOS).The fast reconstruction algorithm based on deep learning with single channel image as input completely ignores spatial information and weight information.In order to solve the above problems,this paper proposes a new parallel reconstruction algorithm based on end-to-end deep learning with multiple-channel magnetic resonance images as input,which makes full use of the relative spatial position information and weight information between each channel image.After comparative experimental analysis,this method has better experimental results.