Spectral unmixing for in vivo fluorescence imaging based on accurate target-to-background estimation

Spectral unmixing is a useful technique in fluorescence imaging for reducing the effects of background fluorescence (BF), also called autofluorescence (AF), and separating multiple fluorescence probes. But it is complicated by the significant overlap of the fluorophore emission spectra, and the strong BF signal is often highly mixed with all multi-target fluorescences and can have a confusing effect on the measurement of the multi-target fluorescences. In this work, we introduce a spectral unmixing algorithm tailored for in vivo optical imaging, which effectively separates the multi-target fluorescence from the BF without any hardware-based BF acquisition or a prior knowledge of in-vitro spectra. First, we use kernel maximum autocorrelation factor analysis (kMAF) to accurately detect and separate multi-target fluorescence regions from the BF in sparse multispectral observation data. The observation data being outside of the target regions only contain BF, so we can get accurate spectral estimation of the BF. With the accurate target-to-background fluorescence estimation, the multi-target fluorophores and BF can be easily unmixed in simulated and in vivo experimental data by using multivariate curve resolution-alternating least squares method (MCR-ALS).