CT labels your inspectors can trace

Connect 3D defect labelling, expert review and training datasets, so your team can trace a model's inputs back to each inspector's decision.

Request a walkthroughRead the docs (opens in a new tab)

Video still: NASA Johnson Space Center, Public domain (NASA) (opens in a new tab), graded. Sources

Why it matters

Which labels trained this defect model?

Find the exact scans, label revisions and review decisions behind a training run, without piecing together folders and review notes.

  1. Fraunhofer Me 163 V5, industrial CT: the section around a pressure bottle.
  2. A plane opens the scan at axial slice 300.
  3. Axial slice 300, in the scan's own greys.
  4. Dashed: SAM 2.1's proposal, from a recorded run outside Nitsor.
  5. White: published instance 138, the bottle wall (Fraunhofer EZRT).
  1. A section of an aircraft fuselage around a pressure bottle, rendered from its industrial CT scan.

    1/5Fraunhofer Me 163 V5, industrial CT: the section around a pressure bottle.

  2. A plane cuts the section open, showing the round wall of the pressure bottle.

    2/5A plane opens the scan at axial slice 300.

  3. One CT slice with the bright round wall of the pressure bottle.

    3/5Axial slice 300, in the scan's own greys.

  4. A close view of the bottle wall with the model's proposal as a dashed outline.

    4/5Dashed: SAM 2.1's proposal, from a recorded run outside Nitsor.

  5. The bottle wall with its published outline in white beside the model's dashed proposal.

    5/5White: published instance 138, the bottle wall (Fraunhofer EZRT).

Real industrial CT and its published outline: Gruber et al., Fraunhofer EZRT, CC BY 4.0 (opens in a new tab), modified (rendered, cut, windowed, outline drawn). Model proposal from SAM 2.1, Meta AI, Apache 2.0 (opens in a new tab), run outside Nitsor.

How it works here

Review the finding. Fix the training set.

Your inspectors mark porosity, cracks, inclusions and delamination on the volume itself, at the scan's own spacing. Every mark carries the name of whoever drew it.

  1. Register

    Connect the scan in your storage.

    The volume stays in your object storage. Nitsor keeps the path, size, checksum and scan settings. The browser reads viewed slices from that storage through short-lived signed links, and model jobs work on temporary copies.

  2. Label

    Mark the defect in 3D.

    Brush, contour or threshold across slices for porosity, shrinkage, cracks, inclusions, delamination and lack of fusion.

  3. Review

    A second inspector decides.

    A reviewer accepts, corrects or rejects each finding. Nitsor keeps the author, reviewer and revision together, including rejected work.

  4. Pin

    Freeze the reviewed set.

    Save a dataset version, a fixed selection of labelled scans. Train on that version while inspectors continue their work.

What the record holds

Show which revision the inspector checked.

Each verdict on a label records the version of the label it judged. A dataset version fixes the scans and labels used for training.

The Nitsor editor on an industrial CT crop of an aircraft part: stepping through the slices, a blue label covers the wall of a ring-shaped part; the header reads VGStudio volume (.vgi/.vol) with a voxel size of 0.33 x 0.33 x 0.60 mm; and the linked planes cross at one point with their position lines.
What each frame shows
  1. Step through the scan slice by slice
  2. A part's outline is a label
  3. The header gives the format and voxel size
  4. Linked planes cut through the same point
  5. The position lines follow the slice

Real industrial CT: Gruber et al., Fraunhofer EZRT, CC BY 4.0 (opens in a new tab), modified: cropped to 384 x 384 x 384 voxels. Label outline from the dataset's published instance masks.

See how dataset versions work

Finding: porosity, scan housing-0142

Proposed by
SAM 2.1, as a suggestion
Corrected by
M. Osei, inspector, 3 slices
Accepted by
D. Novak, senior inspector
Review of
the corrected outline
In version
dsv_0f3a9c21

Worked example on public scans. Names, decisions and version IDs are examples.

  • One record per scan

    Its settings, its findings, who drew them, who reviewed them and every change since.

  • Measurements use the scan's spacing

    The import keeps the voxel size, so a distance you measure is in millimetres.

  • Training data that stays fixed

    A dataset version fixes its scans and label revisions. New edits do not change a training job's inputs.

Which scans open

Bring reconstructed volumes as DICOM Part 10 (uncompressed) or VGStudio .vgi/.vol. They open in the browser at their own spacing.

What gets scanned

  • Aerospace

    Airframe structures and fasteners.

  • Automotive

    Castings, welds and assemblies.

  • Energy

    Turbine blades, battery cells and pipe welds.

  • Additive manufacturing

    Printed parts and lattices.

Worked example

One public scan, 169 labelled parts.

Inspect skin panels, frames, stringers, pipe clamps and rivets in a 512-slice volume, each part with its own published label. Beside it, a carbon-fibre laminate shows damage labelled by hand. Bring your own castings or composites to a walkthrough.

An industrial CT slice of an airframe section with every published part outlined in white and the six largest numbered by their label IDs.

Me 163 V5, slice 300: published instances

On this slice
18 parts
In the volume
169 parts
Labels
one ID per part

IDs are shown for the six largest parts on the slice.

Real industrial CT and instance labels: Gruber et al., Fraunhofer EZRT, CC BY 4.0 (opens in a new tab), modified (one slice windowed, parts outlined).

A render of the same airframe section with 3D boxes around its six largest parts, derived from the published labels.

Me 163 V5: 3D boxes derived from the published masks

Boxes
the 6 largest parts
Made from
the published instance masks

The render is a still from one fixed camera.

Real industrial CT: Gruber et al., Fraunhofer EZRT, CC BY 4.0 (opens in a new tab), modified (rendered, boxes derived from the published labels).

A synchrotron CT slice of a carbon-fibre laminate with its hand-labelled damage in three classes shown in white at three strengths.

Carbon-fibre laminate, synchrotron CT: published damage labels

  • Damage in 0° plies
  • Damage in ±45° plies
  • Damage in 90° plies

Labelled by trained people. The scan is dimmed so the labels read.

Real synchrotron CT and damage labels: Kopp et al., MIT, CC0 (opens in a new tab), modified (cropped, dimmed, labels drawn in white).

Questions

Before you choose Nitsor

Our customer asks which labels trained the model. What do we show them?
Show the dataset version the training job read. It identifies the scans, label revisions, authors and recorded reviews, including model proposals.
Does Nitsor decide whether a part passes?
No. Nitsor records the labels and reviews. Your quality authority makes the call. Nitsor also never decides whether a part passes inspection, never certifies regulatory compliance, and never sells annotation labour.
Does Nitsor's integrity check say the labels are right?
No. It says they have not changed since the version was fixed. Whether they are right is your reviewers' call, and it is on record.
Our volumes are large. Do we upload them?
No. Register them where they sit. The browser reads viewed slices from there through short-lived signed links, and model jobs work on temporary copies while they run.
Can outside inspectors review without raising the bill?
Yes. Nitsor does not charge per person. Give outside inspectors the access their review requires.

Start with one part.

See a model proposal, an inspector's correction and the dataset version used for training. Start with a public CT volume.

Request a walkthroughRead the docs (opens in a new tab)