Combined non-negative matrix factorization and Otsu method for advanced analysis of Raman hyperspectral images acquired at nano- and microscale

Abstract

Micro- and nanospectroscopic hyperspectral imaging approaches provide comprehensive information about the studied material. While the obtained experimental data are characterized by high dimensionality it may lead to many difficulties in the interpretation of the results. A common approach overcoming this problem is the application of multivariate data analysis, which enables data dimensionality reduction and extraction of crucial information. In this paper, we present our proprietary algorithm to investigate hyperspectral maps of biological samples at various levels of organization. The proposed approach is based on a combination of non-negative matrix factorization and the Otsu method of image binarization. We demonstrated its application in the investigation of a TER hyperspectral map of a DPPE monolayer, which enables the automatic differentiation of spectra corresponding to nanoscopic domains in the lipid film. Moreover, to examine the proposed computational algorithm in the analysis of more complex types of samples, we applied it to a confocal Raman hyperspectral map of a single HeLa cell. The NMF-Otsu analysis automatically distinguished the spectra of lipid droplets and the cellular membrane.