Removing non-resonant background from CARS spectra via deep learning

Broadband Coherent Anti-Stokes Raman Scattering (B-CARS) is a powerful label-free nonlinear spectroscopy technique allowing one to measure the full vibrational spectrum of molecules and solids. B-CARS spectra, however, suffer from the presence of a spurious signal, called non-resonant background (NRB), which interferes with the resonant vibrational one, distorting the line shapes and degrading the chemical information. While several numerical techniques are available to remove this unwanted contribution and extract the resonant vibrational signal of interest, they all require the user's intervention and sensitively depend on the spectral shape of the NRB, which needs to be measured independently. In this work, we present a novel approach to remove NRB from B-CARS spectra based on deep learning. Thanks to the high generalization capability offered by the deep architecture of the designed neural network, trained through realistic simulated spectra, our fully automated model is able to process B-CARS spectra in real time and independently of the detailed shape of the NRB spectrum. This results in fast extraction of vibrational spectra without requiring user intervention or the measurement of reference spectra. We expect that this model will significantly simplify and speed-up the analysis of B-CARS spectra in spectroscopy and microscopy.