Simple phasor-based deep neural network for fluorescence lifetime imaging microscopy.

Area of Science:

• Biophysics
• Microscopy
• Computational Biology

Background:

• Fluorescence lifetime imaging microscopy (FLIM) provides insights into fluorophore environments.
• Current FLIM analysis relies on time-consuming fitting methods.
• Existing deep learning models require extensive training and are limited to specific lifetime ranges.

Purpose of the Study:

• To develop a computationally efficient and accurate method for analyzing FLIM data.
• To accelerate the analysis of fluorescence lifetime images using a simple neural network.
• To enable precise characterization of molecular environments in live cell imaging.

Main Methods:

• A simple neural network, Phasor-Net, utilizing fully connected layers was designed.
• High-dimensional fluorescence intensity temporal decays were reduced to phasor coordinates, mean, and amplitude-weighted lifetimes.
• The network was trained and validated on simulated biexponential decays within a restricted time interval (12.5 ns) and photon count (<10^6).

Main Results:

• Phasor-Net demonstrated higher precision and reduced bias compared to standard fitting methods for biexponential decays.
• The simple neural network achieved performance comparable to more sophisticated deep learning architectures.
• Phasor-Net exhibited a significantly faster training process (15 minutes vs. 30 minutes).

Conclusions:

• Phasor-Net offers a precise and accelerated approach for FLIM data analysis.
• The method is suitable for live cell imaging applications with limited photon counts and time intervals.
• Successful application to determine biexponential decay parameters in living cells expressing EGFP-mCherry fusion proteins was demonstrated.