Rapid elimination of scattering in three-dimensional fluorescence spectra via deep learning

Three-dimensional fluorescence spectra are often affected by scattering effects, traditional scattering elimination methods rely excessively on parameter settings and cannot automatically eliminate scattering in batches, thereby limiting the application of fluorescence spectroscopy technology in rapid online monitoring and analysis of samples. In this study, we have developed a model based on a deep learning CycleGAN to rapidly eliminate scattering from three-dimensional fluorescence spectra. The proposed model efficiently eliminates scattering by simply inputting single or batches of contaminated fluorescent spectra. By training the CycleGAN using a large dataset of simulated three-dimensional fluorescence spectra and employing data augmentation, to the model can transform fluorescence spectra with scattering into ones without scattering. To validate the effectiveness of the proposed methed, we confirmed its generalization and reliability by eliminating scattering from two sets of previously unseen real experimental three-dimensional fluorescence spectra. We evaluated the effectiveness of scattering elimination across various noise levels and scattering widths, using metrics such as the mean absolute error, peak signal-to-noise ratio, structural similarity and cosine similarity. Furthermore, we conducted a component analysis using PARAFAC on the spectra post-scattering elimination, yielding correlation coefficients of >0.97 when compared to that in case of actual components. Finally, we compared the proposed model with traditional mathematical methods, such as blank subtraction and Delaunay triangulation. Results showed that the proposed model can automatically and efficiently eliminate scattering from fluorescence spectra in batches, substantially improving the efficiency of scattering elimination.