Robust blind spectral unmixing for fluorescence microscopy using unsupervised learning.

Area of Science:

• Microscopy and Imaging Science
• Computational Biology
• Machine Learning Applications

Background:

• Fluorescence microscopy suffers from spectral bleed-through, complicating multi-fluorophore imaging and leading to false positives.
• Existing spectral unmixing methods have limitations, including requiring prior spectral knowledge or restricting fluorophore numbers to detection channels.
• Multiphoton microscopy further complicates spectral unmixing with overlapping emission spectra.

Purpose of the Study:

• To develop a robust and flexible spectral unmixing method for fluorescence microscopy.
• To overcome the limitations of current spectral unmixing techniques, particularly for multi-photon microscopy.
• To enable simultaneous imaging of more fluorophores than hardware limitations typically allow.

Main Methods:

• Developed Learning Unsupervised Means of Spectra (LUMoS), an unsupervised machine learning clustering approach.
• LUMoS learns spectral signatures directly from mixed images, enabling blind channel separation.
• Integrated LUMoS into ImageJ for user-friendly application in fluorescence imaging.

Main Results:

• LUMoS successfully separates multiple fluorophore channels without prior spectral information or channel restrictions.
• Demonstrated robustness in multi-channel separations for two-photon microscopy images via experimental and simulated data.
• Extended LUMoS for background/autofluorescence removal and colocalization analysis.

Conclusions:

• LUMoS provides a flexible and robust solution for spectral unmixing in fluorescence microscopy.
• The method enhances spectral resolution and image clarity without requiring hardware upgrades.
• LUMoS expands the capabilities of two-photon microscopy for simultaneous multi-fluorophore imaging.