Assessment of unmixing approaches for the quantitation of SERS nanoparticles in highly multiplexed spectral images

Surface-enhanced Raman scattering nanoparticles (SERS NPs) offer powerful optical contrast features for imaging assays. Their gold core enhances the inelastic scattering cross section, allowing highly sensitive and rapid detection, and their characteristic sets of narrow spectral bands give them unsurpassed multiplexing capabilities. Multiplexed hyperspectral images are commonly unmixed using a compensation matrix of reference spectra to produce quantitative image channels illustrating the distribution of each material. It is these unmixed channels that are fit for interpretation from assays utilizing SERS NP contrast agents. Some factors that may impact SERS NP quantitative and dynamic range capabilities may include endogenous background heterogeneity, the ability of unmixing algorithms to account for signal variances, and linear system conditioning imposed by contrast agent signals. We report on hyperspectral Raman imaging of mixtures of SERS NPs from an expanded library of contrast agents. We study increasing plexity and varying degrees of system conditioning as inputs to a diverse set of classical, non-negatively constrained, and regularized regression algorithms to investigate which signal features and unmixing methods deliver the most promising quantitation performance with the least error. Raman imaging of SERS NP mixtures is performed on controlled substrates and representative biological specimens, and experimental results are compared against ground truth data. We evaluate spectral fitting fidelity, quantitation, and specificity correlations with system conditioning. Spectral unmixing with a regularized hybrid of least squares regression with principal component analysis (HLP) algorithm approximated spectra with 3.5× better fitting fidelity and 3× better quantitation robustness with tissue background compared with simpler unmixing routines.