SofiaKaps/FLIMngo_data

FL🦩Mngo Training Dataset: deep learning for fluorescence lifetime predictions

This dataset contains the training data used in FLIMngo, a YOLO-based deep learning network for predicting fluorescence lifetime values from raw TCSPC-FLIM data.

Dataset Summary

The dataset comprises:

• 1242 .h5 files, each storing:
  • Simulated TCSPC-FLIM data in the form of 3D tensors (t, x, y)
  • Ground truth lifetime maps as 2D tensors (x, y) with per-pixel fluorescence lifetimes (in nanoseconds)
• File organisation the dataset is split across 18 .tar files to aid distribution. Each tar file is approx. 4.5 GB with the whole dataset being 78 GB.

Source data

The fluorescence intensity images used as the basis for these simulations are from the Human Protein Atlas (HPA). Each intensity image includes four colour channels representing microtubules, endoplasmic reticulum, nucleus and the target protein.

A fluorescence decay curve was simulated for every pixel of the intensity images as described in the FLIMngo Git repository and publication.

Simulation parameters

• Laser repetition rate: 40 MHz
• Time bins: 256, at 0.0977 ns per bin (~25 ns full window)
• Photon counts: 10-2500 photons per pixel
• Normalisation: each decay curve has been normalized to have values between 0 and 1
• IRF: each image is convolved with one IRF drawn at random from a library of 12 experimental or simulated IRFs, with laser jitter applied per image.
• Lifetimes: up to four exponential components per pixel, corresponding to microtubules, proteins, nucleus and endoplasmic reticulum channels of the source images. Component ratios are set by Perlin noise.
• Noise: Poisson noise applied after scaling to the per-pixel photon count.

Usage

import h5py
import numpy as np
import torch

# path to an extracted .h5 file
file_path = '013f0f3a-bbad-11e8-b2ba-ac1f6b6435d0_0.h5'

# internal paths used in the .h5 files
img_internal_path = 'tpsf'
label_internal_path = 'tau_maps'

# load data
with h5py.File(file_path, 'r') as f:
    raw_data = np.asarray(f[img_internal_path])
    ground_truth = np.asarray(f[label_internal_path])

# convert to torch tensors
raw_data_tensor = torch.from_numpy(raw_data).permute(0, 2, 1).float() 
ground_truth_tensor = torch.from_numpy(ground_truth).permute(1, 0).float()

Licence

Released under CC-BY-4.0. The FLIMngo source code is separately licensed under MIT.

Citation

If you found FLIMngo helpful, please consider citing our work! 😊

@article{kapsiani2025deep,
  title={Deep learning for fluorescence lifetime predictions enables high-throughput in vivo imaging},
  author={Kapsiani, Sofia and Läubli, Nino F and Ward, Edward N and Fernandez-Villegas, Ana and Mazumder, Bismoy and Kaminski, Clemens F and Kaminski Schierle, Gabriele S},
  journal={Journal of the American Chemical Society},
  year = {2025},
  publisher={ACS Publications}
}