Identifying minimally redundant wavenumbers for vibrational microspectroscopic image analysis

Recent approaches to multispectral microscopy such as infrared, Raman and CARS microscopy produce large amounts of high-dimensional spectra at high spatial resolution. In this context, we propose and validate a method for unsupervised feature selection. Unsupervised feature selection is of relevance in several applications of multispectral imaging techniques, most notably in reducing the measurement time of CARS microscopic experiments. Our feature selection is based on minimizing a mutual-information based measure of redundancy, and can be seen as the unsupervised version of the well established minimal-redundancy-maximal-relevance approach to supervised feature selection. We compare our approach to previously proposed unsupervised feature selection approaches and demonstrate its advantages on two types of multispectral imaging techniques as well as on synthetic data.