NMR Metabolite Spectrum Denoising with Improved U-Net++

Nuclear Magnetic Resonance (NMR) spectroscopy serves as a significant tool for capturing essential information in chemical reactions. However, in practical applications, NMR signals are frequently afflicted by Gaussian noise, leading to challenges in sequence analysis and quantification. In this paper, we have devised a deep neural network named NMRDnet, with the Unet++ architecture at its core, incorporating modules such as CBAM and soft-thresholding. This network was trained using a simulated dataset based on the exponential characteristics of NMR spectroscopic signals. The experimental results show that the NMRDnet network can effectively remove the noise and restore the original signal. It is superior to the low-rank Hankel matrix denoising method in maintaining peak signal strength and reducing signal distortion.