Micro S plit : semantic unmixing of fluorescent microscopy data

Area of Science:

• Biophotonics and Computational Imaging

Background:

• Fluorescence microscopy faces limitations in speed, resolution, and phototoxicity due to optical constraints and photon budgets.
• Current methods often require trade-offs between imaging parameters, hindering simultaneous visualization of multiple structures.

Purpose of the Study:

• To introduce MicroSplit, a novel deep learning-based computational multiplexing technique for simultaneous imaging and unmixing of multiple cellular structures.
• To overcome the limitations of conventional fluorescence microscopy by enabling faster and more photon-efficient imaging.

Main Methods:

• Developed MicroSplit, a computational multiplexing method utilizing Variational Splitting Encoder-Decoder networks.
• Implemented deep learning to computationally unmix up to four superimposed noisy structures from a single fluorescent channel.
• Incorporated uncertainty-aware prediction and estimation of spatially resolved prediction errors.

Main Results:

• MicroSplit successfully separates superimposed noisy structures into distinct, denoised image channels.
• Demonstrated improved imaging speed and photon efficiency.
• Showcased robust performance across diverse datasets, noise levels, and imaging conditions.
• Validated enhanced downstream analysis and reduced photon exposure.

Conclusions:

• MicroSplit offers a powerful solution for computational multiplexing in fluorescence microscopy.
• The technique enables faster, more photon-efficient imaging while improving data quality.
• Open-sourced methods and models facilitate widespread adoption in biological imaging research.