Skip to content

Build your first workflow

In this exercise you will create a short segmentation graph from an empty canvas. It uses bundled data, so unexpected results are easier to diagnose.

1. Start clean

Choose New workflow.... VIPP opens a new tab and preserves any other workflow tabs. The new graph contains one unbound Image Source on an otherwise empty canvas. Select that source (do not add a second one) and set:

Setting Value
Source sample
Sample VIPP synthetic multichannel volume

Wait for the source thumbnail and metadata to appear.

2. Add and connect nodes

Search the node library and add these operations:

Image Source → Extract Channel → Gaussian Blur → Otsu Threshold
             → Fill Holes → Label Connected Components

Drag from an output port to a compatible input port. Typed ports prevent many invalid connections; they do not detect scientifically inappropriate choices.

Set Extract Channel to the red channel. If channel names are available, prefer the semantic name; otherwise verify the index visually.

3. Tune while looking upstream and downstream

Select Gaussian Blur. Change its scale gradually and compare:

  • small puncta in the extracted channel;
  • speckle in the smoothed image;
  • foreground retained by Otsu Threshold;
  • object merging after connected components.

Restore a conservative value before continuing. There is no universally correct blur or threshold: tune against the structures your assay intends to measure.

New workflows request Auto compute. After calculation, read the compact badge on Gaussian Blur, Otsu Threshold, and Label Connected Components. Auto begins with reviewed safe GPU defaults. Accelerated-only exact compatible history makes the next global Auto run measure CPU once on the same execution surface; a later matching run applies the 1.20x/20-ms gate. A CPU badge can reflect eligibility, the environment, that exploration step, or a learned CPU assignment and is not proof that setup failed. Keep this first tutorial scientifically focused; use Prefer GPU or Custom with the separate CPU/GPU guide before benchmarking providers.

4. Add a measurement branch

Add Measure Objects and connect the labels to it:

Label Connected Components → Measure Objects

Select the measurement node and choose Calculate. Check the row count, column names, units, and whether leading axes are represented as expected.

5. Save an inspectable checkpoint

Choose Save workflow and use a name that communicates its purpose and release, for example:

nuclei-segmentation-vipp-0.15.0a2.json

Also record:

  • the input dataset or acquisition family used for tuning;
  • any manual exclusions;
  • expected axes and physical units;
  • one or more reference images used to judge the result.

Workflow JSON is necessary but not sufficient provenance. See the workflow contract for what it does and does not preserve.

6. Test that the workflow reopens

Create another workflow tab, then load the saved JSON. Open opens another session rather than replacing the active graph. Confirm the source, graph, parameters, connections, and authored compute request before relying on the file as an analysis record.

Next, learn how to use your own images without assuming that the same parameters will transfer unchanged.