hyperPICASSO: adapting the mutual-information-based unmixing technique PICASSO to hyperspectral datasets

Area of Science:

• Biophotonics
• Spectroscopy
• Computational Biology

Background:

• Mutual-information (MI)-based unmixing algorithms, like PICASSO, effectively reduce spectral overlap in multiplexed fluorescent imaging.
• Current methods often use naive binning for hyperspectral data, losing valuable spectral information and reducing unmixing efficiency.

Purpose of the Study:

• To develop a generalized MI-based unmixing algorithm for hyperspectral datasets.
• To improve the accuracy and efficiency of spectral unmixing in highly multiplexed fluorescence imaging.

Main Methods:

• Developed the hyperPICASSO algorithm, extending MI-based unmixing to hyperspectral data.
• Extracted spectral features from annotated images for approximate linear unmixing.
• Iteratively refined unmixing by pairwise minimization of MI.

Main Results:

• hyperPICASSO significantly reduced spectral crosstalk compared to standard PICASSO on binned data.
• Outperformed linear unmixing using measured spectra, especially for overlapping emission spectra.
• Demonstrated effective crosstalk reduction by leveraging hyperspectral data.

Conclusions:

• MI-based unmixing is enhanced by utilizing full hyperspectral data.
• hyperPICASSO offers superior spectral unmixing performance for complex biological samples.
• The algorithm improves the fidelity of quantitative analysis in multiplexed fluorescence imaging.