Fast MRI Algorithm Based On Deep Learning

Magnetic Resonance Imaging(MRI)is widely used in the field of medical imaging due to its non-radiation,multi-parameter and high contrast characteristics.But its long data sampling time limits its application.Since the MRI technology was proposed,various methods for improving the imaging speed have been proposed,such as increasing the maximum magnetic field conversion rate and Parallel Imaging.However,one of the current research hotspots is rapid imaging by under-sampling k-space data.Theoretically,under-sampling of k-space can accelerate the imaging speed exponentially,but if the sampling frequency is lower than the Nyquist-Shannon theorem,the reconstructed image will have serious artifacts.The solution to artifacts is to build a reconstruction algorithm that uses prior knowledge to recover the lost data.Up to now,many reconstruction methods based on compressed sensing have been proposed,such as methods based on dictionary learning.This kind of methods can restore the image well,but it is difficult to select the regularization equation and to adjust the hyperparameters.In recent years,due to the popularity of deep learning,more and more researchers have tried to use deep convolutional neural networks to construct reconstruction algorithms.The existing reconstruction algorithms based on deep learning far exceed the accuracy of traditional algorithms based on machine learning.Although deep learning-based methods require a lot of training data and computing resources,more and more public data sets and the development of GPUs and other hardware are gradually following up.Reconstruction algorithms based on deep learning have made significant progress,but there is still much room for improvement in the field of fast MRI.For example,due to the large amount of data in a single sample of MRI,most of the current methods are based on the reconstruction of a single slice.But MRI is usually stereo imaging,and the data is three-dimensional.Using the correlation between slices as a priori knowledge is an important angle to further improve the reconstruction accuracy.In view of this,this article mainly conducts the following research:(1)A hybrid cascaded convolutional neural network(HC-CNN)based on deep learning is proposed.This method establishes an iterative reconstruction model based on a convolutional neural network.HC-CNN mainly uses 3D and 2D hybrid convolution to reconstruct MRI image sequences.In each iteration,3D convolution is used to extract data redundancy between slices of the image sequence,and then 2D convolution is used to further optimize the quality of a single slice.HC-CNN shares weights in the iterative reconstruction process,in addition to reducing the size of the model,it can also make the network parameter training more adequate.Experiments show that HC-CNN can effectively use the data redundancy between slices and significantly improve the accuracy of image reconstruction.(2)A deep iterative convolutional neural network(DICNN)is proposed.HC-CNN uses3 D convolution to deal with data redundancy between adjacent slices.The 3D convolution takes the weight of the plane direction and the slice direction as equal.When the interval between adjacent slices increases,the 3D convolution will learn harmful information from the adjacent slices.In order to solve this problem,DICNN is proposed to improve HC-CNN.The main method is to replace the 3D convolution in HC-CNN with the structure of cyclic convolutional neural network.In order to enhance the flow of information in the network,DICNN also performs a recursive loop in the iterative direction.Experiments show that DICNN not only improves the reconstruction accuracy significantly,but also can effectively cope with different slice pitches.(3)A new fast MRI reconstruction algorithm SR-Net is proposed.SR-Net has made further improvements on the basis of DICNN,that is,constructing an adjacent slice feature extractor based on deformable convolution.Simple 2D convolution will be difficult to deal with complex redundant relationships when fusing adjacent slices,and when the difference between adjacent slices is large,the shortcomings of 2D convolution fusion will be further increased.In response to this problem,DICNN regards the relationship between the current slice and the adjacent slice as a subordinate relationship,and the feature extraction method also only extracts effective information on the adjacent slice.DICNN greatly avoids the problem of extracting harmful information due to the large slice interval.The three reconstruction methods proposed in this paper prove experimentally that the use of data redundancy between adjacent slices can significantly improve the reconstruction accuracy of MRI images.