2026
Kaggle, Computer Vision, Deep Learning
Biohub3D Cell Tracking in Zebrafish
An end-to-end 3D nuclei detection and tracking pipeline for zebrafish light-sheet microscopy video, submitted as a fully offline Kaggle notebook.
In numbers
- 1,615 / 3,947
- final leaderboard, top 41%
- 0.639 → 0.791
- leaderboard score after moving to a 3D U-Net
- 199
- ground-truth graphs for offline validation
What I did
- 1.
Built an end-to-end 3D nuclei detection and tracking pipeline for zebrafish light-sheet microscopy video, submitted as a fully offline Kaggle notebook.
- 2.
Progressed from a classical CPU detection-and-linking pipeline to a trained 3D U-Net detector, lifting the public leaderboard score from 0.639 to 0.791.
- 3.
Re-implemented the competition's exact evaluation metric to validate offline against 199 ground-truth graphs, running systematic threshold and parameter sweeps to find the true operating point.
- 4.
Ran controlled ablations and documented approaches that didn't hold up under the real metric (e.g. a division-detection heuristic, gap-closing) rather than reporting only positive results.
Notes
Tracking cells through 3D space and time is two problems stacked: find every nucleus in each volume, then link them across frames without losing them when they divide or blur.
The biggest single lever was validation. Re-implementing the competition metric locally against 199 ground-truth graphs turned guesswork into measurable sweeps, and made it obvious which clever ideas (division heuristics, gap-closing) were actually hurting.
Built with
Python · PyTorch · 3D U-Net · scikit-image · NumPy
Next project
Kaggriculture →