Improving fluorescence lifetime imaging microscopy phasor accuracy using convolutional neural networks.

Area of Science:

• * Biomedical imaging
• * Fluorescence microscopy
• * Machine learning in imaging

Background:

• * Fluorescence lifetime imaging microscopy (FLIM) is a powerful biological imaging technique.
• * FLIM faces challenges including low signal-to-noise ratio (SNR), slow acquisition, and complexity.
• * Improving SNR is crucial for accurate analysis and interpretation of FLIM data.

Purpose of the Study:

• * To address the low SNR challenge in FLIM images.
• * To demonstrate the use of pre-trained convolutional neural networks (CNNs) for denoising FLIM data.
• * To enhance the accuracy of fluorophore separation and image segmentation.

Main Methods:

• * Employed pre-trained CNN models for denoising FLIM measurements, eliminating the need for extensive training datasets.
• * Utilized pre-trained networks in the inference stage for rapid computation (milliseconds) and high accuracy.
• * Applied K-means clustering to segmented, denoised images for fluorophore separation.

Main Results:

• * Demonstrated effective noise reduction in FLIM images across various biological samples (in vivo mouse kidney, fixed cells, fixed mouse kidney).
• * Showcased improved segmentation accuracy and enhanced SNR, even in challenging, out-of-distribution conditions (in vivo plant samples).
• * Validated the method's effectiveness in separating fluorophores in noisy FLIM images, particularly for in vivo applications.

Conclusions:

• * The proposed method offers a fast and accurate approach for segmenting FLIM images from any system.
• * Significantly improves the identification of biologically relevant structures in biomedical imaging.
• * Provides a robust solution for denoising FLIM data, especially in scenarios where averaging is not feasible.