patchworks

Tiled processing of arbitrarily large images — any image, any function.

┌──────┬──────┬──────┐     fn(tile) → labels      ┌──────┬──────┬──────┐
│ tile │ tile │ tile │  ─────────────────────►    │  1   │  2   │  3   │
├──────┼──────┼──────┤                            ├──────┼──────┼──────┤
│ tile │ tile │ tile │                            │  4   │  5   │  6   │   globally
├──────┼──────┼──────┤                            ├──────┼──────┼──────┤   consistent
│ tile │ tile │ tile │                            │  7   │  8   │  9   │   labels
└──────┴──────┴──────┘                            └──────┴──────┴──────┘

patchworks splits an image too large for memory into tiles, runs any segmentation function on each tile, and stitches the results into one consistent label image — stored with the image in a single OME-Zarr, with a table of every object.

[!NOTE] On how this was written. Large parts of patchworks were vibe coded — written with heavy LLM assistance rather than line by line. It is covered by a test suite and has been run on real data, so it is not untested, but the usual caveats apply: read the code before you trust it with anything irreplaceable, and please open an issue if something looks off.

Install

pip install patchworks                 # core
pip install "patchworks[cellpose]"     # + Cellpose
pip install "patchworks[bioio,imaris]" # + converting CZI/LIF/ND2/TIFF/.ims to OME-Zarr
pip install "patchworks[napari]"       # + viewing and reviewing in napari
pip install "patchworks[all]"          # everything above and more

GPU options of the DoG plugin need cupy matching your CUDA version (pip install cupy-cuda12x). PlantSeg is on conda-forge only. See Getting started for every extra.

Three ways to use it

Python — any function from a tile to labels:

from patchworks import tile_process
from patchworks.plugins.cellpose import cellpose_fn

tile_process("scan.zarr", cellpose_fn("cyto3", gpu=True, diameter=30))

The labels go into scan.zarr/labels/labels, as a multiscale pyramid next to the image.

Command line — the same building blocks, no script:

patchworks convert scan.czi scan.zarr
patchworks segment scan.zarr --method cellpose --model cyto3 --diameter 30 --gpu
patchworks view scan.zarr

On a cluster — a Snakemake workflow with a pixi environment: convert, segment tile batches as GPU jobs, merge, relate label images (which cell each nucleus is in), and bundle, from one YAML config per segmentation:

cd workflow
pixi run multi-slurm --config my_multi.yaml

What you get

• One OME-Zarr holding the image and every label image, readable by napari, Fiji and any OME-Zarr tool.
• Object tables — size, position, shape — and relations between label images (the cell of each nucleus, the position of each cilium in its cell), exported as spreadsheets.
• Review in napari: patchworks review scan.zarr steps through the objects most likely to be wrong; corrections flow into the tables.
• Plugins for Cellpose, PlantSeg, a nuclei-seeded watershed, difference of Gaussians (cilia, spots) and Noise2Void denoising — or bring your own function.

Documentation

https://imcf.one/patchworks/ — getting started, the cluster workflow, guides, examples and the API reference.

License

GNU General Public License v3.0 (GPL-3.0). See LICENSE.