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

Abstract

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

Effect of non-resonant background on the extraction of Raman signals from CARS spectra using deep neural networks

“… 52 Hence, in future studies, NRB could be better approximated for these kinds of applications by considering the details of the excitation laser, such as spectral envelope and phase delay. 35 Further, the order of the polynomial and range of the coefficients can be optimized in future work for better results. Also, experimentally recorded NRB can be utilized in synthesizing training data.…”

“…Wang et al explored Very Deep Convolutional Autoencoders (VECTOR) for the NRB removal, and the performance is compared with SpecNet. 35 In our recent work, we have also demonstrated that the training with semisynthetic data in addition to the synthetic data improves the DL model performance. 36 The SpecNet model performance was found to be poor when compared with our previous work 36 and VECTOR model, 35 where it could not able to extract the spectral lines with minimal intensities.…”

“…35 In our recent work, we have also demonstrated that the training with semisynthetic data in addition to the synthetic data improves the DL model performance. 36 The SpecNet model performance was found to be poor when compared with our previous work 36 and VECTOR model, 35 where it could not able to extract the spectral lines with minimal intensities. More importantly, they have not evaluated the model accuracy quantitatively.…”

“…[30][31][32] In recent times, these DL algorithms have also been deployed to solve the issue of NRB in CARS measurements. [33][34][35][36] Houhou et al have utilized the Long Short-Term Memory (LSTM) model to extract the Raman signal and compared their results with the KK & MEM computational techniques. 33 Valensise et al employed a convolutional neural network (CNN) model for retrieving the imaginary part of CARS spectra.…”

“…Thus, all investigations reported to date have trained their DL model only by considering the same type of NRB irrespective of their DL model architecture, and no other type is explored. [34][35][36] Even the recent VECTOR model has shown that robust performance is merely achieved for the simulated datasets, whereas it is found to be sensitive when dealing with the actual experimental data. It is attributed to the fact that NRB used in that study is simply modelled as a product of the countervailing sigmoids and hinted that further improvements on NRB are required to handle complex CARS data.…”