Warning: Active development. Cross-validate results with other software before drawing conclusions. API and file formats may change without deprecation.
FLIMKit is a Python toolkit for FLIM data from FLIM microscope systems (or any PTU-based system). Built as a drop-in for FLIM microscope software, with two workflows:
- Reconvolution fitting: mono/bi/tri-exponential lifetime fitting with full IRF deconvolution, per-pixel and summed modes, multi-tile ROI stitching, batch processing, and optional time-varying background correction (FLIMfit-style, from a measured fluorophore-free reference)
- Lifetime distribution fitting: Gaussian and Lorentzian continuous α(τ) distributions (Lakowicz §4.11.2), per-ROI and per-pixel maps with GPU acceleration
- Phasor analysis: calibrated phasor plots, interactive elliptical cursors, spatial filtering (gaussian/median/wavelet), two-component decomposition, automatic peak detection, session save/load
Four entry points: desktop GUI, guided terminal UI, CLI scripts, Python API.
Examples repo | Full documentation
Python ≥ 3.12 required (3.14 recommended — official builds use 3.14).
git clone https://github.com/alex1075/FLIMKit.git
cd FLIMKit
python install.py # auto-detects GPU and installs the right backend
python validate_installation.py # 10 checks - all should passFor development work (PyInstaller + test dependencies):
python install.py --devOr download the compiled app from the Releases tab (no Python needed).
A pre-built image is available on Docker Hub. It runs the full desktop GUI in a browser via xpra (no installation needed on the client).
Pull and run:
docker run -d \
-p 14500:14500 \
-v /path/to/your/data:/data \
--name flimkit \
alex1075/flimkit:latestThen open http://localhost:14500 in your browser. PTU files and data should be placed in the folder you mount to /data - use the file dialog inside the app to navigate there.
TrueNAS SCALE (Custom App):
- Apps → Discover Apps → Custom App
- Paste the contents of
docker-compose.yamlfrom this repo - Edit the volume paths to match your pool (e.g.
/mnt/tank/microscopy:/data) - Deploy - TrueNAS will pull the image automatically
Build from source (required if you want to push your own changes):
python build_docker.py # builds linux/amd64, pushes to Docker HubRequires Docker Desktop with buildx. On Apple Silicon, buildx cross-compiles for linux/amd64 automatically.
python main.pyFive tabs: Single FOV Fit, Tile Stitch / Fit, Batch ROI Fit, Machine IRF Builder, Phasor Analysis. The right panel shows an FOV preview (intensity image + summed decay) and switches to the interactive phasor view when that tab is active.
python main.py --clipython fit_cli.py --ptu data.ptu --machine-irf machine_irf_default.npy --nexp 2
python phasor_cli.py --ptu data.ptu --irf irf.xlsxfrom flimkit.phasor_launcher import launch_phasor
state = launch_phasor('data.ptu', irf_path='irf.xlsx')
# With spatial phasor filtering
state = launch_phasor('data.ptu', irf_path='irf.xlsx',
phasor_filter='gaussian', filter_kwargs={'sigma': 1.5})FLIM data (PicoQuant .ptu, T3 mode):
- PicoHarp T3 (Leica FALCON / STELLARIS)
- HydraHarp v1/v2 T3
- TimeHarp 260 N / P T3
- MultiHarp / generic T3
Imaging PTUs are reconstructed into per-pixel decays from their scan line markers.
Read as a decay only (no image): image reconstruction needs scan line markers, so FLIMKit fits the decay but cannot rebuild a FLIM image from a PTU that has none. This covers single-spot, point, and FCS acquisitions (for example PicoQuant TimeHarp 260P point measurements), and any PTU exported without imaging markers.
IRF sources:
- Measured / scatter IRF from a
.ptu - Leica IRF from the exported
.xlsx(interpolated or analytical model) - PicoQuant SymPhoTime Check file (
.pck) histogram - A built machine IRF, or an analytical Gaussian IRF
A .pck IRF must come from the same instrument and TCSPC resolution as the data being fit.
Not supported:
- T2-mode PTUs. FLIMKit decodes T3 mode only (one TCSPC histogram per sync period). T2 records are raw global timestamps with no per-period decay, so a T2
.ptuwill not produce a meaningful decay. - Older PicoQuant formats (
.pt3,.ht3,.phu,.pt2) and non-PicoQuant TCSPC (Becker and Hickl.sdt/.spc). - Leica
.lifand proprietary LMSCOMPRESSED blocks. Export.ptufrom LAS X instead.
