Highly Undersampled Magnetic Resonance Imaging Reconstruction Using Autoencoding Priors

Image reconstruction from highly under-sampled k-space data is a classical problem in Magnetic Resonance Imaging(MRI),which is of great necessity for reducing acquisition time.However,as the rate of k-space sampling rate declines,we may face different challenges such as the noise amplification,blurred object edges and aliasing artifacts due to underdetermined issues.Very recently,deep learning has been adopted to assist fast imaging and the general image restoration problems,which could be roughly categorized into two types based on the consistency between the training data and testing data.The first category is supervised learning approaches,which usually need input-output pair in training stage and start from the same type of input data in testing stage.In another category of unsupervised learning approaches for MR imaging,they aim to learn the probability distribution of the fully sampled images by network training.After that,the network-learned image priors are applied to the constrained image reconstruction framework as an explicit constraint.we exploit the recent concept of network-learned priors as a regularization term for MRI reconstruction.The Denoising autoencoder(DAE)is utilized as effective priors in our iterative reconstruction procedure,due to its flexible representation extension and excellent robustness abilities in image restoration.The main contributions of this work are as follows:(1)To our best knowledge,this is the first work to introduce the DAE prior for MRI reconstruction.Unlike the recent deep CNN-based methods employing an end-to-end learning fashion,we used the network learning as a tool to learn general prior information and incorporate it into the constrained reconstruction framework.Therefore,once the network-learned image prior is obtained,it can be applied to reconstruction tasks with different sampling trajectories and acceleration factors and can guarantee promising results.(2)More importantly,two advanced strategies are proposed to enhance the naive DAE prior,termed EDAEP.First,considering that the high-dimension manifold learning may favor more accurate representation of image prior,we learn the prior in higher-dimensional scenario for network training by means of variables augmentation technique,and employ it for original single-channel image reconstruction task.Second,recognizing that noise distribution is the most important parameter which affects the DAE prior,and different implementation of noise levels may favor different image features,the two-sigma rule and average techniques are jointly employed to improve the prior robustness.Experimental results under varying sampling trajectories and acceleration factors consistently demonstrate the superiority of the enhanced autoencoding priors,in terms of Peak Signal-to-Noise Ratio(PSNR),Structural Similarity(SSIM)and High-Frequency Error Norm(HFEN).This technique can effectively improve the reconstruction quality and has superior performance in terms of detail preservation compared with state-of-the-art methods.