Generalizable Yeast Cell Segmentation
StarDist + HoVerNet fine-tuned for yeast microscopy. F1 = 0.96 on the held-out test set. Ships training scripts and a `predict.py` CLI.
What this is
Yeast cell segmentation in microscopy images. Two pre-trained backbones fine-tuned for this domain, plus a small classification head to label each segmented instance.
How it works
StarDist and HoVerNet run in parallel because they fail in different ways:
- StarDist: star-convex polygon predictor. Strong when cells are round and well-separated.
- HoVerNet: horizontal/vertical distance maps for instance segmentation. Better in clusters and on shapes that aren’t star-convex.
Two interchangeable training scripts: train_stardist.py and train_hover.py. Inference is a predict.py CLI: pick a backend, point at images, get masks back.
Results
F1 = 0.96 on the held-out test set. Trained weights and a public yeast dataset are linked from the repo. The actual training data is a private microscopy collection awaiting publication.
CS-433 Machine Learning at EPFL, team “3-packs”.