Deep learning enhanced fast fluorescence lifetime imaging with a few photons

Abstract

We present a deep learning (DL) framework, termed few-photon fluorescence lifetime imaging (FPFLI), for fast analysis of fluorescence lifetime imaging (FLIM) data under highly low-light conditions with only a few photons per pixel. FPFLI breaks the conventional pixel-wise lifetime analysis paradigm and fully exploits the spatial correlation and intensity information of fluorescence lifetime images to estimate lifetime images, pushing the photon budget to an unprecedented low level. The DL framework can be trained by synthetic FLIM data and easily adapted to various FLIM systems. FPFLI can effectively and robustly estimate FLIM images within seconds using synthetic and experimental data. The fast analysis of low-light FLIM images made possible by FPFLI promises a broad range of potential applications.

Perspectives on label-free microscopy of heterogeneous and dynamic biological systems

“…In fact, machine learning has recently been used to process label-free FLIM images acquired in low SNR conditions. 94 , 95 This allows reliable recovery of FLIM decays with low photon counts, hence enabling fast image acquisition in deep tissue. For example, lifetime estimates with high accuracy were achieved with 50 times fewer photons per pixel (i.e., for exogenous fluorescence and 30 to for autofluorescence from live cells) compared to ground-truth.…”