live-foci 1.0.0

LiFT is an end-to-end pipeline for detecting and tracking DNA-damage foci (e.g. γH2AX, 53BP1) in live-cell fluorescence microscopy timelapse data.

LiFT – Live Foci Tracking

LiFT is an end-to-end pipeline for detecting and tracking DNA-damage foci (e.g. γH2AX, 53BP1) in live-cell fluorescence microscopy timelapse data. Starting from raw multi-frame TIFF stacks, it segments and tracks nuclei, crops each cell into its own image stack, corrects for motion, detects foci frame-by-frame, and tracks individual foci over time to quantify repair kinetics.

---

Pipeline overview

The pipeline runs in seven sequential steps:

Step	Name	Description
1	Nuclei segmentation	Detect and segment nuclei in every frame using Cellpose-SAM or Cellpose-V3
2	Nuclei tracking	Link segmented nuclei across frames to build single-cell trajectories
3	Cell cropping	Crop each tracked nucleus into a separate image stack
4	Registration	Correct for translational and rotation within each cropped cell stack
5	Foci detection	Detect DNA-damage foci in the registered stacks
6	Foci tracking	Track individual foci over time to quantify repair kinetics

---

Five ways to run LiFT

1. Interactive GUI — LiFT_app.py

A browser-based interface. Each step has a dedicated page where you can tune parameters, run a preview on a randomly selected cell, and then launch the full run on all data with a progress bar.

python LiFT_app.py

Then open http://localhost:5000 in your browser. Work through steps 1–6 in order. When you click Run all for any step, the current parameters are automatically saved to parameters.yml inside your data folder.

---

2. Jupyter notebook — LiFT_notebook.ipynb

The same pipeline exposed as a step-by-step notebook. Useful for exploratory analysis, custom preprocessing, or integrating the pipeline into a broader workflow. Each section contains a user-options cell where you set parameters, a preview cell to visualise results on a subset, and a run-all cell that saves parameters to parameters.yml and processes the full dataset.

jupyter notebook LiFT_notebook.ipynb

---

3. Command-line runner — LiFT.py

For this method, you can simply clone the entire repository and execute from that local directory.

Reads the parameters.yml written by the GUI or notebook and re-runs the full pipeline on a new data folder with identical settings. More experienced users can also write a custom parameters.yml with their desired settings. Intended for batch reproduction and HPC submission.

# Run all steps
python LiFT.py data/

# Run specific steps only
python LiFT.py data/ --steps 1 2 5 6

Programmatic use:

import LiFT

LiFT.run("data/")
# or specific steps only by adding: steps=[1, 2, 5, 6]

---

4. Python Package

LiFT is published on PyPI as live-foci — install it and it's available in your codebase as lift.

pip install live-foci



# with an optional extra, e.g. GPU segmentation

pip install "live-foci[cp-sam]"

Confirm it installed correctly:

lift info

Every pipeline step is available as a Python function. Parameters can come from parameters.yml or be passed explicitly — explicit arguments always win over config.

import lift

 

# run the full pipeline from config

lift.run("data/experiment_name")

lift.run("data/experiment_name", steps=[1, 2, 5, 6])

 

# or call individual steps

lift.segment("data/experiment_name")

lift.segment("data/experiment_name", method="cellpose_v3", diameter=120, min_area=500)

 

lift.track_nuclei("data/experiment_name")

lift.track_nuclei("data/experiment_name", method="IOU", min_length=15)

 

lift.crop("data/experiment_name", margin=50)

 

lift.register("data/experiment_name")

lift.register("data/experiment_name", method="stackreg")

 

lift.detect("data/experiment_name")

lift.detect("data/experiment_name", method="TopHat", threshold=36, sigma=0.6, radius=3.0)

 

lift.track_foci("data/experiment_name")

lift.track_foci("data/experiment_name", method="GNN", max_distance=8.0, gap_closing=3)

---

CLI reference

lift info                                         # show installed optional packages

lift init [data_path]                             # generate parameters.yml (default: cwd)

lift config [data_path] [key.path=value ...]      # interactive editor or set values directly

lift run <data_path> [--steps N ...]              # run pipeline

lift <data_path> [--steps N ...]                  # shorthand for lift run

lift init

Probes your environment for installed optional packages and writes a parameters.yml with the best available defaults, also checks currently installed packages and inserts these as options:

lift init                          # writes to current directory

lift init data/experiment_name     # writes to specified path

lift config

With no arguments, opens an interactive editor that walks through every parameter step by step, showing the current value and available options. Press Enter to keep a value unchanged, or type a new one. Ctrl+C saves what has been changed so far and exits.

lift config                        # interactive editor in current directory

lift config data/experiment_name   # interactive editor at path

To set values directly without the interactive editor:

lift config step5_detection.threshold=40

lift config step5_detection.method=TopHat step5_detection.params.radius=3.0

lift config step1_segmentation.segmentation.min_area=500

lift config step4_registration.preprocessing=null

lift config data/experiment_name step5_detection.threshold=40

With no key=value arguments, prints the current config.

---

5. Docker Image

Docker

LiFT ships two Dockerfiles: Dockerfile.base (the heavy layer — OS packages + live-foci from PyPI, published to Docker Hub) and Dockerfile (a thin layer on top that adds whatever EXTRAS you want). This means Docker Hub only ever hosts two images — cpu-base and gpu-base — and you can build any combination of extras locally in seconds, since the heavy layer is already pulled and cached. There's no fixed set of presets to choose from.

Pull the base image

docker pull <dockerhub-user>/lift:cpu-base   # python:3.11-slim, no GPU needed

docker pull <dockerhub-user>/lift:gpu-base   # CUDA runtime, requires NVIDIA container runtime



# pin to a specific released version instead of the rolling tag

docker pull <dockerhub-user>/lift:0.1.0-cpu-base

The base image alone runs, but has no segmentation/tracking extras installed — you'll want to add extras next.

