A deep learning-based method to spectrally separate overlapping fluorophores based on their fluorescence lifetime.

To investigate biological samples via fluorescence microscopy, fluorophores should be carefully chosen in order to avoid spectral overlap. This poses constraints on the number and type of fluorophores that can be used at the same time on the sample. The fluorescence lifetime is a useful feature to demultiplex the fluorescence signal of spectrally overlapping fluorophores. So far, several methods have been proposed to separate fluorophores based on temporal or spectral fingerprints, including phasor approach and SPLIT. However, these methods usually rely solely on linear separation. During the last decade, the development of machine learning approaches offers great possibilities in terms of data analysis and processing speed. In this contribution, we exploited the deep learning benefits to separate two spectrally overlapping fluorophores using their fluorescence lifetime. Taking into account also the non-linear contributions, our solution offers improved performance with respect to the state-of-the-art approaches. The training was implemented on synthetic images of cells stained with two spectrally overlapped fluorophores, while the test was performed on both synthetic and real images.