Skip to content

Skeleton And Network Analysis

Skeleton workflows analyze curvilinear structures such as mitochondria, neurites, vessels, fibers, and similar networks.

Basic QC Workflow

binary mask
  -> Skeletonize
  -> Skeleton Graph Overlay

Use the overlay to check endpoints, junctions, isolated voxels, and branch structure before trusting tables.

Choose 2D or volumetric 3D skeletonization

Select Skeletonize and review both Spatial processing and Method:

  • Auto from axes resolves declared YX as 2D and declared ZYX as one 3D volume. An ambiguous QYX source is rejected until its page meaning is declared.
  • 2D YX processes each leading YX plane independently.
  • 3D ZYX processes each ZYX block once as a volume. For TZYX, each time point is an independent 3D block; time points are never joined.

Auto chooses Zhang for 2D and Lee for 3D. Zhang 2D cannot be combined with a resolved 3D ZYX volume. The output preserves shape, effective axes, calibration, and leading dimensions, and its history records the resolved method and whether processing used a 2D plane or a volumetric block. The 3D Lee route uses a 3x3x3 neighbourhood with the image boundary treated as background.

Slice-wise and volumetric results answer different questions

Independent YX thinning can change cross-Z connectivity, endpoints, junctions, cycles, and branch lengths even when every individual slice looks plausible. Use 3D only when Z semantics and spacing are independently supported by the acquisition record.

For visual QC, duplicate a representative mask branch and run one Skeletonize node as 2D YX and the other as 3D ZYX. Inspect XY, XZ, and YZ views in napari, then compare overlays, component counts, endpoints, junctions, and branch lengths. A bundled synthetic phantom is useful for regression, but it does not validate a biological segmentation or the acquisition's Z scale.

Measurement Workflows

Join skeleton and morphology per object

New in 0.16.0a2

Skeletonize Labels and Analyze Skeleton per Label preserve the object IDs needed for a reliable table join. Existing skeleton-first workflows remain supported.

Choose Open example… → Skeletons & Networks → Per-label Skeleton & Morphology to try a three-object synthetic image with calibrated spacing.

  1. Label the segmented mask and apply object filtering once.
  2. Connect the filtered labels to both Measure Objects and Skeletonize Labels. Inspect the skeleton image: surviving skeleton voxels keep their original object label.
  3. Connect those same filtered labels to Analyze Skeleton per Label → Original labels and the labeled skeleton to its optional Skeleton input.
  4. Inspect the Objects summary table and Skeleton components output. The summary keeps one row for every original object; component details retain both label_id and a component ID within that object.
  5. Connect morphology and the per-label summary to Merge Tables. For this single 3D image, join on label_id. Across images, timepoints, channels or slices, include those identity columns too.
  6. Open the example's scatter plot to compare each object's volume with its own skeleton length. Keep the full raw tables before selecting analysis features.
flowchart TB
    A[Filtered object labels] --> B[Measure Objects]
    A --> C[Skeletonize Labels]
    A -->|Original labels| D[Analyze Skeleton<br/>per Label]
    C -->|Skeleton| D
    B --> E[Merge Tables]
    D -->|Objects summary| E

The sample contains a branched network, a separate line and one isolated voxel. Its synthetic spacing is 0.45 micrometer on all three axes. Raise the minimum volume from 1 to 2 voxels and calculate again: the isolate is removed from both branches. These settings demonstrate consistent object selection; they are not segmentation settings validated for acquired images.

Retain connected-component and isolate measurements. They describe different features: skeleton_component_count counts pieces within an object; isolated_node_count counts skeleton voxels with no neighbors. The latter does not count every separate mitochondrial object. For the cell's complete network, keep a separate Measure Overall Skeleton Network branch as needed. Adjacent skeletons from different labels can connect when treated as one binary mask, so its topology need not equal a sum of per-label metrics.

See the per-label reference for empty skeletons, calibration, graph connectivity and table identity.

Keep skeleton-first analysis when that is the question

Use Skeletonize → Label Skeleton Components when you want to discover connected pieces of a binary skeleton. Its IDs describe skeleton pieces and remain independent of any earlier object labels. That existing behavior is unchanged. The label-first route above is preferred when the aim is to join measurements to original object morphology.

Per-component measurements:

binary mask
  -> Skeletonize
  -> Analyze Skeleton

Branch-level measurements:

binary mask
  -> Skeletonize
  -> Measure Skeleton Branches

Branch summaries:

binary mask
  -> Skeletonize
  -> Measure Skeleton Branches
  -> Summarize Skeleton Branches

Whole-network summary:

binary mask
  -> Skeletonize
  -> Measure Overall Skeleton Network

Explicit graph tables:

binary mask
  -> Skeletonize
  -> Skeleton Graph Tables

Cleanup Workflow

binary mask
  -> Skeletonize
  -> Prune Skeleton Branches
  -> Analyze Skeleton / Measure Skeleton Branches

Pruning removes short terminal spurs and optionally isolated skeleton voxels.

Reference Workflows

Workflow Purpose
synthetic-per-label-skeleton.json New in 0.16.0a2: retain original object labels, inspect skeleton summaries and components, and join morphology before plotting.
synthetic-skeleton-qc.json Compact 3D skeleton QC with keypoints, branch/component labels, pruning, and tables.
synthetic-advanced-skeleton-network.json Time-indexed 3D skeleton/network stress test with loops, fragments, graph tables, and anisotropic scale.

What To Check

The skeleton node reference lists each node's outputs, units, graph terms, and pruning behavior.

  • Is the input a binary mask or an already skeletonized mask?
  • Should analysis be 2D per slice or 3D volumetric?
  • Is z-spacing correct before reporting physical lengths?
  • Are short spurs biological or threshold/skeletonization artifacts?
  • Are endpoint and junction counts plausible in the graph overlay?