Before fitting, build a machine IRF for your system once and reuse it across sessions. You need matched .ptu + .xlsx pairs, 10-20 is a good number.
In the GUI, go to Machine IRF Builder, point it at your pairs folder, and save as machine_irf_default. From source this goes to flimkit/machine_irf/; compiled app saves to ~/.flimkit/machine_irf/.
from flimkit.FLIM.irf_tools import build_machine_irf_from_folder
build_machine_irf_from_folder(
folder="/path/to/pairs",
align_anchor="peak",
reducer="median",
save=True,
output_name="machine_irf_default",
)Per-pixel fitting uses a batched matrix solver that runs on GPU when a supported backend is detected. This applies to both single-FOV and tile-ROI pipelines. The same fit_per_pixel() function is used in both.
| Backend | Hardware | Notes |
|---|---|---|
| MLX | Apple Silicon (M1/M2/M3/M4) | Detected automatically |
| PyTorch MPS | Apple Silicon | Fallback if MLX not installed |
| PyTorch CUDA | NVIDIA | pip install torch --index-url https://download.pytorch.org/whl/cu126 |
| PyTorch ROCm | AMD | pip install torch --index-url https://download.pytorch.org/whl/rocm6.2 |
python install.py detects your hardware and installs the right backend automatically. No extra flags needed at runtime.
Limitations: --free-tau-perpixel mode uses batched Adam on GPU when a backend is available (n_exp ≥ 2). fit_summed (single global fit) is always CPU - it's fast enough not to matter.
Compiled app and GPU: The compiled app bundles whatever GPU libraries are installed on the build machine. A binary built on Apple Silicon will have MLX/MPS; one built on a CUDA machine will have CUDA. If you need GPU in the compiled app, build it yourself on the target hardware. See Compiled App.
Not strictly necessary, but useful after making code changes.
python install.py --dev # install test requirements first
cd flimkit_tests
python run_tests.py # all tests
python run_tests.py -c # with coverage report
python run_tests.py integration # integration tests only| Format | Description |
|---|---|
| PNG | Intensity and lifetime map images |
| OME-TIFF | Lossless export with metadata - opens in Fiji/ImageJ |
| GeoJSON | ROI geometries and stats - imports directly into QuPath |
| CSV | Fit summaries and per-ROI statistics |
| NPZ | Session files for restoring analysis state |
Done: single FOV fitting, tile stitching, batch ROI processing, phasor analysis, GUI, session restoration, compiled app, ROI analysis with QuPath export.
Up next: config persistence, stat histograms, auto-region detection, batch n-exp in GUI. Chemical validation and publication pending.
See ROADMAP.md for details.
Fitted lifetimes from FLIMKit will typically read slightly higher than FLIM microscope software for the same data. This is a consequence of IRF placement: FLIMKit anchors the IRF at the steepest-rise point of the leading edge, which differs from how FLIM microscope software places the IRF. The difference is systematic and reproducible across acquisitions.
Lifetime distribution fitting - Gaussian and Lorentzian α(τ) models:
Lakowicz, J.R. (2006). Principles of Fluorescence Spectroscopy (3rd ed.). Springer. §4.11.2, pp. 141-144.
Time-varying background correction - measured background decay B = V·b(t) + Z:
Görlitz, F. et al. (2017). Open Source High Content Analysis Utilizing Automated Fluorescence Lifetime Imaging Microscopy. J. Vis. Exp. (119), e55119. https://doi.org/10.3791/55119
PhasorPy: Gohlke, C. et al. Zenodo. https://doi.org/10.5281/zenodo.13862586
Tile stitching: Preibisch et al. (2009). Bioinformatics 25(11). https://doi.org/10.1093/bioinformatics/btp184
Cellpose-SAM: Pachitariu & Stringer (2025). bioRxiv. https://doi.org/10.1101/2025.04.28.651001
FLIMKit is designed, developed, and maintained by Alex Hunt. Anthropic's Claude AI was used as an assistant for parts of the GUI implementation, compiled app builds, code debugging, and Docker packaging; all scientific design, fitting/phasor methods, validation, and the overall architecture are the author's own work.
Alex Hunt: alexander.hunt@ed.ac.uk