Medical AI annotation

Medical AI Data Annotation for Imaging, Pathology and Surgical Video

Organ and lesion masks, slide regions and instrument tracks, labelled to your clinical guideline and reviewed by experts you approve, on de-identified data handled under your compliance rules.

Start a Pilot Explore Services

Compatible workflows

Tools we can work with

Roboflow
SuperAnnotate
Darkhorse
Roboflow
SuperAnnotate
Darkhorse

TL;DR

We segment organs and lesions in DICOM and NIfTI volumes, with rules for slice-by-slice consistency and full 3D masks.
Pathology slides and surgical video get region, cell and instrument labels with written definitions for ambiguous tissue and smoke-filled frames.
Clinical judgement comes from expert reviewers you approve. Annotators are calibrated against them before production.
We work on de-identified data under your own compliance requirements. We don't claim certifications or clinical accuracy; your team owns clinical validation.

Radiology imaging

DICOM and NIfTI segmentation: organs, lesions and 3D volumes

Scans such as CT and MRI arrive as stacks of slices. A good mask has to be correct on each slice and make sense as a three-dimensional shape.

Organ and lesion masks

Organs and structures are segmented as labelled masks, using a class list and boundary definitions written with your clinical team. Lesions are labelled separately, with attributes such as lesion type or confidence when your protocol defines them. Where a boundary is genuinely unclear, the guideline says whether to include, exclude or flag the uncertain region.

We keep the original orientation, spacing and series information, and deliver masks that line up with the source volume. Window and level settings used for labelling are written down so reviewers look at the same contrast.

Slice-by-slice consistency

A mask drawn slice by slice can jump or vanish between neighbouring slices. Reviewers scroll through the stack and check the shape in other planes, such as sagittal and coronal views, to catch gaps, sudden jumps and islands that do not belong to the structure.

Working in 3D

For volume models, masks are checked as 3D objects: one connected structure where anatomy says it should be, smooth surfaces where the tissue is smooth, and separate instances for separate lesions. Interpolation between slices can speed work up, but interpolated slices are always reviewed, never accepted blind.

Our general mask, polygon and keypoint rules are explained on the image annotation page.

Digital pathology

Pathology slide annotation

Whole-slide images are very large and full of fine detail. Labelling them well depends on clear definitions and a fixed magnification for each decision.

Regions, cells and grades

Projects can include tissue region polygons, cell-level points or boxes, and slide or region classes that follow your grading scheme. The guideline fixes the magnification for each label type, because a boundary drawn at low zoom looks different from one drawn at high zoom.

Ambiguous areas, such as tissue folds, out-of-focus patches and staining artefacts, have their own labels, so the model is not trained on them as if they were real tissue.

Surgical video

Surgical video tracking and phase labelling

Surgical footage combines fast instrument motion, smoke, blood and reflections. It needs video rules, not still-image rules.

Instruments and anatomy over time

Instruments are segmented or boxed and keep one track ID while they stay in view, including when they briefly leave the frame and return. Visible anatomy can be segmented on keyframes with reviewed interpolation in between.

Phases, steps and events

Procedure phases and steps are labelled as time spans with written start and end cues. Frames hidden by smoke, lens fog or a blood-covered lens are tagged so they can be excluded or studied separately.

Track IDs, interpolation and event boundaries are covered in depth on our video annotation page.

Expert review

Expert reviewer calibration

Medical labels need clinical judgement. Our role is to turn that judgement into written rules that can be applied the same way across thousands of cases.

Calibrating annotators to your experts

During the calibration pilot, annotators and the expert reviewers you approve label the same set of cases. Disagreements are discussed and settled by the experts, and each decision becomes a rule with example images. Annotators move to production only after their work matches the expert reference on the calibration set.

Cases that need clinical judgement beyond the written rules are escalated to the expert reviewers, not decided by the annotation team.

Data handling

Privacy-conscious workflows on de-identified data

We work with de-identified data, under NDA, and under the compliance requirements your organisation sets. Those requirements are agreed before any data moves.

What we ask for and how we handle it

We ask clients to remove direct identifiers from files and metadata before sharing, including burned-in text on images. If an identifier is found during labelling, work on that item stops and it is reported to you. Access is limited to the project team, and retention and deletion follow your instructions.

VIPRO does not claim HIPAA compliance, certifications or clinical accuracy. Your team stays responsible for clinical validation and regulatory decisions about the model.

Quality

Two-stage QA for medical labels

Every batch passes an annotator self-check and an independent reviewer audit, with expert escalation for clinical questions.

Review in practice

Stage one checks masks against the written rules: classes, boundaries, slice continuity and required attributes. Stage two is an independent audit by a reviewer who did not label the item. Rejections come back with the rule broken, and guideline changes are versioned so every delivered mask can be traced to the rules it followed.

Medical AI annotation FAQ

Can you work with DICOM and NIfTI files directly?

Yes. We label volumes in their source formats and keep orientation, spacing and series information so masks line up with the original scans. The exact tools and export format are agreed during the calibration pilot.

How do you keep masks consistent between slices?

Reviewers scroll through each stack and check the structure in other planes to catch gaps, jumps and stray islands. Interpolated slices are always reviewed, and 3D masks are checked as connected structures where the anatomy calls for it.

Who makes the clinical decisions?

Expert reviewers you approve. Annotators are calibrated against their reference labels before production, and cases that need clinical judgement beyond the written rules are escalated to those experts.

Are you HIPAA certified?

We don't claim HIPAA compliance or any certification. We work on de-identified data under NDA and under the compliance requirements your organisation sets, with restricted access and retention and deletion on your terms. Clinical validation of the model stays with your team.

Test us on a de-identified sample set.

Start with a Free Calibration Pilot