Research On Magnetic Resonance Spectroscopy Reconstruction Algorithms Based On Deep Learning

Nuclear magnetic resonance(NMR)has been widely used as an effective analysis technique in the structural identification of chemical substances.Normally,the data in the indirect dimensions of NMR spectra is rather time-consuming to record,and the acquisition time increases exponentially with the number of dimensions.Non-uniform sampling technique can shorcten the experimental time by reducing the number of sampling data in the indirect dimension of NMR spectra.If a non-uniformly sampled(undersampled)signal is processed directly by zero-filling,a large number of overlapping artifacts will be formed in the spectrum.Therefore,data processing algorithms should be used to reconstruct undersampled NMR spectra.Traditional NMR spectra reconstruction methods use the prior knowledge of NMR spectra(such as sparsity,low rank,etc.)to regularize the established optimization equation,and then combine optimization algorithms to obtain approximate solutions.Although these methods can provide good reconstructions,these methods have the disadvantages of large computational complexity and long reconstruction time.Deep learning has strong nonlinear fitting capability,and because of the support of hardware,deep learning has fast calculation speed.In view of this,this paper proposes two deep neural networks for NMR spectra reconstruction,including an encoder-decoder high-resolution neural network that uses virtual echo to reduce training time and a model-driven iterative self-adaptive thresholding network.In this paper,a deep Encoder-Decoder High-Resolution Networks(EDHRN)is proposed.A block called Encoder-Decoder(ED)is introduced into High-Resolution Networks(HRNet)and it has been shown that ED structure is effective for removing artifacts.Then,this paper innovatively brings the virtual echo algorithm to data preprocessing,which can change the imaginary part of NMR spectral data to zero.Thus,the number of channels of the output layer in EDHRN only needs to be set to one to represent the real part of NMR spectra data.Experimental results show that this method not only has faster reconstruction speed than traditional NMR spectra reconstruction algorithms such as hmsIST and SMILE,but also has higher reconstruction quality than classical convolutional neural networks such as DenseNet and U-Net.At the same time,compared with the existing deep learning-based NMR spectra reconstruction methods,the training time of our method is also shorter.Many deep neural networks lack theoretical support and interpretability,while traditional NMR spectra reconstruction algorithms have a solid theoretical foundation.Therefore,this paper draws on the theory of the Iterative Shrinkage Thresholding Algorithm(ISTA),and designs a novel iterative self-adaptive thresholding network(ISATN)driven by a mathematical model.This method uses the mathematical operation layer and the convolutional layer as a module to represent a step of the calculation process in ISTA.Multiple identical modules constitute the overall structure of the neural network.Experiments show that this method can obtain high-quality NMR reconstructed spectra with fewer model parameters and shorter reconstruction time.