Target-To-Background Separation Based Spectral Unmixing For In-Vivo Fluorescence Imaging With Sparse Channels

In-vivo fluorescence imaging is a rapidly growing field with broad applications in life science due to its low cost and ease of use. However, unlike the in-vitro experiment, the target in-vivo fluorescence is often mixed with the tissue fluorescence, also known as background fluorescence, which makes the results difficult to recognize and classify. Tissue fluorescence is widely distributed in visible range and fully mixed with target fluorescence in all spectral bands which can't be treated as an endmember. Hence, the current linear spectral unmixing methods often fail to separate target fluorescence probes from BF in multispectral observations. Until now, people mainly use hardware equipment matched with special fluorescence dyes to eliminate the background fluorescence, but the operation could be infeasible and none of the method solves the problem. So because the limitation of fluorescence channels, and background signal is fully mixed with target fluorescent, so even the traditional spectrum unmixing method for the minimum convex analysis cannot effectively extract the pure target fluorescence without background signal.In this work, we treat the background fluorescence as a constant and add a new matrix B to the classic linear spectra unmixing model. We exploit the intrinsic accumulation contrast in target-to-background fluorescence to detect and separate multi-target fluorescence areas from the background in sparse multispectral observation data utilizing kernel maximum autocorrelation factor analysis. Then we use the fast marching-based image inpainting method to patch up the removed target fluorescence areas and reconstruct the multispectral BF. With the BF matrix being reconstructed and subtracted from the original observation data, the multiple target fluorophores could be easily unmixed from the subtracted observation data using current spectral unmixing method. Unlike other methods, our technology is fully automated and unsupervised for sparse spectral unmixing without any prior knowledge about the spectra. Under a set of experiments with multiple fluorescent probes at the visible region, our proposed method demonstrated excellent performances to detect multiple target fluorescences while other state-of-art methods failed to get desired results.