Deep learning for extracting Raman signal from coherent anti-Stokes Raman scattering spectrum

Abstract

Coherent anti-Stokes Raman scattering (CARS) microscopy has been shown to be very successful in various applications such as biomedical imaging, cell biology, and material science. One challenge in extracting the Raman signal from the CARS spectrum is the non-resonant background (NRB) contribution, which deforms the spectral line shape and deteriorates the quantitative information. To solve this problem, several numerical techniques alongside other algorithmic methods have been introduced. A state-of-the-art approach to extract the Raman signal from the CARS spectrum is deep learning (DL), which forms the basis of three studies in this dissertation.

In the first study, a pre-trained convolutional neural network (CNN) with synthetic data has been fine-tuned, utilising two sets of semi-synthetic spectra. The model shows 86% accuracy on prediction with regard to the Raman signal from the semi-synthetic CARS spectrum. Regarding the experimental CARS spectra prediction, the fine-tuned model was found to be better in comparison to the original model.

In the second study, the hyperparameters of SpecNet, a CNN model from the literature, has been optimised to avoid two major problems of the SpecNet. The first problem is not finding all peaks of the spectrum, especially the ones at the edges of the signal, and the second problem relates to not matching the exact intensity of the finding peaks across the spectrum. The results show that on both synthetic and experimental spectra, the optimized model works much better than the original one.

In the third study, the extraction of the Raman signal from the CARS spectrum has been explored via bidirectional long short-term memory (Bi-LSTM) for the first time, and has been compared to three DL models including LSTM, CNN, and a very deep convolutional autoencoder (VECTOR). All four models have been trained with synthetic spectra, whilst three different NRB were used to prepare the data. Bi-LSTM model surpasses the performance of the other models on extracting the Raman signal from the CARS spectra with both synthetic and experimental spectra.

In general, the automatic extraction of the Raman signal from the CARS spectrum has been investigated by utilising DL models to avoid the need for manually set parameters of typical methods. These models perform comparably with the standard numerical methods and enable automating an important processing step of CARS spectra.