EFLIM

A self-supervised, event-based deep learning method for FLIM under extreme low light.

Introduction

EFLIM (Event-Based First-Photon FLIM) is a self-supervised deep learning method for fluorescence lifetime imaging microscopy (FLIM). Unlike conventional approaches that rely on photon histograms, EFLIM treats each excitation event as a binary process — either no photon is emitted, or a single first-arrival photon is detected with precise timing. By leveraging spatial and temporal context, EFLIM estimates lifetime under extremely low-light conditions (PPP<1), reducing photon demand by about three orders of magnitude. This enables fast, high-fidelity intravital imaging with strong robustness to intensity fluctuations. EFLIM opens new opportunities for studying dynamic biological processes in deep tissue.

Overview

• Event-based representation: avoids histogram construction and directly models each excitation event.
• Self-supervised learning: robust training without the need for paired datasets or ground-truth lifetimes.
• Extreme low light capability: accurate lifetime estimation even below 1 photon per pixel (PPP).
• Robustness to artifacts: stable performance despite photobleaching, motion, or intensity fluctuations.
• Broad applicability: enables fast, minimally invasive imaging of dynamic molecular processes in vivo.

Workflow

1. Photon-arrival dataset preparation

• Rearrange raw photon data into a sequence of frames that contain the arrival times of all photons.
• For Becker & Hickl systems: convert .SPC files into .tif format.
• For PicoQuant systems: convert .PTU files into .tif format.
• We recommend using at least 500 frames with PPP > 0.1 in regions of interest.

1.1. Simulation (if raw data are unavailable)

If you do not have experimental raw data, you can generate synthetic photon arrivals using the provided MATLAB scripts.

Example:

./0_simulations/run_simu_USAF1951.m

This script simulates 500 frames with PPP = 0.5 and saves the dataset to:

./simu_USAF1951_PPP0.5

Inside this folder, you will find:

• Photon-arrival frames : ./simu_USAF1951_PPP0.5/raw/frame*.tif
• Ground truth: ./simu_USAF1951_PPP0.5/lt_gt/lt_gt.tif
• FastFLIM (center-of-mass method, CMM): ./simu_USAF1951_PPP0.5/lt_fastflim/lt_fastflim.tif
• Intensity-weighted lifetime visualizations, saved with the default colormap suffix _weddingdayblues (e.g., ./simu_USAF1951_PPP0.5/lt_fastflim/lt_fastflim_lt500-3500_in0-0.5_weddingdayblue.tif)

1.2. Experimental dataset

We provide a representative raw SPC dataset (Becker&Hickl, HPM-100-07, SPC-QC-104) for reproducibility and testing purposes.

The raw SPC file can be downloaded from:

https://drive.google.com/file/d/1cnEcXbqJvVVJ8ZdCexfEFmzAmhLR2nbM/view?usp=drive_link

After downloading, place the .SPC file into:

./spc2tiff/ExampleData/

Then run the MATLAB script:

./spc2tiff/spc2tiff.m

This script converts the raw SPC photon stream into a sequence of TIFF frames that can be directly used for EFLIM training and inference.

The converted TIFF files will be saved to:

./spc2tiff/output/

2. Python training and inference

Once the dataset is prepared, you can train and evaluate EFLIM using the provided Python code.

Requirements

• Python ≥ 3.9
• GPU support (CUDA-enabled GPU recommended)
• Additional Python packages:
  • numpy
  • scipy
  • tifffile
  • tqdm
  • matplotlib

Create a new conda environment (recommended):

conda create -n eflim python=3.10 -y
conda activate eflim

GPU support

To run EFLIM efficiently on a GPU, make sure you have a working CUDA toolkit installed. The recommended way is to install PyTorch together with the matching CUDA version directly from the official PyTorch website.

For example, on a machine with CUDA 11.8, you can install PyTorch with:

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118

Install dependencies:

pip install -r requirements.txt

Training example:

python run_EFLIM.py \
  --folderName .//simu_USAF1951_PPP0.5//raw

If you want to train on a specific GPU (e.g., GPU 2):

CUDA_VISIBLE_DEVICES=2 \
python run_EFLIM.py \
  --folderName .//simu_USAF1951_PPP0.5//raw

This script will:

• Load the raw data
• Perform training and inference for both lifetime and intensity
• Output a lifetime video, an intensity video and an intensity-weighted lifetime visualization