Skip to content

Grow regions from seeds

Use numbered seed objects to divide a foreground region into separate objects. For example, nuclei can seed cell regions in an actin image. The seeds supply object identity; the image guides boundaries; the mask limits where growth is allowed.

New in 0.16.0a1

Grow Regions from Seeds — CellProfiler Propagation is a new CPU node for one aligned YX plane. Marker-Controlled Watershed is already available for 2D and 3D data.

Choose the method

Need Method
The 2D CellProfiler Propagation rule, using appearance and distance from seeds Grow Regions from Seeds — CellProfiler Propagation
Seeded flooding of an elevation image, including a full ZYX volume Marker-Controlled Watershed
Growth by geometric distance without an intensity image Expand Labels

These methods can produce different boundaries from the same seeds. Choose the method before tuning parameters, and validate the result on held-out images.

Grow 2D regions with CellProfiler Propagation

  1. Prepare one scalar YX image. For multichannel or volumetric data, use Extract Channel and Select Axis Slice to select the intended channel, time point and plane. Confirm that the result has exactly two axes, Y and X.
  2. Create a labels image for the seeds, with zero as background and a distinct positive integer for each object. Nucleus labels are one possible source.
  3. Create a Boolean foreground mask with a threshold or other reviewed segmentation. Inspect it: growth cannot cross excluded pixels.
  4. Add Grow Regions from Seeds — CellProfiler Propagation. Connect the intensity image to Guidance image, labels to Seed labels, and the mask to Foreground mask. All three must share the same shape, axes and physical grid.
  5. Start with Distance regularization at 0.05, then click Calculate. Higher values give pixel distance more influence relative to image appearance. Compare the output over the input image using label boundaries.
  6. Save the chosen preprocessing, mask and regularization in the workflow. Apply those frozen settings to held-out fields before using the labels for measurements.

The node keeps seed IDs and returns an int32 labels image. A seed outside the mask remains in the output but does not start growth. A disconnected masked region with no reachable seed stays zero. Inspect both cases when labels seem missing or extend beyond the foreground.

Intensity scale matters. The node does not normalize the image. A setting of 0.05 on raw uint16 intensities is not equivalent to 0.05 on an image scaled to 0–1. Make any rescaling an explicit upstream step and save its settings. The node uses manual/cached execution; recalculate it after changing an input or parameter.

For the statistics paper's complete nuclei/cell/cytoplasm segmentation, follow Reproduce the statistics-paper compartments. It adds the CellProfiler-specific smoothing, seed exclusions, hole filling and cytoplasm definition around this growth step.

Use watershed on a 3D volume

  1. Prepare aligned ZYX image, marker-label and Boolean mask volumes. Select a channel and time point explicitly if the source also contains C or T.
  2. Connect all three to Marker-Controlled Watershed and select 3D ZYX for its spatial mode. Check the input-axis interpretation before running.
  3. Choose the elevation image deliberately. For separating touching objects in a mask, use Euclidean Distance Transform and H-Maxima Markers, then set Image meaning to Distance map (invert). Use Elevation image for a supplied elevation whose low values should form basins.
  4. Inspect the resulting labels through Z and in orthogonal views. Compare object identity and split/merge errors, not only a maximum projection.

The repository's seeded-3d-watershed.json provides a runnable graph, initially bound to the bundled synthetic volume. Download and open that workflow, then replace its image source with a reviewed ZYX volume to try the same steps. Its saved parameters are a validation example, not established settings for every mitochondrial acquisition.

The existing watershed uses voxel adjacency, and the existing distance transform uses voxel-index distances. They do not use physical voxel spacing to adjust those calculations. Calibration is retained for later measurements, but unequal Z and XY spacing can still affect the segmentation. Sobel Edges operates per YX plane and should not be described as a volumetric gradient.

For exact input rules and the boundary of CellProfiler compatibility, see the seeded segmentation reference. The validation status separates acquired-volume backend checks from biological validation.