Deep learning for removal of non-resonant background in CARS hyperspectroscopy

In this work, a deep learning approach proposed by Valensise et al. [3] for extracting Raman resonant spectra from measured broadband CARS spectra was explored to see how effective it is at removing NRB from our experimentally measured "spectral-focusing"-based approach to CARS. A large dataset of realistic simulated CARS spectra was used to train a model capable of performing this spectral retrieval task. The non-resonant background shape used in creating the simulated CARS spectra was altered, to mimic our experimentally measured NRB response. Two models were trained: one using the original approach (Specnet) and one using the updated NRB "Specnet Plus", and then tested their ability to retrieve the vibrationally resonant spectrum from simulated and measured CARS spectra. An error analysis was performed to compare the model's retrieval performance on two simulated CARS spectra. The modified model's mean squared error value was five and two times lower for the first and second simulated CARS spectra, respectively. Specnet Plus was found to be more effective at extracting the resonant signals. Finally, the NRB extraction abilities of both models are tested on two experimentally measured CARS hyperspectroscopy samples (starch and chitin), with the updated NRB model (Specnet Plus) outperforming the original Specnet model. These results suggest that tailoring the approach to reflect what we observe experimentally will improve our spectral analysis workflow and increase our imaging potential.