VECTOR: Very deep convolutional autoencoders for non?resonant background removal in broadband coherent anti?Stokes Raman scattering

Rapid label-free spectroscopy of biological and chemical specimen via molecular vibration through means Broadband Coherent Anti-Stokes Raman Scattering (BCARS) could serve as a basis for robust diagnostic platform wide range applications. A limiting factor CARS is the presence non-resonant background (NRB) signal, endemic to technique. This multiplicative with chemically resonant meaning perturbation it generates cannot be accounted simply. Although several numerical approaches exist account remove NRB, they generally require some estimate NRB in form separate measurement. In this paper, we propose deep neural network architecture called VECTOR (Very dEep Convolutional auTOencodeRs), which retrieves analytical Raman-like spectrum from spectra training simulated noisy spectra, without need an reference comprised encoder decoder. The aims compress input lower dimensional latent representation losing critical information. decoder learns reconstruct compressed representation. We also introduce skip connection bypasses decoder, benefits reconstruction performance deeper networks. conduct abundant experiments compare our proposed previous literature, including widely applied Kramers-Kronig method, well two another recently methods that use