Fast fluorescence lifetime imaging and sensing via deep learning

Student thesis: Doctoral Thesis

Abstract

Fluorescence lifetime imaging microscopy (FLIM) has become a valuable tool in diverse disciplines. This thesis presents deep learning (DL) approaches to addressing two major challenges in FLIM: slow and complex data analysis and the high photon budget for precisely quantifying the fluorescence lifetimes. DL's ability to extract high-dimensional features from data has revolutionized optical and biomedical imaging analysis. This thesis contributes several novel DL FLIM algorithms that significantly expand FLIM's scope. Firstly, a hardware-friendly pixel-wise DL algorithm is proposed for fast FLIM data analysis. The algorithm has a simple architecture yet can effectively resolve multi-exponential decay models. The calculation speed and accuracy outperform conventional methods significantly. Secondly, a DL algorithm is proposed to improve FLIM image spatial resolution, obtaining high-resolution (HR) fluorescence lifetime images from low-resolution (LR) images. A computational framework is developed to generate large-scale semi-synthetic FLIM datasets to address the challenge of the lack of sufficient high-quality FLIM datasets. This algorithm offers a practical approach to obtaining HR FLIM images quickly for FLIM systems. Thirdly, a DL algorithm is developed to analyze FLIM images with only a few photons per pixel, named Few-Photon Fluorescence Lifetime Imaging (FPFLI) algorithm. FPFLI uses spatial correlation and intensity information to robustly estimate the fluorescence lifetime images, pushing this photon budget to a record-low level of only a few photons per pixel. Finally, a time-resolved flow cytometry (TRFC) system is developed by integrating an advanced CMOS single-photon avalanche diode (SPAD) array and a DL processor. The SPAD array, using a parallel light detection scheme, shows an excellent photon-counting throughput. A quantized convolutional neural network (QCNN) algorithm is designed and implemented on a field-programmable gate array as an embedded processor. The processor resolves fluorescence lifetimes against disturbing noise, showing unparalleled high accuracy, fast analysis speed, and low power consumption.
Date of Award8 Jun 2023
Original languageEnglish
Awarding Institution
  • University Of Strathclyde
SponsorsUniversity of Strathclyde
SupervisorDavid Li (Supervisor) & Yu Chen (Supervisor)

Cite this

'