Deep learning for real-time removal of the non-resonant background from broadband CARS spectra

Abstract

We present a novel approach to remove the unwanted non-resonant background from Broadband Coherent Anti-Stokes Raman Scattering (B-CARS) spectra, based on deep learning. The unsupervised model is built as a convolutional neural network with seven hidden layers. After training on synthetic data, our model was able to process experimental B-CARS spectra and correctly retrieve all the relevant vibrational peaks. The retrieval time is 100 microseconds per spectrum, faster than the time required to record it. We expect that this model will significantly simplify and speed-up the analysis of B-CARS spectra, allowing real-time retrieval of the vibrational features.