Live imaging: lineages broken by identity swaps
A lab saw unexpectedly variable cell-cycle times in sister cells. Most of the variability came from tracking errors in crowded fields.
- Client
- Stem cell biology lab
- Data
- 72 h time-lapse, FUCCI reporter, 48 fields
- Timeline
- 3 weeks
- Outcome
- Asymmetric division finding validated
Anonymized composite case. Details are combined and altered to protect client confidentiality.
The problem
Automated tracking with default settings reported sister cells with very different cycle lengths, suggesting asymmetric division. The effect was strongest late in the movies, when colonies became dense.
What we found
- Out-of-the-box segmentation merged touching nuclei, causing identity swaps when cells separated.
- Manual annotation of 200 divisions showed 23% of lineage links were wrong in the final 24 hours.
How we rescued it
- Fine-tuned a Cellpose model on 150 annotated frames, which reduced merge errors.
- Re-tracked with ultrack and gap closing, then curated low-confidence links in napari.
- Reported tracking accuracy per time window and limited statistics to windows above 95% link accuracy.
The result
Lineage error fell from 23% to about 3%. A real but smaller asymmetry in G1 duration remained, linked to a reporter-defined subpopulation, and the lab confirmed it with an independent marker.
| Measure | Result |
|---|---|
| Lineage link errors before | 23% |
| After | ~3% |
| Validated finding | Smaller, real asymmetry |