Skip to content

Getting started

Install

pip install patchworks

Python 3.11 or later. Add what you need:

Extra For
cellpose (cellpose3, cellpose4 to pin) the Cellpose plugin
bioio, imaris converting CZI, LIF, ND2, TIFF, .ims to OME-Zarr
napari viewing and reviewing results
review object tables as pandas, Excel export
dog deconvolution before the difference-of-Gaussians plugin
careamics Noise2Void denoising
workflow the cluster workflow (Snakemake)
gpu, distributed, remote GPU memory sizing; Dask clusters; S3/GCS/HTTP stores
all everything 
pip install "patchworks[cellpose,bioio,napari]"

GPU options of the DoG plugin and of label dilation need cupy for your CUDA version (pip install cupy-cuda12x). PlantSeg is on conda-forge only (conda install -c conda-forge plant-seg).

Your first run

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

to_ome_zarr("scan.czi", "scan.zarr")  # once: a pyramidal OME-Zarr

tile_process(
    "scan.zarr",
    cellpose_fn("cyto3", gpu=True, diameter=30),
    channel=0,
    tile_shape=(1, 2048, 2048),  # or "auto", sized to your memory
    overlap=20,                  # halo: about one object diameter
)

The labels are written into scan.zarr/labels/labels, next to the image. Look at them:

from patchworks.plugins.napari import view_in_napari

view_in_napari("scan.zarr")

Any function

fn is any callable from a tile (a NumPy array) to integer labels of the same shape:

from skimage.filters import threshold_otsu
from skimage.measure import label


def my_fn(tile):
    return label(tile > threshold_otsu(tile)).astype("int32")


tile_process("scan.zarr", my_fn)

Each tile is segmented independently; objects crossing tile boundaries are joined into one label afterwards.

Next