Fig 1

The architecture of LD-Net with L-layer downsampling and upsampling blocks for NMR spectroscopy denoising. 1D convolution (K, C, P) is used in which K, C, P denote the kernel size, channel number, the number of zero paddings, respectively. The original code scripts, trained models, as well as the example datasets would be freely available at https://github.com/HaolinZhan/LD-Net-for-NMR-spectroscopy-denoising."

Fig 2

1D 13C NMR spectra of 0.1 mol∙L–1 cholesterol in CDCl3. (a) A noisy spectrum which is acquired with 4 scans. (b) The reference spectrum acquired with 512 scans. (c, d) The spectra using LD-Net denoising with the layers of 3 and 12, namely L = 3 (c) and L = 12 (d), respectively. (e) The spectra adopting DN-Unet denoising as a contrast. The spectral SNRs after denoising are enhanced by factors of 218 (c), 153 (d), and 277 (e) relative to the noisy input (a), respectively. All reconstruction times were counted on the identical computing hardware."

Table 1

Performance comparison between different methods for NMR spectroscopy denoising a."

Fig 3

1D 13C NMR spectra of a mixture containing 0.5 mol∙L–1 quinine and 0.05 mol∙L–1 menthol in CDCl3. (a) A noisy spectrum acquired with 4 scans. (b) The reference spectrum acquired with 512 scans. (c, d) The spectra using LD-Net denoising with the layers of 3 and 12, namely L = 3 (c) and L = 12 (d), respectively. (e) The spectra using DN-Unet denoising as a contrast."

Fig 4

2D 1H-13C HSQC spectra of 0.2 mol∙L-1 azithromycin in CDCl3. (a) A noisy input acquired with 4 scans and additional Gaussian white noise with standard deviations of 0.02 is added. (b) 2D spectrum collected with 32 scans as a reference. (c, d) 2D spectra after 12-layer LD-Net denoising (c) and DN-Unet denoising (d). (e, f) 1D slices for each of 2D spectra from the indicated regions. The 2D contour levels for panels c and d remain identical for a fair comparison, and panels a and b are contoured with the smaller levels for the better visualization. The black arrows denote the peaks with negative phases in panels b and e."

Fig 5

2D 1H-15N HSQC spectra of 0.25 mmol∙L–1 15N-labeled BRM bromodomain. (a, e, i) Noisy inputs acquired with 8 scans (a) and adding random Gaussian white noise with standard deviations of 0.02 (e) and 0.03 (f). (b, f, j) 2D spectra after 3-layer LD-Net denoising on inputs of panels a, e, i, respectively. (c, g, k) 2D spectra after 12-layer LD-Net denoising on inputs of panels a, e, i, respectively. (d, h, l) 2D spectra after DN-Unet denoising on inputs of panels a, e, i, respectively. The 2D contour levels for all panels apart from the panel l remain identical for a fair comparison, and the panel l is contoured with a smaller level for the better visualization."