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.

Package discovery requires only bioimageflow-core. Classical image dependencies and heavy model dependencies are declared in worker EnvironmentSpec objects and imported only inside process_row. Importing this package does not require ImageIO, NumPy, SciPy, scikit-image, Cellpose, TensorFlow, StarDist, or other model packages 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.