Cytoland

Robust virtual staining of landmark organelles from label-free microscopy.

Part of the VisCy monorepo.

Paper: Robust virtual staining in Nature Machine Intelligence

Installation

# From the VisCy monorepo root
uv pip install -e "applications/cytoland"

Usage

Training and prediction use the shared viscy CLI provided by viscy-utils:

# Training (pick a model-specific config)
uv run --package cytoland viscy fit -c examples/configs/vscyto3d/finetune.yml

# Training with Spotlight loss
uv run --package cytoland viscy fit -c examples/configs/vscyto3d/train_spotlight.yml

# Prediction
uv run --package cytoland viscy predict -c examples/configs/vscyto3d/predict.yml

The YAML config determines which model and data module to use via class_path:

model:
  class_path: cytoland.engine.VSUNet
data:
  class_path: viscy_data.hcs.HCSDataModule

Tutorials and demos

Scripts and tutorials live under examples/:

All demo scripts are written as jupytext-style percent-cell .py files. Regenerate paired .ipynb notebooks with jupytext --to ipynb solution.py if you prefer the notebook UI.

Models

FNet3D note: All spatial dimensions (Z, Y, X) must be divisible by 2^depth (default depth=4 requires divisibility by 16). See examples/configs/fnet3d/fit.yml.

Benchmark note: FNet3D and SEC61B benchmarks now launch from applications/dynacell/. Cytoland copies are transitional legacy — see examples/configs/dynacell/ and examples/configs/fnet3d/.

References

@article{liu2025robust,
    title = {Robust virtual staining of landmark organelles},
    author = {Liu, Ziwen and Hirata-Miyasaki, Eduardo and Pradeep, Soorya and others},
    journal = {Nature Machine Intelligence},
    year = {2025},
    doi = {10.1038/s42256-025-01046-2},
}