Decision-tree-based non-resonant background removal for enhancing chemical-selective contrast in hyperspectral CARS microscopy

Coherent anti-Stokes Raman scattering (CARS) is a nonlinear optical process used for spectroscopy and chemical imaging. CARS signals can be orders of magnitude stronger than those of its incoherent counterpart, spontaneous Raman scattering, thus enabling substantially faster acquisition speeds. This attribute has positioned CARS as a desirable alternative to spontaneous Raman scattering as a contrast mechanism for label-free hyperspectral chemical imaging due to its traditionally long acquisition times. The presence of a non-resonant background (NRB) that distorts the shapes and intensities of resonant peaks and introduces spurious signal to non-resonant spectral regions, however, complicates spectral analysis and degrades chemical-selective image contrast. The NRB has and continues to hinder the widespread adoption of CARS despite its clear advantages. NRB removal techniques that retrieve Raman-like signals from CARS spectra have long been a central focus of CARS research, with "deep" machine learning approaches being most recently explored. Here, we present an original "shallow" machine learning approach to NRB removal based on gradient-boosted decision trees obtained using the open-source gradient-boosting framework XGBoost. We find that the tree-based model accurately retrieves Raman-like spectra, both in simulated and experimental CARS spectra. When applied to experimental hyperspectral CARS microscopy, the tree-based model significantly improves chemical-selective contrast, thereby facilitating the spatiospectral analysis of samples within the region of interest. This work establishes tree-based gradient boosting as an effective and viable tool for NRB removal for enhancing chemical-selective contrast in hyperspectral CARS microscopy.