Improvement in Signal-to-Noise Ratio of Liquid-State NMR Spectroscopy via a Deep Neural Network DN-Unet

Nuclear
magnetic resonance (NMR) is one of the most powerful analytical
tools and is extensively applied in many fields. However, compared
to other spectroscopic techniques, NMR has lower sensitivity, impeding
its wider applications. Using data postprocessing techniques to increase
the NMR spectral signal-to-noise ratio (SNR) is a relatively simple
and cost-effective method. In this work, a deep neural network, termed
as DN-Unet, is devised to suppress noise in liquid-state NMR spectra
to enhance SNR. It combines structures of encoder–decoder and
convolutional neural network. Different from traditional deep learning
training strategy, M-to-S strategy is developed to enhance DN-Unet
capability that multiple noisy spectra (inputs) correspond to a same
single noiseless spectrum (label) in the training stage. The trained
1D model can be used for denoising not only 1D but also high dimension
spectra, further improving DN-Unet’s performance. 1D, 2D, and
3D NMR spectra were utilized to evaluate DN-Unet performance. The
results suggest that DN-Unet provides larger than 200-fold increase
in SNR with weak peaks hidden in noise perfectly recovered and spurious
peaks suppressed well. Since DN-Unet developed here to increase SNR
is based on data postprocessing, it is universal for a variety of
samples and NMR platforms. The great SNR enhancement and extreme excellence
in differentiating signal and noise would greatly promote various
liquid-state NMR applications.