Your First Pipeline¶
A five-minute end-to-end run: load a hyperspectral hand recording, compute a
per-frame NDVI projection that highlights blood perfusion, and render the
result as MP4. No Python required — the entire flow uses the
restore-pipeline CLI against a pre-authored YAML.
NDVI was first developed to map chlorophyll in satellite imagery, but hemoglobin has a similar spectral footprint: it absorbs visible light strongly (~566 nm) and transmits near-infrared (~750 nm). The same normalized-difference formula doubles as a qualitative perfusion index:
If you have not installed cuvis-ai yet, follow the Installation Guide first.
1. Get sample data¶
The XMR_Demo_Blood_Perfusion dataset (~11 GB) is a 568-frame hyperspectral recording of a human hand:
This drops .cu3s files under data/XMR_Demo_Blood_Perfusion/. Other demo
datasets live in the datasets catalog.
2. Inspect the pipeline you're about to run¶
The pre-authored NDVI blood-perfusion pipeline lives at
configs/pipeline/medical/blood_perfusion/ndvi.yaml. It is a three-node
graph that loads each cube, computes NDVI at 750 nm vs 566 nm, and writes
the colormapped frames to an MP4:

View the node graph and port wiring from the CLI:
3. Run inference¶
Point the pipeline at any .cu3s file from the dataset — Auto_005.cu3s is a
good first run:
uv run restore-pipeline \
--pipeline-path configs/pipeline/medical/blood_perfusion/ndvi.yaml \
--cu3s-file-path data/XMR_Demo_Blood_Perfusion/Auto_005.cu3s
A few minutes later you have output/blood_perfusion/ndvi_projection.mp4 —
the colormapped NDVI video at 15 fps:
4. Where to go next¶
- Blood Perfusion Tutorial — the spectral theory, plus an extension that adds a qualitative SpO₂ proxy.
- Notebook walkthrough — the same pipeline constructed in Python, with custom-node extensions.
- Concepts — the mental model: nodes, ports, pipelines, training.
- Workflows → Build Pipeline (YAML) — author a YAML pipeline from scratch.
- Workflows → Restore Pipeline — full reference for the
restore-pipelineCLI. - Tutorials — the rest of the gallery: RX Anomaly, Channel Selection, Deep SVDD, AdaCLIP.