Deep Learning-Based Restoration of Distorted Transmission Raman Spectra through Biological Tissue.

Retrieving optical information from photons traversing scattering media is essential in fields relating to detection and imaging. Raman spectroscopy offers high chemical specificity for molecular analysis, and the emerging deep Raman techniques, such as transmission Raman spectroscopy (TRS), enable subsurface probing. However, as biological tissues are highly scattering media, in vivo Raman applications are severely limited by tissue-induced signal attenuation and spectral distortion, which undermine quantitative accuracy. Here, we developed a deep learning-based framework to restore Raman spectra acquired after propagating through biological tissue. We built a comprehensive data set of 4410 paired pre- and post-transmission Raman or surface-enhanced Raman scattering (SERS) spectra from 18 Raman-active samples. Using this data set, we systematically characterized tissue-induced spectral distortions and trained a 1D U-Net model to learn the inverse transformation. This model effectively restored attenuated intensities, suppressed noise, and reconstructed spectral profiles. On the independent testing set, the restored spectra exhibited remarkably improved similarity to ground-truth profiles, achieving >95% average cosine similarity and substantially reduced distortions in both absolute and relative intensities. Furthermore, restoration enhanced molecular quantification, yielding clearer concentration-response relationships for mixed SERS nanoparticles. These results demonstrate the effectiveness of our model and show that data-driven full-spectrum restoration can effectively counteract tissue-induced degradation, improving the accuracy of Raman-based quantification through scattering of biological media.