Cell tracking in live-cell imaging: segmentation, linking, lineages and validation
How to track cells in time-lapse microscopy: Cellpose segmentation, TrackMate and ultrack linking, lineage trees, accuracy checks, common errors.
Key takeaways
- Tracking quality is limited by segmentation quality. Fix segmentation first, often by fine-tuning on your images.
- Identity swaps in dense regions silently corrupt lineages and timing statistics.
- Validate tracks against manual annotation and report accuracy before computing biology.
- Analyze single-cell distributions; population averages hide subpopulations.
Live-cell imaging measures behavior over time: how fast cells move, when they divide, how reporter signals pulse, what fates daughters choose. Every one of those measurements depends on following the same cell from frame to frame. Tracking errors do not crash the analysis. They change the biology quietly.
Step 1: Acquisition decisions that make tracking possible
- Frame interval: cells should move less than about half a diameter between frames.
- Nuclear marker: a nuclear reporter (H2B-FP, SiR-DNA) makes segmentation and tracking far more robust than brightfield alone.
- Phototoxicity: more light means better images and sicker cells. Watch for exposure-dependent slowing of division and migration.
- Field density: plan for confluence late in the movie, which is where tracking usually breaks.
Step 2: Segmentation
Deep-learning segmentation (Cellpose, StarDist) handles variable morphology, low contrast and touching nuclei far better than thresholding. Generalist models are a strong start. Fine-tuning on 50 to 200 annotated frames from your own data often removes most merge and split errors, and that pays off directly in tracking quality.
Step 3: Linking
- LAP tracking (TrackMate, after Jaqaman and colleagues): frame-to-frame linking plus gap closing, splitting (divisions) and merging, solved as an assignment problem.
- ultrack: considers multiple segmentation hypotheses jointly with tracking, which helps when segmentation is ambiguous.
- btrack: Bayesian tracking with motion models, suited to dense fields.
- Single-particle tracking for vesicles, receptors or chromatin loci, with mean-squared-displacement analysis of motion types.
Step 4: Divisions and lineages
Division detection turns tracks into lineage trees: mother, daughters, granddaughters. Lineages support cell-cycle timing (with FUCCI reporters), sister-cell correlations and fate decisions. They are also the most fragile output, because one swapped identity rewires the tree.
What our experts avoid: computing cell-cycle or sister-cell statistics without measuring tracking accuracy. In one rescued project, 23% of lineage links were wrong in the last day of imaging. That error created an apparent asymmetry in sister-cell cycle times.
Step 5: Validate, then curate
- Manually annotate a representative subset: early and late frames, sparse and dense regions.
- Compute link accuracy, division detection precision and recall, and identity-switch rates.
- Curate low-confidence links in napari.
- Restrict statistics to time windows and regions where accuracy meets a threshold you set in advance, for example 95% of links correct.
Step 6: Measurements
- Migration: speed, persistence, directionality, mean squared displacement and chemotactic index.
- Cell cycle: phase durations from FUCCI, inter-division times and their inheritance.
- Reporter dynamics: single-cell intensity time series, peak detection, pulsing frequency and oscillation analysis.
- Morphodynamics: shape changes and protrusion activity over time.
- Events: death, differentiation and other transitions, classified from features or learned models.
Always look at distributions and subpopulations. A population-mean migration speed can hide a fast and a slow subpopulation that respond differently to treatment.
How OmicsQuest can help
Our live-cell imaging and tracking service delivers segmentation and tracking with reported accuracy, curated lineages, annotated movies and statistics, for 2D, 3D light-sheet and organoid data.
Frequently asked questions
What is the best software for cell tracking?
For most 2D and 3D fluorescence data we combine deep-learning segmentation (Cellpose or StarDist) with TrackMate's LAP tracker, ultrack or btrack for linking, and napari for curation. The best choice depends on density, frame rate, divisions and dimensionality, so we test on a manually annotated subset.
How often should I image cells for tracking?
Often enough that cells move less than about half a cell diameter between frames, and divisions are captured in several frames. For typical mammalian cells that is every 5 to 15 minutes, balanced against phototoxicity.
Work with OmicsQuest. Related services:
References
- Stringer C, Wang T, Michaelos M, Pachitariu M. Cellpose: a generalist algorithm for cellular segmentation. Nat Methods. 2021;18:100-106.
- Ershov D, Phan MS, Pylvanainen JW, et al. TrackMate 7: integrating state-of-the-art segmentation algorithms into tracking pipelines. Nat Methods. 2022;19:829-832.
- Jaqaman K, Loerke D, Mettlen M, et al. Robust single-particle tracking in live-cell time-lapse sequences. Nat Methods. 2008;5:695-702.