Target/Background Classification Regularized Nonnegative Matrix Factorization for Fluorescence Unmixing

Abstract

Nonnegative matrix factorization (NMF) is usually applied to multispectral fluorescence imaging for fluorescence unmixing. Unfortunately, most NMF-based fluorescence unmixing methods fail to take advantage of spatial information in data. Besides, NMF is an inherently ill-posed problem, which gets worse in the sparse acquisition of multispectral data (from a small number of spectral bands) due to its insufficient measurements and severe discontinuities in spectral emissions. To overcome these limitations by exploiting the spatial difference between multiple-target fluorophores and background autofluorescence (AF), we propose improved normalized cut to automatically classify all multispectral pixels into target fluorophores and background AF groups. We then initialize NMF by extracting the endmember spectra of target/background fluorescent components in the two groups, and impose a $L_{1/2}$ -norm partial sparseness constraint on merely the abundances of target fluorophores within hierarchical alternating least squares framework of NMF. Experimental results based on synthetic and in vivo fluorescence data show the superiority of the proposed algorithm with respect to other state-of-the-art approaches.