Deep-learning autoencoders for unsupervised BCARS chemical imaging

Abstract

The application of Broadband CARS to cell imaging studies has thus far been limited to those where highcontrast features are present, such as lipids and exogenously introduced tags. This is due to the inherent low SNR obtained in BCARS from the low density of oscillators in single cells coupled with the non-resonant background present in all media which distorts the measured signal. In this paper, we show that an autoencoder which we named VECTOR2, trained on simulated spectra, can accurately perform NRB removal of recorded BCARS images of unstained biological specimen. This allows cell imaging comparable in time to spontaneous Raman imaging with high bandwidth and resolution. The introduction of standard baseline flattening prior to NRB removal preserves the image structure while removing artefacts from raster scanning and optical noise. This results in a hyperspectral image of the NRB-free BCARS signal which is linear in the sample concentration and has a spectrum that is very similar to the spontaneous Raman spectrum.