Deep unfolding blind source unmixing for multicolor fluorescence imaging

Multicolor fluorescence imaging enables simultaneous visualization of multiple cellular components through specific labeling. However, an increase in the number of fluorophores leads to severe spectral overlap, necessitating spectral unmixing to eliminate crosstalk between channels. Because reference spectra are difficult to accurately acquire and multicolor mixed images inherently contain noise, the difficulty of spectral unmixing is substantially increased. Here, we propose a deep unfolding blind source unmixing method, termed DuBsUnmix. The optimization problem is first decomposed via non-negative matrix factorization into sub-problems for the spectral matrix and the abundance matrix. These two sub-problems are then alternately unrolled into a deep unfolding network, where projected gradient descent is employed for the spectral learning module, and proximal gradient descent combined with residual network components is used for the abundance learning module. By this design, DuBsUnmix effectively overcomes the challenges of spectral distortion and image noise in spectral unmixing. The method achieves superior unmixing performance across various simulated scenarios. In experiments on real samples, including eight-color fluorescent beads, seven-color mouse brain sections, and six-color live-cell dynamic imaging, DuBsUnmix demonstrates high accuracy, robust performance, and strong morphological generalization capability. This work provides an accurate and robust solution for blind source unmixing in multicolor fluorescence imaging, offering significant value for biomedical research.