A multiscale approach for surface-enhanced Raman spectroscopy (SERS) spectrum representation and its application to bacterial discrimination

A multiscale approach for peak representation and identification is proposed in the context of surface-enhanced Raman spectroscopy (SERS). The SERS spectrum is naturally composed of peaks with different preferred scales, and some tiny but reproducible peaks can be important distinguishing factors. However, if only a single scale is employed, there is a significant probability that a tiny peak will be overlooked because its preferred scale might be far away from the chosen scale. In contrast to many established approaches, the proposed method is able to adaptively characterize each peak on its preferred scale so that relevant information is not overlooked. We demonstrate the merit and biological relevance of multiscale peak identification with spectra from real biological experiments. In addition, the practicability of the proposed approach is evaluated on two bacterial discrimination tasks. The proposed method outperforms the single-scale approach, which is optimized trial-by-trial in the experiments.