Magnetic Resonance Image Reconstruction Based On Multi-prior Network Learning

Magnetic resonance imaging(MRI)is a noninvasive imaging technique with good spatial resolution and abundant tissue information.It is one of the most important means of medical imaging.One big challenge of MRI is the long imaging time.Existing MRI acceleration methods,including designing fast imaging sequences and developing advanced hardware(parallel imaging),have already reached their respective upper limits utilizing the available techniques,and it is difficult to make further breakthroughs.Compressed sensing(CS)-based MR image reconstruction(CS-MRI)methods have achieved promising performance in fast MRI.However.the fixed sparse transform used in the traditional CS-MRI approaches cannot achieve the expected ideally sparse MR signal,which fundamentally limits the performance.Besides,it is difficult to adjust the parameters manually.The iterative reconstruction process also takes a long time.In addition,CS-MRI can only use the underlying structure information of a single image without considering the multi-source prior knowledge of the accessible big data.In recent years,the rapid development of deep learning technology has provided a new opportunity for fast MRI.Nevertheless,there are few relevant studies for dynamic MRI.These few methods do not make full use of the spatiotemporal redundant information of dynamic MR images and the prior knowledge in different domains.Another issue is that the reconstructed images with high acceleration tend to be over-smooth for existing methods.This project combines the advantages of CS-MRI and multi-source prior network learning to reconstruct dynamic MR images for fast dynamic MRI.The major research contents are as follows:1.A deep residual sparse and cross-domain reconstruction network is proposed for dynamic MRI reconstruction.Inspired by the traditional k-t FOCUSS method,we propose a network framework similar to residual learning in the k-t space to decompose the unknown signal into a prediction signal and a residual signal.Then,the convolutional network module of the cascaded model is utilized to perform the sparse coding of the residual signal.The real signal is restored in the x-f space.In addition,to fully exploit the prior knowledge in different domains,we propose a U-Net-like network to reconstruct the signal in the image space.The proposed method learns the signal recovery by switching the reconstruction process between the x-f space and the image space in an iterative manner.2.To improve the reconstruction performance,we introduce the idea of image super-resolution to MRI reconstruction.Furthermore,inspired by the success of the image pyramid structure and the recursive convolutional neural networks,we propose a recursive multi-scale progressive learning and supervision-based reconstruction method.This method performs multi-scale and multi-resolution supervision and feature learning on the intermediate reconstructed image in the iterative process.In this way,a finer high-resolution MR image in the image domain can be obtained.Experimental results show that the proposed method has a certain improvement in image details.