Novel automatic de-scattering method for three-dimensional fluorescence spectra based on deep learning

Abstract

Three-dimensional excitation–emission matrix fluorescence spectroscopy is often plagued by scattering effects, mainly Rayleigh and Raman scattering, and thus must be pre-processed for de-scattering before spectral analysis. In this study, a new method based on deep learning was proposed to automaticly remove the scattering from fluorescence spectra. First, the semantic segmentation model Deeplabv3+ was used to realize the segmentation of the scattering region in the original fluorescence spectrum and compared with other segmentation models. The results show that the mean intersection over union (mIoU) of this model is 89.58 % and its Accuracy is 95.40 %, which are better than those of the other models. Then, the Globally and Locally Consistent Image Completion (GLCIC) network was used to repair the spectral energy of the scattering areas. Model effectiveness in scattering removal is evaluated using metrics such as mean absolute error, mean relative error, structural similarity, and peak signal-to-noise ratio. Furthermore, this study employs PARAFAC for component analysis of the spectra after scattering removal by the model, comparing them with actual components, all yielding correlation coefficients greater than 0.99. The results show that compared with the traditional scattering-removal method based on the mathematical model, the proposed method can automatically remove scattering areas in batch samples’ fluorescence spectra, without the need for manual parameter setting or other processes. This improves scattering removal efficiency, enhances the simplicity, and speed of fluorescence detection technology.