bioimageflow-segmentation-tools

Segmentation-focused tool package for BioImageFlow.

Tools

• Cellpose3: Cellpose v3 pretrained model wrapper.
• CellposeSAM: Cellpose-SAM pretrained model wrapper.
• StarDistSegmenter: StarDist 2D pretrained model wrapper.
• InstanSegSegment: selected-target nuclei or cell segmentation with named or local InstanSeg models.
• Nagini3DSegment: volumetric NAGINI-3D segmentation with probability, parametric surfaces, and curvature outputs.
• ThresholdSegment: threshold an intensity image and label connected foreground objects.
• OtsuThresholdSegment: compute a global Otsu threshold and label foreground objects.
• LocalThresholdSegment: compute a Sauvola local threshold and label foreground objects.
• WatershedSegment: split foreground regions from marker labels or connected components.
• DistanceWatershedSegment: split foreground with marker-free distance-transform watershed.
• SplitTouchingObjects: split clumped labels using distance-transform watershed.
• FilterLabels: remove labels by area, border contact, intensity, and shape constraints.
• PostprocessLabels: remove small labels and relabel label images sequentially.

Classical connected-component tools use face connectivity by default and treat image > threshold as foreground when above is enabled. Label inputs are validated as finite, integral, non-negative arrays, and object counts are based on distinct positive IDs rather than the largest ID. Default label-output paths use TIFF rather than inheriting an input extension that may be lossy or unable to store UInt32 label IDs.

Heavy model dependencies are declared in isolated EnvironmentSpec objects and imported only inside process_row. Importing this package does not require Cellpose, TensorFlow, StarDist, or other model packages to be installed in the main process. Cellpose3, CellposeSAM, and StarDistSegmenter lazily keep one model per worker-side tool instance. Repeated calls with the same model selection reuse the weights; changing model_type or model_name replaces the cached model, and clear_model_cache() releases it explicitly. Applications can invalidate remote worker caches by stopping the corresponding Wetlands environment.

NAGINI-3D license boundary

Nagini3DSegment installs and executes the third-party nagini3D runtime, which is distributed under AGPL-3.0. The BioImageFlow wrapper remains BSD-4-Clause, but downstream distributors must review and comply with NAGINI-3D's license before publishing an environment or product containing that runtime. NAGINI model weights and datasets are not bundled in this package.

The former nnInteractive wrapper was removed because nnInteractive requires a stateful volumetric inference session; its public point-list-to-2D-mask contract did not represent the upstream API.

Example

from bioimageflow_core import Arguments
from bioimageflow_segmentation_tools import ThresholdSegment

segment = ThresholdSegment()
result = segment.process_row(
    Arguments(
        input_image="input.tif",
        threshold=128.0,
        labels="labels.tif",
        above=True,
    )
)

Workflow graph construction with segment(...) requires installing the main-process bioimageflow orchestrator alongside this package.