Fast reconstruction of non-uniform sampling multidimensional NMR spectroscopy via a deep neural network

Multidimensional nuclear magnetic resonance (NMR) spectroscopy is used to examine the chemical structures of the studied systems. Unfortunately, the application of NMR spectra is limited by their long acquisition time, especially for 3D, 4D, and higher dimensional spectra. Non-uniform sampling (NUS) has been widely recognized as a powerful tool to reduce the NMR experimental time. But the quality of NUS spectra depends on appropriate reconstruction algorithms. As an effective data processing method, deep learning has been widely used in many fields in recent years. In this work, a deep learning-based strategy for fast reconstruction of non-uniform sampling NMR spectra is proposed. In our experiments, the proposed deep neural network has better performance in removing artifacts and preserving weak peaks than typical convolutional neural networks of U-Net and DenseNet. Besides, a novel approach of generating training data is utilized to reduce the computational burden of neural networks, and thus training our network can be easier and faster than previous deep learning-based works. Compared with the two currently available methods, SMILE and hmsIST, our strategy can provide comparable reconstruction quality in terms of peak intensities and the fidelity of peak shape. The reconstruction time of our methods is also comparable to or faster than the two methods, especially for 3D spectra.