Abstract
Time lapse fluorescence microscopy can be used to monitor the behaviour of individual cells over time, but the imaging results can be affected by the inherent problem of overlapping emission spectra, commonly referred to as cross-talk. Various spectral unmixing methods have been adopted to solve this problem, thereby estimating the original signal from the observed mixture. The methods for evaluating the quality of spectral unmixing are usually based on the reference objects: a set of pixels where one of the signals is present while another is absent. However, reference objects are not always readily available, particularly in experiments that involve multi-colour fluorescence reporters. Moreover, there is little discussion about how we select training data to estimate the cross-talk coefficients of spectral unmixing methods. Our purpose is to address the challenge of missing reference objects by proposing a new quality evaluation approach, and to consider different strategies of training data selection. To this end, we develop a correlation-based quality measure that employs spatiotemporal information, and compare spectral unmixing results obtained using different training strategies.