Self-Supervised\nLearning for Generic Raman Spectrum\nDenoising

Raman and surface-enhanced Raman scattering (SERS) spectroscopy\nare highly specific and sensitive optical modalities that have been\nextensively investigated in diverse applications. Noise reduction\nis demanding in the preprocessing procedure, especially for weak Raman/SERS\nspectra. Existing denoising methods require manual optimization of\nparameters, which is time-consuming and laborious and cannot always\nachieve satisfactory performance. Deep learning has been increasingly\napplied in Raman/SERS spectral denoising but usually requires massive\ntraining data, where the true labels may not exist. Aiming at these\nchallenges, this work presents a generic Raman spectrum denoising\nalgorithm with self-supervised learning for accurate, rapid, and robust\nnoise reduction. A specialized network based on U-Net is established,\nwhich first extracts high-level features and then restores key peak\nprofiles of the spectra. A subsampling strategy is proposed to refine\nthe raw Raman spectrum and avoid the underlying biased interference.\nThe effectiveness of the proposed approach has been validated by a\nbroad range of spectral data, exhibiting its strong generalization\nability. In the context of photosafe detection of deep-seated tumors,\nour method achieved signal-to-noise ratio enhancement by over 400%,\nwhich resulted in a significant increase in the limit of detection\nthickness from 10 to 18 cm. Our approach demonstrates superior denoising\nperformance compared to the state-of-the-art denoising methods. The\nocclusion method further showed that the proposed algorithm automatically\nfocuses on characterized peaks, enhancing the interpretability of\nour approach explicitly in Raman and SERS spectroscopy.