Photonic Data Science

Fluorescence emission has been one of the major obstacles to apply Raman spectroscopy in biological investigations. It is usually several orders more intense than Raman scattering and hampers further analysis. In cases where the fluorescence emission is too intense to be efficiently removed via routine mathematical baseline correction algorithms, an alternative approach is needed. One alternative approach is shifted-excitation Raman difference spectroscopy (SERDS), where two Raman spectra are recorded with two slightly different excitation wavelengths. Ideally, the fluorescence emission at the two excitations does not change while the Raman spectrum shifts according to the excitation wavelength. Hence the fluorescence is removed in the difference of the two recorded Raman spectra. For better interpretability a spectral reconstruction procedure is necessary to recover the fluorescence-free Raman spectrum. This is challenging due to the intensity variations between the two recorded Raman spectra caused by unavoidable experimental changes, as well as the presence of noise. Existent approaches suffer from drawbacks like spectral resolution loss, fluorescence residual, and artefacts. In this contribution, we proposed a reconstruction method based on non-negative least squares (NNLS), where the intensity variations between the two measurements are utilized in the reconstruction model. The method achieved fluorescence-free reconstruction on three real-world SERDS datasets without significant information loss. Thereafter, we quantified the performance of the reconstruction based on artificial datasets from four aspects: reconstructed spectral resolution, precision of reconstruction, signal to-noise-ratio (SNR), and fluorescence residual. The artificial datasets were constructed with varied Raman to fluorescence intensity ratio(RFIR), SNR, full-width at half-maximum (FWHM), excitation wavelength shift, and fluorescence variation between the two spectra. It was demonstrated that the NNLS approach provides a faithful reconstruction without significantly changing the spectral resolution. Meanwhile, the reconstruction is almost robust to fluorescence variations between the two spectra. Last but not the least the SNR was improved after reconstruction for extremely noisy SERDS datasets.