Before you choose Nitsor

What is Nitsor?

Nitsor is software for 3D scan labelling, expert review and versioned training datasets. It keeps model suggestions and people's decisions with the revisions they changed. We call it the system of evidence for AI-assisted inspection.

How does it run?

Community runs on your infrastructure for free. Team is hosted by Nitsor at app.nitsor.com for $649 per workspace per month. Enterprise covers agreed deployment and support requirements.

Nitsor is source-available. Not open source, and we will not call it that.

Which screens does Nitsor support?

Labelling and review work on laptop and desktop screens, 1280 px wide and up. Tablet and phone screens are not supported.

How do we start?

Choose Community to self-host, or request a walkthrough to start on hosted Team. Paid checkout and usage charging are not active. Use a public scan to show us your workflow.

Can we label ultrasound images?

The ultrasound workflow covers uncompressed single-frame DICOM ultrasound images for research. PixelSpacing or valid Ultrasound Regions calibration supplies physical scale; uncalibrated images have no physical scale. Images without spatial geometry stay individual images. Compressed transfer syntaxes, multi-frame, cine and native 3D or 4D acquisitions are outside this workflow.

Where do the scans live?

Existing scans stay in connected S3-compatible storage. Upload to your storage or a Nitsor-managed bucket, or let Nitsor manage all storage. Hosted uploads default to Nitsor storage, which the owner can switch off. The browser uses short-lived signed links, and model jobs use temporary copies.

Do you keep a copy of my scans?

Existing scans stay in their connected storage. Hosted uploads default to Nitsor-managed storage; the owner can switch it off or use a write-enabled customer bucket. Read-only connections keep source metadata, point, box and class labels, reviews, audit events and dataset version manifests in Nitsor, never scan pixels or derived files. Masks, stored previews and model outputs require write access. Uploads also need an owner or admin to enable Accept uploads after the storage checks pass. Browser reads use short-lived signed links; model jobs use temporary copies.

Can the storage sit inside my own network?

Yes, if your Nitsor deployment can reach it and the storage supports the required S3 operations. Community runs on infrastructure you manage.

What happens if a scan changes after it was labelled?

The next read notices. Registration records each object's content identity, and a read that finds the file changed is refused with an error that says it was replaced after registration, rather than showing different bytes under the same labels.

Which formats does Nitsor open?

The viewer and editor open DICOM Part 10 (uncompressed) for CT and MRI series; VGStudio .vgi/.vol for industrial CT volumes.

These register as raw acquisition data and do not open as images: MRD for ISMRMRD magnetic-resonance acquisition data; TIFF projection stacks for CT projections that come with a Nikon .xtekct descriptor. Other industrial CT vendor formats are recognised by name but do not open.

Spacing, orientation and origin survive the import, and each series keeps its own grid.

What comes out?

A dataset version: a fixed selection of labelled scans with digests, pinned by the jobs that read it. The API serves its members and history.

What happens to our data if we stop paying?

Scans referenced in customer storage stay there. Nitsor does not charge an exit fee. Hosted retention and access after cancellation follow the published plan terms.

How do I prove who labelled a 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.

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.

A training job can reference one fixed dataset version and the saved label revisions behind it.

How do I see what changed in a revision?

Open the project's History tab. Each revision shows what it added, removed and changed, who saved it and when.

In the viewer, the Revision control opens an earlier revision on the same slice, so you read the labels as they stood.

What does a reviewer actually do?

Accepts, corrects or rejects each piece of work. The verdict is bound to the version they saw, and both names stay on it.

Rejected work can be sent back with the reviewer's note. Corrections follow the configured workflow, and the verdict stays in the record either way.

Can several readers label the same study?

Yes. A consensus read sends a study to 2 to 16 readers, blind to each other by default. Nitsor scores their agreement: Dice or IoU for masks, distance in millimetres for points and exact match for each field.

When reads miss your agreement criteria, such as a minimum Dice or a maximum distance in millimetres, an adjudicator takes one reader's labels, keeps the original or edits their own. Each read and decision is recorded in History and available through the API.

What can an auditor see?

Give an auditor read-only access to the relevant labels, reviews, changes and dataset versions without a per-person charge. The records show who did what. Your quality process determines whether the work meets its requirements.

Can I see how things were last March?

Yes. Every commit is kept, and each carries the class list in force when it was made.

Can a model label first and a person check?

On Team and Enterprise. A model proposes on its own branch, and a person accepts, corrects or rejects each proposal. Community has no hosted compute and no usage meter. Team includes 250 model-assisted volumes in each monthly billing period on Nitsor's GPUs. One model-assisted volume is one completed scan/model result under the existing scan identity rules, with every object the model finds included.

How do we choose what goes to review?

Set a score threshold or sampling rule for the step into review. Nitsor records which work was routed and why. A routing decision does not establish that a label is correct.

Do you publish a speed number?

No. Measure it on your own scans: every proposal, accept and reject is on record.

What does it cost?

Community is free and self-hosted. Team is hosted at $649 per workspace per month, or $6,228 upfront per workspace per year ($519 per month equivalent), with 250 model-assisted volumes included in each monthly billing period. Each extra volume costs $0.25. Included volumes reset at the start of each monthly billing period on both Team plans; unused volumes do not carry over. The yearly plan does not pool the allowance for the year. Enterprise starts at $12,000 per year, billed yearly, plus a volume commitment at a rate set in your contract. Model runs are priced per completed scan/model result, never per person. These are launch prices. Paid checkout and usage charging are not active. The subscription is paid upfront. Extra usage is settled after each monthly billing cycle. A cycle follows your billing dates and can start partway through a calendar month.

For example, 1,000 volumes in a monthly billing period cost $649 + 750 × $0.25 = $836.50 on the monthly Team plan. You see and confirm the quote before a run starts. Within one billing period, the first successful run of a model version on a scan counts one volume. Later runs of the same model version on the same scan in that period are free. A new billing period, a new model version or a different scan counts again. A failed run never counts. A workload outside the supported models and scan sizes, including an exceptionally expensive run, needs a separate fixed quote agreed before it runs, with no surprise runtime charge.

Do we pay for each annotator or reviewer?

No per-seat fees. Add annotators and reviewers without paying for each person. Team is billed per workspace, with model-assisted volume usage under its plan terms. Fair use applies.

Is anything limited on Team?

Team is unlimited for people, labels and reviews under fair use. Fair use also covers Nitsor-managed storage, including uploaded scans, label records, history and dataset versions, and the volume of API traffic, including from agents and scripts. Existing scans referenced in customer storage remain in that storage.

Fair use is not a meter, and Nitsor does not bill for it. If a workspace uses far more than a typical team, we talk to its owner and agree a plan that fits. We give notice before we change this policy.

What about compliance with my standard?

Nitsor makes no compliance claim. It records authors, review decisions, class definitions and dataset history. Your organisation evaluates those records against its requirements. Your quality authority makes the call.

Nitsor also never decides whether a part passes inspection, never certifies regulatory compliance, and never sells annotation labour.

Is there an API?

Yes: an HTTP API described by an OpenAPI 3.1 document. Agents and scripts call it with a scoped credential that expires, issued in the hosted app on Team and Enterprise. People, models and agents write under named accounts with the limits you set.