A comparative study of robustness to noise and interpretability in U-Net-based denoising of Raman spectra

Raman spectroscopy is a valuable analytical technique for molecular characterization, but its practical application is often restricted by low signal-to-noise ratio (SNR), especially at short integration times. These limitations are critical in time-sensitive applications where rapid spectral acquisition is required. Deep learning, in particular, U-Net-based models are becoming routine tools for denoising the spectra. However, such models are typically applied as black boxes that are only evaluated using performance metrics. In this study, we investigate how training strategies using spectra acquired at different integration times and thus varying noise levels, affect model generalization. Specifically, we compare two models trained with the same U-Net architecture: a Single-Condition (SC) model trained on spectra acquired with a single integration time, and a Multi-Condition (MC) model trained on a dataset combining multiple integration times. Besides the quantitative evaluation of the models using Root Mean Squared Error (RMSE) and Pearson Correlation Coefficient (PCC), we apply interpretability techniques to gain deeper insight into how the models process spectral data. Saliency maps and Jacobian matrix analysis enable us to visualize which spectral regions the models focus on during denoising. This study highlights that training with diverse integration times significantly improves model generalization and denoising robustness. Moreover, our use of interpretability techniques reveals how different training strategies influence model focus and decision-making, offering a novel perspective on designing explainable and noise-resilient deep learning models for Raman spectroscopy.