Yusuke Uchida's Cell Tracker Scores 0.9762 Across 199 Embryo Videos
New arXiv paper proposes a 3D cell tracking pipeline designed for sparse annotations: arxiv.org/abs/2611.06624 . Combines 2.5D U-Net and 3D SegResNet detectors with loss masking around likely unannotated cells.
Takeaways
- New arXiv paper proposes a 3D cell tracking pipeline designed for sparse annotations: arxiv.org/abs/2611.06624 .
- Combines 2.5D U-Net and 3D SegResNet detectors with loss masking around likely unannotated cells.
- Scores candidate edges using dense flow, learned cell matching, and dedicated division classifiers.
- Jointly optimizes links, divisions, and node inclusion using a metric-aligned surrogate objective.
- Reaches 0.9762 in 5-fold CV on 199 zebrafish embryo videos; joint selection beats sequential by 0.0105.
- Code available at github.com/yu4u/kaggle-biohub-3rd .
Tracking cells through a 3D time-lapse embryo requires detecting each nucleus, linking it across frames, and recording every division. Sparse annotations make that reconstruction harder because many real cells appear in the images without labels.
Yusuke Uchida addresses that constraint in the paper Learning to Track Cells from Sparse Annotations in 3D Time-Lapse Microscopy . Developed for the CZ Biohub Cell Tracking During Development competition on Kaggle, the method combines sparse-label detection, learned correspondence scores, explicit division models, and global lineage optimization. A public implementation covers training, inference, and evaluation.
Sparse labels create false negatives
A 3D light-sheet microscopy sequence can contain tens of thousands of nuclei moving and dividing over time. Exhaustive annotation is prohibitively expensive, so researchers label a subset. Standard supervised training may interpret every unlabeled location as background, teaching a detector to suppress genuine cells that annotators omitted.
Producing a complete lineage graph requires the pipeline to solve three connected tasks:
- Detection: locate every cell center in each 3D volume.
- Linking: connect the same cell across consecutive frames.
- Division identification: connect one parent to its two daughter cells.
Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.
Read original report
