A Simulation-Driven Deep Learning Approach for Artifact-Resistant Raman Spectral Unmixing

Raman hyperspectral imaging provides label-free chemical contrast, but quantitative analysis is limited by weak signals, fluorescence background, instrumental noise, and overlapping component spectra. Here, we present a simulation-driven cascaded attention U-Net (AU-Net) for Raman spectral denoising and targeted component unmixing directly from raw spectra. The training data were generated from measured pure-component references combined with randomized fluorescence baselines, nuisance Raman features, reference-spectrum perturbations, and mixed Poisson-Gaussian noise. This design provides paired raw inputs, denoised spectra, and component-level targets without requiring experimentally annotated mixture spectra. AU-Net was evaluated on simulated lipid-protein-DNA mixtures, sugar-mixture spectra, and HeLa-cell Raman hyperspectral images and was compared with NNLS, MCR-ALS, and an autoencoder-based Raman unmixing baseline under explicit preprocessing and reference-use assumptions. The benchmarks show that the raw-input neural workflow improves component recovery under the tested noise, baseline, and interference conditions, while additional stress tests define limitations under reference-spectrum variation, bilinear mixing, and changed target libraries. These results establish AU-Net as a reference-anchored workflow for repeated Raman analyses with a stable component library rather than a blind or universally transferable unmixing method.