Segmentation and mask review

Use Omnipose to turn phase-contrast images into labeled cell masks. If masks already exist, skip this stage and load them through create_image_objects().

from bactoscoop import ImageCollection
ic = ImageCollection(r"C:\path\to\working_copy", px=0.065)
ic.load_phase_images(phase_channel="C1")
ok = ic.segment_images(
    mask_thresh=1, minsize=100, n=range(len(ic.images)),
    model_name="bact_phase_omni",
)
if not ok:
    raise RuntimeError(ic.processing_summary())
ic.create_image_objects(phase_channel="C1")

Selection and thresholds

n selects image positions; None selects all loaded images in the wrapper. Selected indices must be unique, valid integers. An empty selection is rejected. The wrapper saves masks only for the selected images, so an incomplete mask set cannot subsequently satisfy a whole-collection phase/mask match.

minsize is the minimum mask size in pixels. mask_thresh becomes Omnipose’s mask_threshold; changing it changes how much of the predicted signal is included. The walkthrough uses 1, but it is not a universal threshold.

The current wrapper fixes channels to [0, 0], rescale to None, flow threshold to 0.0, omni=True, cluster=True, resample=True, diameter=None, and batch size to 1. These are wrapper implementation choices, not keyword arguments exposed by ImageCollection.segment_images.

Review the result

import bactoscoop.plot as bplt
fig = bplt.plot_segmentation_review(ic.image_objects, crop_size=500)

Look for merged cells, split cells, missing poles, and labels crossing the image boundary before constructing meshes. A failed segmentation sets the collection’s failure state; the same collection refuses to silently reload old masks from disk afterward. Start a fresh collection when deliberately resuming from a known mask set.

API: bactoscoop.imagecollection.ImageCollection.segment_images(), bactoscoop.omni.Omnipose, and Images, channels, and masks.