SabreSDN: A Complex-Valued 1D U-Net for Denoising SABRE-Hyperpolarized NMR Spectra at Low Concentration

Signal Amplification By Reversible Exchange (SABRE) can enhance nuclear magnetic resonance (NMR) signals by several thousand-fold through para-hydrogen-derived hyperpolarization. However, at low substrate concentrations (50~$\mu$M and below), the signal-to-noise ratio (SNR) remains insufficient for reliable spectral analysis, particularly in time-resolved measurements where signal averaging is precluded by the finite hyperpolarization lifetime. Traditional denoising methods such as exponential windowing and Cadzow filtering rely on manually tuned parameters and fail to preserve fine spectral features. We introduce SabreSDN (SABRE Signal Denoising Network), a complex-valued 1D U-Net that operates directly on complex NMR spectra (8192~points, real and imaginary channels). The network incorporates four key architectural components: (1)~complex convolutions (CConv1d) that encode the absorption--dispersion Hilbert-transform relationship as a structural prior, halving parameter count relative to unconstrained real-valued convolutions; (2)~CBAM1d attention modules that jointly model channel-wise and frequency-axis feature importance; (3)~a MultiScaleBlock bottleneck with parallel dilated convolutions (dilation~1,~4,~8) to simultaneously capture narrow singlets, $J$-coupled multiplets, and broad spectral features; and (4)~gated skip connections that selectively suppress noisy encoder features during decoding. Training uses a three-component composite loss that allocates $4\times$ higher weight to peak-containing regions ($\sim$2.4\% of spectral points) together with frequency-domain amplitude matching and baseline smoothness regularization. Trained on 50~$\mu$M SABRE-hyperpolarized $^{19}$F NMR data (20,000 noisy--clean pairs, 80/20 train/val split across 50~epochs), the model reduces validation loss from 0.162 to 0.024. Denoised spectra show clear recovery of $J$-coupling multiplet fine structure with minimal baseline distortion. With 7.71~million parameters and $\sim$2~ms GPU inference per spectrum, SabreSDN is practical for integration into existing SABRE NMR acquisition pipelines. Code and model weights are available at \url{ https://github.com/kimariyb/SABRE-Denoise }.