Accelerated MRI Research Based On Dictionary Learning And Deep Learning

Magnetic resonance imaging(MRI)is a very important medical imaging method,but its long time data acquisition process restricts its clinical application.In order to solve the problem that the time of data acquisition is too long in MRI,a method of undersampling K spatial data is proposed,which has speeded up the acquisition process of MRI in theory,but undersampling K spatial data violates the Nyquist-Shannon(Nyquist Shannon)theorem.This will lead to serious artifacts and blurred image details caused by the reconstructed images based on the undersampled K spatial data.The difficulty of solving this problem is to find an algorithm that uses prior knowledge to make up for missing data and reconstruct undersampled images.At present,a lot of MRI acceleration methods based on the principle of compressed sensing and parallel imaging have been proposed.Most of these methods can well reconstruct magnetic resonance.But the core algorithm of these methods generally adopts the traditional machine learning method,and its algorithm effect still has much room for improvement.Since the birth of AlxNet,deep learning has been paid much attention by many researchers and made outstanding contributions in many fields.In the field of MRI acceleration,some people also propose a solution based on deep learning.Most of these methods need to train a large cascade convolution neural network.Because the training of such networks often requires a large number of high quality data sets,and the trained networks often tend to be over fitting,so performance of these method is not good enough.The method of accelerated magnetic resonance imaging proposed in this paper is an optimization method of the learning joint-sparse codes for calibration-free parallel MR imaging.Firstly,in the paper we compare LINDBERG with some other methods in a relatively fair environment.I have participated in the research and experiment of LINDBERG,and the identity of the third author.The publication of the paper has been completed.By comparing the visual effect,quantitative index and error map of the reconstructed magnetic resonance image,we find that LINDBERG surpasses the other methods in many aspects.Then we use the deep learning method to further optimize the magnetic resonance image reconstructed by LINDBERG.Because of the existence of the initial reconstruction of magnetic resonance images,the problem of the neural network to be treated is simplified,so this paper only designs a shallow hierarchical convolution neural network,and the network has no full connection layer.The parameters of the network are reduced,so the neural network proposed in this paper is easy to train and reduces the requirement for training data sets.Finally,we find that the proposed method can further optimize the LINDBERG by comparing the magnetic resonance images reconstructed by LINDBERG and the proposed method.Moreover,the new idea that combines traditional machine learning with deep learning proposed in this paper will complement each other and enlighten other fields.