IQUANA is a research tool for AI-assisted segmentation and annotation of scientific image datasets — with a hierarchical label space, reproducible measurements in physical units, and a human correction loop around every model prediction.
Most annotation tools stop at the mask. IQUANA carries the same objects through measurement and review, so what you export is an analysis rather than a pile of polygons — and every number stays traceable to the object it came from.
Datasets with typed per-image metadata you can filter and group by.
Run segmentation models across a dataset, one model per label if you need it.
Review and correction queues surface what the model most likely got wrong.
Per-object metrics in physical units, driven by configurable profiles.
Measurements, provenance and label hierarchy out to your own analysis.
Built for research groups who need the measurements to be defensible, not just the masks to look right.
Prompt with points, boxes, circles or polygons — positive and negative — on top of foundation models including SAM and DINOv3, then refine the result by hand.
Labels nest into parent–child trees that say which kinds of thing are part of which, so annotated structure is nested the same way the objects are. An optional LLM assistant drafts the hierarchy from a description of your domain.
Area, perimeter, circularity and diameter per object, selected per dataset through quantification profiles. The metric registry is extensible.
Snapshot queues route predictions to reviewers in a chosen order, track what was fixed versus deliberately left, and keep model output separable from human edits.
Calibrate images so measurements come out in millimetres rather than pixels, with per-dataset defaults for consistent acquisition setups.
Queue model runs across whole datasets in the background, ordered by your label hierarchy, and patch or replace existing annotations.
Fine-tune on your corrected annotations. Runs, metrics and model versions are tracked in MLflow and registered for reuse.
Role-based access with per-dataset membership, so annotators, curators and owners see the right surface and provenance stays attributable.
IQUANA was built with marine biologists at our research partner HIFMB to segment and measure corals and polyps in reef survey photography — thousands of objects per survey, where hand-drawing every contour was never going to scale. The pipeline is general, but it was shaped against a real dataset and a real measurement question.
Semi-automated segmentation of dense, overlapping structures
Point, box, circle and polygon prompting
Positive and negative prompts for precise control
Multiple overlapping annotations per image
Zoom and pan for fine detail work
Downloadable quantification reports
One installer sets up every component. Linux and macOS are supported directly; on Windows, run it inside WSL 2.
git clone https://github.com/Iquana-tool/iquana-tool.git
cd iquana-tool
./install.shThe installer asks for the release channel, ports, CUDA support and your
HuggingFace token, then clones the components, installs dependencies and offers to start
everything. When it finishes, IQUANA is at http://localhost:3000.
| Requirement | Why |
|---|---|
git | fetching the component repositories |
uv ≥ 0.10 | Python environments for the backend and AI service |
bun | the React frontend |
| Docker or Podman | PostgreSQL and Redis run as containers |
An NVIDIA GPU is optional but strongly recommended — without one, inference and training fall back to the CPU. Full installation, update and troubleshooting notes are in the documentation, along with guides for annotation, calibration, review and quantification.
IQUANA is built around image datasets today. The annotation, quantification and review machinery is not inherently two-dimensional, and we are actively exploring what it takes to bring the same loop to other scientific modalities.
Interested in your own domain? We are a small research team at DFKI and we would rather generalise the tool against a real problem than guess at one. If you have a dataset that needs annotation and quantification, or you are interested in a research collaboration, get in touch.