AutoUnmix: an autoencoder-based spectral unmixing method for multi-color fluorescence microscopy imaging.

Area of Science:

• Biomedical imaging
• Microscopy
• Computational biology

Background:

• Multiplexed fluorescence microscopy is crucial for biomedical research.
• Spectral leaks and overlapping in simultaneous multi-fluorophore imaging degrade image quality and analysis.
• Current spectral unmixing methods are computationally intensive and rely on reference spectra.

Purpose of the Study:

• To develop a deep learning-based blind spectral unmixing method for improved fluorescence microscopy.
• To create a method that overcomes limitations of existing spectral unmixing techniques.
• To enhance image quality and analytical capabilities in multi-fluorophore imaging.

Main Methods:

• Proposed AutoUnmix, a deep learning-based blind spectral unmixing method.
• Implemented a transfer learning framework for adaptability across imaging systems.
• Validated performance on synthetic datasets and biological samples.

Main Results:

• AutoUnmix achieves real-time unmixing, up to 100-fold faster than existing methods.
• Demonstrated superior reconstruction performance with the highest SSIM of 0.99 in three- and four-color imaging.
• Outperformed popular unmixing methods by nearly 20% in quantitative performance.

Conclusions:

• AutoUnmix offers data-independent and superior blind unmixing performance.
• The method significantly enhances image quality and analysis in multi-fluorophore microscopy.
• AutoUnmix is a powerful tool for studying interactions of organelles labeled by multiple fluorophores.