Skip to content

Cells from a membrane stain

With a membrane marker and a nuclear dye, Cellpose 4 (cpsam) tends to segment the bright nuclei, and on the membrane alone a faint wall merges two cells. patchworks has plugins for this, all ordinary custom functions:

Method Cells come from Best when
Nuclei-seeded watershed the membrane, flooded from the nuclei every cell has one nucleus: start here
PlantSeg a boundary U-Net, then GASP / multicut, optionally seeded by nuclei the membrane is uneven or noisy
Cellpose, membrane only Cellpose in 3-D you want to stay with Cellpose

Any of them can run on denoised tiles.

Watershed and PlantSeg give every voxel to a cell, so neighbours touch at every tile seam: they need stitch: "iou" (the workflow refuses the config without it; on the command line, pass --stitch iou).

Nuclei-seeded watershed

Nuclei are easy to find; flooding the membrane from them grows each nucleus to its cell's walls, one cell per nucleus, even across a gap in a wall.

channel: 0              # membrane
nuclei_channel: 1       # each tile becomes [membrane, nuclei]
method: "custom"
label_name: "cyto_labels"
stitch: "iou"
custom:
  module: "patchworks.plugins.watershed"
  kwargs:
    nuclei_min_size: 200    # voxels
    foreground: "otsu"      # stop at the tissue edge
    # max_radius_um: 15     # or: at most this far from the nucleus
    # nuclei_threshold: 800 # if Otsu misses dim nuclei, or two touching nuclei give one seed

The halo (overlap) must hold a whole cell.

PlantSeg

PlantSeg predicts cell boundaries with a U-Net trained on membrane stains and partitions them into cells. It is on conda-forge only, so it has its own environment (on a login node without a GPU, set CONDA_OVERRIDE_CUDA first, see Set up):

pixi install -e plantseg
pixi run -e plantseg fetch my_config.yaml   # its models, once, with internet (the login node)

method: "custom"
stitch: "iou"
custom:
  module: "patchworks.plugins.plantseg"
  kwargs:
    model: "generic_confocal_3D_unet"   # or generic_light_sheet_3D_unet, ...
    segmentation: "gasp"                # gasp | mutex_ws | multicut | dt_watershed
    beta: 0.6                           # lower merges more, higher splits more
    foreground: "otsu"

With nuclei, segmentation: "nuclei_watershed" floods the U-Net's boundaries from them, and "lifted_multicut" keeps supervoxels of one nucleus together. Tiles are resampled to the model's training voxel size automatically.

Cellpose nuclei, then PlantSeg cells

The workflow ships this pairing: Cellpose segments the nuclei, PlantSeg grows one cell from each, and the two are related. Set input and work_dir in config/common.yaml, then:

pixi run -e plantseg multi-plantseg-dry
pixi run -e plantseg multi-plantseg-slurm

config/multi_plantseg.yaml lists config_nuclei.yaml and config_cyto_plantseg.yaml; the second one seeds the cells with the first's labels:

seed_labels: "nuclei_labels"        # each tile becomes [membrane, nuclei labels]
custom:
  module: "patchworks.plugins.plantseg"
  kwargs:
    segmentation: "nuclei_watershed"
  • The cells start once the nuclei are done; other segmentations run alongside. Run your own multi config with PlantSeg from the same environment: pixi run -e plantseg multi-slurm --config my_multi.yaml.
  • Mistakes in the nuclei carry over to the cells. Correct them first with patchworks review, write them with --write-labels nuclei_labels, and seed from nuclei_labels_reviewed.
  • Re-segmenting the nuclei does not re-run the cells: delete <work_dir>/cyto_labels and image.zarr/labels/cyto_labels first.
  • Without PlantSeg, the same pairing works with module: "patchworks.plugins.watershed" in the default environment.

Cellpose, membrane only

Give Cellpose only the membrane (nuclei_channel: null) and set do_3D: true: it combines three orthogonal views, so a wall faint in one plane is found in the others.

Denoising first

Noise2Void learns to remove noise from the image itself, without clean ground truth. Train once per channel (on a GPU node):

pixi install -e careamics
pixi run -e careamics denoise-train <work_dir>/image.zarr --channel 0 --out n2v_membrane.ckpt

then add to any segmentation config, and run from the careamics environment (plantseg-careamics for both):

denoise:
  model: "/path/to/n2v_membrane.ckpt"
  # nuclei_model: "/path/to/n2v_nuclei.ckpt"   # for nuclei_channel

Only the segmentation sees the denoised tiles; nothing extra is stored. Command line: patchworks segment ... --denoise n2v_membrane.ckpt. Try it on one tile with patchworks.plugins.careamics.denoise(tile, model=...).