Deep Learning for Fluorescence Lifetime Predictions Enables High-Throughput In Vivo Imaging.

Area of Science:

• Biomedical Optics
• Machine Learning in Microscopy
• Fluorescence Spectroscopy

Background:

• Fluorescence lifetime imaging microscopy (FLIM) is crucial for studying microenvironment changes in biomedical research.
• Traditional FLIM methods require high photon counts, leading to long acquisition times and limited throughput for live samples.
• Current techniques struggle with photon-starved data, hindering FLIM's application in dynamic *in vivo* studies.

Purpose of the Study:

• To introduce FLIMngo, a deep learning model for accurate FLIM data quantification in photon-starved conditions.
• To enable high-throughput FLIM analysis with reduced data acquisition times and phototoxicity.
• To enhance the applicability of FLIM for live, dynamic biological specimens.

Main Methods:

• Development of FLIMngo, a deep learning model leveraging both temporal and spatial information in raw FLIM data.
• Quantification of FLIM data from decay curves with fewer than 50 photons per pixel.
• Benchmarking against traditional phasor plot analysis and other deep learning methods using simulated and experimental data.

Main Results:

• FLIMngo accurately predicts fluorescence lifetimes from photon-starved FLIM data, outperforming existing methods.
• The model reduces FLIM data acquisition times to seconds, enabling higher throughput and minimizing phototoxicity.
• Demonstrated successful application in quantifying protein aggregates in live *Caenorhabditis elegans*.

Conclusions:

• FLIMngo significantly enhances FLIM's utility as a high-throughput tool for live biological samples.
• The model's ability to analyze photon-starved data opens new avenues for longitudinal studies in organisms like *C. elegans*.
• FLIMngo is open-source and readily implementable, requiring no retraining for diverse applications.