Build your combination locally

This is the step that replaces "pick a preset" — any comma-separated combination from pyproject.toml's [project.optional-dependencies] (cp-sam, cp-v3, trackastra, spotiflow, notebook, app) works:

# CPU segmentation only

docker build --build-arg EXTRAS="cp-v3" -t lift-cpu .



# GPU segmentation + trackastra, no spotiflow — not something a fixed preset could offer

docker build \

  --build-arg BASE=<dockerhub-user>/lift:gpu-base \

  --build-arg EXTRAS="cp-sam,trackastra" \

  -t lift-gpu .



# everything available in Docker (elastix/NGMA/NND are Windows-only, excluded — see Platform support above)

docker build \

  --build-arg BASE=<dockerhub-user>/lift:gpu-base \

  --build-arg EXTRAS="cp-sam,trackastra,spotiflow" \

  -t lift-full .

This is fast — Docker pulls <dockerhub-user>/lift:cpu-base/gpu-base once, caches it, and each build after that just adds a single pip install layer on top.

Running

Mount your data folder into /data and pass the same commands as the CLI:

# generate config

docker run --rm -v $(pwd)/data:/data lift-cpu init /data/experiment_name



# run all steps

docker run --rm -v $(pwd)/data:/data lift-cpu run /data/experiment_name



# run specific steps

docker run --rm -v $(pwd)/data:/data lift-cpu run /data/experiment_name --steps 5 6



# GPU run

docker run --rm --gpus all -v $(pwd)/data:/data lift-gpu run /data/experiment_name



# check installed packages

docker run --rm lift-cpu info

On Windows use %cd% instead of $(pwd):

docker run --rm -v %cd%/data:/data lift-cpu run /data/experiment_name

Using docker compose

# build your combination (any BASE + EXTRAS you want)

BASE=<dockerhub-user>/lift:gpu-base EXTRAS=cp-sam,trackastra docker compose --profile custom build custom

docker run --rm --gpus all -v $(pwd)/data:/data lift-custom run /data/experiment_name



# maintainer-only: rebuild the images that get pushed to Docker Hub

docker compose --profile build-base build

Notes

NVIDIA runtime is required for GPU images. Install NVIDIA Container Toolkit and ensure nvidia-smi works inside a container before using GPU builds.

Image sizes: the GPU base is large (~5–8 GB) due to the CUDA base and torch. The CPU base is around 1–2 GB. Extras add a small amount on top since the base is cached.

---

Parameter reproducibility

Every time you click Run all in the GUI or execute the run-all cell in the notebook, the full parameter set is written to <data_folder>/parameters.yml. This file records all settings for every step that has been run, grouped by step:

step1_segmentation:
  segmentation:
    method: cellpose_sam
    flow_threshold: 0.0
    cellprob_threshold: -0.5
    min_area: 1000
    cellpose_sam:
      scale_factor: 0.5
    cellpose_v3:
      diameter: 140.0
  preprocessing:
    method: None
    wavelet_filtering:
      scales: 5
      w_factor: 1.5
      start_scale: 2
    contrast_adjuster:
      sigma: 3.0
      c_factor: 6.0

:
:
:

step5_detection:
  method: TopHat
  threshold: 36.0
  return_segmentation: false
  params:
    sigma: 0.6
    radius: 6.0
# ... steps 2–4 and 6 follow the same pattern

Experienced users can also write a minimal parameters.yml by hand — any sub-keys that are absent fall back to function defaults, so only the settings that differ from defaults need to be specified.

---

Supported algorithms

Nuclei segmentation (Step 1)

Method	Notes
cellpose_sam	Cellpose with SAM backbone; scale-factor controls effective nucleus size
cellpose_v3	Cellpose V3 with diameter parameter

Optional preprocessing before segmentation to transform the foci signal into a signal resembling a nuclear (DAPI like) staining:

• wavelet_filtering — multi-scale wavelet coefficient thresholding
• contrast_adjuster — Gaussian-based contrast adjustment

Nuclei tracking (Step 2)

Method	Notes
IOU	Intersection-over-union based linking
NND	Greedy Nearest Neighbour Diffusion
trackastra	deep-learning tracker (Trackastra)

Registration (Step 4)

Method	Notes
stackreg	StackReg rigid translation (PyStackReg)
elastix	Elastix; supports MSE, MI, and NCC loss functions

Optional preprocessing for registration: wavelet_denoise, threshold, DOG_filter.

Foci detection (Step 5)

Method	Notes
TopHat	top-hat morphological filter
LOG	Laplacian of Gaussian
Hessian	Hessian-based blob detector
Wavelets	Isotropic undecimated wavelet transform with thresholding of the coefficients
HDome	H-dome transform
HDome-smal	H-dome with LOG pre-filter
MPHD	Maximum Possible Height Dome
Spotiflow	deep-learning spot detector (Spotiflow)

Foci tracking (Step 6)

Method	Notes
GNN	Global Nearest Neighbour
NGMA	Non-iterative Greedy Multi-frame Assignment
trackastra	deep-learning tracker

---

Expected data structure

LiFT expects input data organised as:

data_folder/
  condition/
    experiment_date/
      Pos001/
        raw/ 
          0000.tif
          0001.tif   
          0002.tif  
          :       
      Pos002/
        raw/
      ...
  parameters.yml          ← written automatically; can be pre-populated

Intermediate and final outputs are written alongside the input inside each Pos*/results/ folder.

---

Installation

Note: Detailed installation instructions will be added here.

---

Citation

If you use LiFT in your research, please cite:

Citation will be added here upon publication.