What is Nitsor?
Nitsor is software for labelling CT and MRI scans, reviewing model suggestions and keeping the exact dataset behind each training run, with authors and reviews on record.
Real scans: Gruber et al., Fraunhofer EZRT (CC BY 4.0 (opens in a new tab)), Armato et al., "Data From LIDC-IDRI", TCIA (CC BY 3.0 (opens in a new tab)), Schut et al., CWI and GREEFA (CC BY 4.0 (opens in a new tab)), modified. White: published labels; dashed: model proposals from SAM 2.1, outside Nitsor. Sources
What problem does it solve?
Which labels trained this model, and who checked them?
Teams often piece the answer together from folders of masks and a spreadsheet of reviews. In Nitsor, every revision keeps its author, and each verdict on a label records the version of the label it judged. Two guides show it on public scans: who drew each of four outlines of one nodule, and how a training job pins a dataset version.
- Who labelled this scan?
- When someone submits their labels, Nitsor saves them as a commit. The commit names who submitted them, the revision before it and the time.
- Who checked it?
- When a reviewer finishes a review, Nitsor saves it as a commit of its own under the reviewer's name. Each verdict on a label records the version of the label it judged.
- What did a training job read?
- A training job can reference one fixed dataset version and the saved label revisions behind it.
What does it do?
Label, review and keep the record.
Nitsor keeps model suggestions and people's decisions with the revisions they changed. We call it the system of evidence for AI-assisted inspection.
- 3D workspace
- The browser editor where people label 3D scans.How the workspace works
- Model proposals
- Labels a model suggests, recorded on their own branch for a person to accept, correct or reject.How model proposals work
- Review
- A person's verdict on a piece of work: accept, correct or reject.How review works
- Routed review
- Sending work to reviewers by model score or by a sampling rule.How routed review works
- Versioning
- Submitted labels are saved as commits, each with its author and the revision before it.How versioning works
- Dataset versions
- A fixed selection of labelled scans that a training job keeps using while labelling continues.How dataset versions work
Who is it for?
Teams that train models on 3D scans.
One record serves the quality lead, the ML engineer and the person who runs the labelling queue.
- Quality, clinical and research leads
- Every mark keeps its author, and every review keeps its reviewer.
- ML engineers
- Choose a dataset version, a fixed selection of labelled scans, for each training job. Read it back by ID.
- Data operations
- Labelling and review run as one pipeline, with no bill per person.
Where does the data go?
Your storage keeps the scans. Nitsor keeps the record.
The browser reads viewed slices from your storage through short-lived signed links, and model jobs work on temporary copies while they run.
- In your storage
- Your source scans, the masks people draw and the viewer copies Nitsor makes of registered scans, under a prefix you choose.
- In Nitsor's records
- File locations, checksums, scan geometry, labels, reviews and the history of the work. The records hold no image data.
Who decides?
Your quality authority makes the call.
Your organisation evaluates the record against its own requirements.
- What it records
- Authors, review decisions, class definitions and dataset history.
- What it never does
- Nitsor also never decides whether a part passes inspection, never certifies regulatory compliance, and never sells annotation labour.
How does it run?
Self-host it, or let Nitsor host it.
Choose the plan that matches who operates the software. No plan charges per person.
- Community
- Free and self-hosted on your own infrastructure, on-premises or in your private cloud.
- Team
- Hosted by Nitsor, with managed updates and backups of the hosted workspace data.
- Enterprise
- Hosted by Nitsor, or by agreement a private deployment on your premises or in your private cloud, set up with our help.
Show us what you inspect.
Start with a public CT or MRI scan. See how a model suggestion becomes a reviewed label and a fixed training dataset.