Research On Rapid Magnetic Resonance Imaging Based On Deep Learning

Magnetic Resonance Imaging(MRI)technology has been widely used in clinical application because of its ability of non-invasive,non-radiative and arbitrary tomographic imaging.However,its slow data acquisition time limits its further development in dynamic imaging application,etc..MRI data acquisition time can be shortened in three methods.The first one is to improve the performance of MRI hardware,enhance the main magnetic field strength and switching speed of gradient for the magnetic resonance scanner,but due to the physiological effects of the human body,it is impossible to increase the magnetic field strength and accelerate the switching rate of the magnetic field gradient without any limitations;The second one is to use parallel imaging technology,but due to the limited coil sensitivity,the reconstruction quality is significantly reduced if the acceleration acquisition factor is large;The third one is to reduce amount of the k-space data,and improve the image reconstruction quality of under-sampled data by studying various reconstruction algorithms.This paper mainly focused on improving the acquisition speed of magnetic resonance data by reducing the amount of data collected in k-space.Compressed Sensing(CS),Singularity Function Analysis(SFA)and convolution-based deep learning methods are used to reconstruct under-sampled k-space data with high image quality and short reconstruction time.Both compressed sensing and singular spectrum analysis techniques require that the image data to be reconstructed is sparse,however,the image reconstruction based on deep learning technology does not require the sparseness of image data,which makes it have great application potential in the field of MRI image reconstruction.This article mainly does the following work:Study the theory of compressed sensing and apply it in fast magnetic resonance imaging.The CS reconstruction algorithm based on wavelet sparse transform and with TV regularization terms are studied,under the constraint of data consistent fidelity,the under-sampled magnetic resonance data is reconstructed by obtaining the minimum value of the L1 norm in the sparse transformed domain.The simulation results show that,with different under-sampling ratio,the quality of CS reconstructed images is better compared with that reconstructed with directly zero-padding method.The quality of CS reconstructed images is further improved after adding TV regularization term.Study the singular spectral function analysis model and its reconstruction method.The reconstruction method based on singular spectral function analysis is applied to the simultaneous rapid imaging of magnetic resonance angiography and venography.The singular points can be extracted from the difference image.The singular spectral function is constructed using the singular points.The un-acquired k-space data can be recovered using the weighing of singular spectral function.Simulation experiments show that the singular spectrum analysis reconstruction can improve the image quality of simultaneous angiography and venography imaging.However,this method has the disadvantages of long reconstruction time and large memory consumption.Study the rapid magnetic resonance imaging based on deep convolutional neural networks.The U-net convolutional neural network is constructed and improved.A fast magnetic resonance imaging method based on complex U-net convolutional neural network is proposed and applied to the rapid imaging of magnetic resonance for the brain.Constructing a complex U-net convolutional neural network,a Root Mean Square Propagation algorithm is used to optimize the network parameters with a large number of magnetic resonance brain in complex format.Using the optimized parameters,U-net convolutional neural network can be used to reconstruction of undersampled k-space data.Simulation experiments show that the fast magnetic resonance imaging method based on complex U-net convolutional neural network can effectively improve the image reconstruction quality and significantly shorten the image reconstruction time.At each undersampling ratios,especially low ratios,compared with CS reconstruction and SFA reconstruction,the reconstruction quality of U-net convolutional neural network method is best,and its short reconstruction time can satisfy the requirements of real-time online imaging.Study the rapid magnetic resonance imaging based on deep residual U-net Convolutional Neural Network.Aiming at the problems of gradient disappearance,explosion and over-fitting which often occur in U-net convolution neural network,residual module is added to U-net convolution neural network.The simulation results show that The fast magnetic resonance imaging method based on residual U-net convolution neural network can prevent the gradient disappearance of conventional U-net convolution neural network and further improve the imaging quality.