Simulation-Driven Deep Learning Framework for Raman Spectral Denoising Under Fluorescence-Dominant Conditions

TL;DR

This work tackles Raman spectroscopy in fluorescence-rich biological tissues by engineering a simulation-driven denoising pipeline that integrates a physics-based noise model with deep learning. A dual-stage AUnet denoiser first suppresses stochastic detector noise in the frequency domain, then removes fluorescence baselines to recover faithful Raman spectra; the system is trained and validated entirely on synthetically generated data calibrated to realistic noise and tissue-like fluorescence. Key findings show superior SNR improvement, more accurate peak recovery, and improved downstream skin-component quantification compared with traditional methods, demonstrated on simulated human skin spectra. The approach promises faster, more reliable Raman-based tissue analysis and provides a versatile framework for extending denoising to other spectroscopic contexts, pending real-data validation and domain adaptation.

Abstract

Raman spectroscopy enables non-destructive, label-free molecular analysis with high specificity, making it a powerful tool for biomedical diagnostics. However, its application to biological tissues is challenged by inherently weak Raman scattering and strong fluorescence background, which significantly degrade signal quality. In this study, we present a simulation-driven denoising framework that combines a statistically grounded noise model with deep learning to enhance Raman spectra acquired under fluorescence-dominated conditions. We comprehensively modeled major noise sources. Based on this model, we generated biologically realistic Raman spectra and used them to train a cascaded deep neural network designed to jointly suppress stochastic detector noise and fluorescence baseline interference. To evaluate the performance of our approach, we simulated human skin spectra derived from real experimental data as a validation case study. Our results demonstrate the potential of physics-informed learning to improve spectral quality and enable faster, more accurate Raman-based tissue analysis.

Figures (8)

• Figure 1: Overview of the dual-stage AUnet-based denoising architecture. The first stage removes stochastic noise via DCT, and the second stage refines the denoised spectrum to match the clean Raman signal.
• Figure 2: Simulated spectra examples. Left columns are two examples of pure Raman spectrum and fluorescence spectrum examples. Right columns are the combination the Raman and the fluorescence with the noise added.
• Figure 3: SNR improvement comparison. (a) Example Raman spectra before and after denoising using Savitzky-Golay (SG) filter, wavelet filter, and the proposed deep learning (DL) model. Two representative input spectra with initial SNRs of 12.5 and 6.4 are shown. DL achieves a visually smoother result that more closely matches the ground truth. (b) Quantitative comparison of SNR improvement across 500 $(\text{r2f}, \text{SNR})$ pairs. Each point represents the average improvement over 5 signals per pair. DL consistently outperforms SG and wavelet filters across different noise and fluorescence conditions.
• Figure 4: Raman peak evaluation across prominence thresholds. (a) Missing peak ratio and artifact peak ratio for DL and PolyFit models. DL consistently produces fewer false negatives across all prominence levels. But for the false positives, becaus (b) Peak value bias and peak shift of matched peaks. DL achieves lower intensity bias. Error bars represent standard deviation.
• Figure 5: Denoising results on simulated human skin spectra at different SNR levels. Each panel shows the original noisy input (gray), traditional polynomial baseline result (blue), our DL model output (green), and ground truth (orange). Highlighted regions indicate Raman bands of interest.