Spatial resolution improved fluorescence lifetime imaging via deep learning.

Area of Science:

• Biomedical Imaging
• Computational Biology
• Deep Learning

Background:

• Fluorescence Lifetime Imaging (FLIM) systems often produce low-resolution images, limiting detailed cellular analysis.
• Acquiring high-resolution FLIM data can be time-consuming and technically challenging.
• Existing methods struggle to reconstruct fine spatial details from limited pixel resolution in FLIM.

Purpose of the Study:

• To develop a deep learning approach for generating high-resolution (HR) fluorescence lifetime images from low-resolution (LR) FLIM data.
• To create a robust method for generating semi-synthetic FLIM data suitable for neural network training.
• To introduce a novel hybrid neural network, SRI-FLIMnet, for simultaneous lifetime estimation and LR-to-HR image transformation.

Main Methods:

• Proposed a theoretical framework for training neural networks using semi-synthetic FLIM data with diverse cellular features and lifetime characteristics.
• Developed a degrading model to generate paired LR-HR FLIM images for supervised learning.
• Designed and implemented the spatial resolution improved FLIM net (SRI-FLIMnet), a hybrid neural network for enhanced FLIM reconstruction.

Main Results:

• SRI-FLIMnet demonstrated superior performance in reconstructing spatial information from LR FLIM images.
• Evaluated performance using semi-synthetic data and experimental images of bacterial-infected mouse macrophage cells.
• The proposed data generation method and SRI-FLIMnet successfully achieved superior spatial resolution in FLIM applications.

Conclusions:

• The developed deep learning approach and SRI-FLIMnet provide an effective solution for obtaining HR FLIM images.
• The method enables faster acquisition of high-resolution FLIM data, crucial for detailed biological studies.
• This work advances FLIM capabilities by improving spatial resolution and image quality.