Self-Supervised Denoising Network for Quantitative Stimulated Raman Scattering Microscopy in Cross-Type Samples

Stimulated Raman scattering (SRS) microscopy has emerged as a vital analytical tool in the fields of biology and medicine. However, the prevalent noise in SRS images has limited its ability for high-sensitivity and high-speed imaging. Here, we introduce DenoiseGAN, a self-supervised network that can reduce random noise while maintaining structural consistency in SRS images. DenoiseGAN was demonstrated in different sample types, including gastric cancer exfoliated cells, bladder cancer tissues, and fungal cells, achieving an up-to 7-fold increase in the signal-to-noise ratio and a structural similarity index measure greater than 0.81. Moreover, DenoiseGAN can preserve quantitative molecular information, such as the protein/lipid ratio in gastric cancer exfoliated cells and bladder cancer tissues, allowing for the differentiation of positive and negative tumor specimens, as well as the heavy water (D₂O) metabolism in fungal cells, enabling rapid antifungal susceptibility testing with SRS images. This work underscores the potential of DenoiseGAN as a broadly applicable denoising network for SRS images.