Nitsor journal

Practical guides for teams labelling 3D scans. Each one works through public CT data, published annotations and recorded model runs.

Slice 300 of the airframe scan: a pressure bottle, a stringer and frames, each part outlined in white with its published ID.A frame junction, a small ring and the edge of the pressure bottle in the airframe scan, each part traced by its published outline.
Real industrial CT with its published part outlines, Gruber et al., Fraunhofer EZRT, CC BY 4.0 (opens in a new tab), modified. Sources

How do we prepare industrial CT labels for training?

Industrial CT annotation: prepare component labels for machine learning

One public airframe scan with 169 published part labels, and the decisions a team makes before any label reaches a training set: what it means, how it is checked and which version a model reads.

7 min read

  1. A lung nodule on a chest CT slice, outlined four times in white by four radiologists and once, dashed, by a model.

    How do we check label quality when experts disagree?

    CT annotation quality control: measure reader agreement and review disagreements

    8 min read

  2. A CT slice of an apple with a browned core, and a model's proposal dashed around the core.

    Can we reproduce the data behind a model we trained last quarter?

    Medical imaging dataset versioning: reproduce the scans and labels behind a training run

    7 min read

  3. The chest CT rebuilt as one volume and opened on a coronal plane through both lungs.

    Why do our masks land on the wrong slice?

    DICOM annotation workflow: check slice order, spacing and mask alignment

    7 min read

Pictures: real industrial CT, Gruber et al., Fraunhofer EZRT, CC BY 4.0 (opens in a new tab); real clinical CT, Armato et al., "Data From LIDC-IDRI", TCIA, CC BY 3.0 (opens in a new tab); food X-ray CT, Schut et al., CWI and GREEFA, CC BY 4.0 (opens in a new tab); all modified. White outlines are published labels, and dashed ones are SAM 2.1 proposals, run by Nitsor outside the product. Sources

Every number in these guides is calculated from public data, and each guide's notes name the dataset, the licence and the date. Found an error? Send us the detail and the source, and we will check it and say what changed. Contact us.

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.

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