Medical data annotation
Labels a radiologist would sign.
Segmentation, detection, classification and report structuring on CT and MRI, read twice, adjudicated by radiologists, and shipped with the agreement scores attached. On our catalogue or on your data.
What we annotate
From voxel masks to structured reports
Medical labelling fails on expertise long before it fails on volume. Every task below is read by people who can read the scan, and reviewed by people who sign reports for a living.
Organ and lesion segmentation
Voxel-level masks on CT and MRI: liver, kidneys, pancreas, spleen, lung lobes, vertebrae, intracranial haemorrhage, focal lesions. Delivered as DICOM SEG, RTSTRUCT or NIfTI with the label map documented.
Detection and localisation
Bounding boxes and key slices for nodules, fractures, haemorrhage, free fluid and incidental findings, with the radiologist report linked to each mark.
Study and series classification
Normal versus abnormal, finding present or absent, protocol and phase labels, image-quality grades. Fast, high-volume labels for triage and QC models.
Report structuring and extraction
Section labelling, finding extraction, laterality, size and location, assertion and negation, mapped to RadLex or your own schema, from de-identified radiology reports.
Measurement and grading
Lesion diameters, volumes, Agatston scores, spinal canal measurements and standard grading scales, applied consistently by trained readers and checked by radiologists.
Model output review
Radiologists grade your model’s outputs, rank alternatives and write critiques: expert preference data for clinical fine-tuning and a hard evaluation set that is not in your training data.
Quality
Agreement is reported, not averaged away
A label set is only as good as the disagreements it resolved. We make those disagreements visible: how often readers differed, who adjudicated, and what the agreement score was for every batch.
Schema and guideline first
Every project starts with a written label schema, inclusion rules and worked examples agreed with your clinical lead. Ambiguity is resolved in the guideline, not in the data.
Double read and adjudication
Specialty work is read independently twice. Disagreements go to a senior radiologist for adjudication, and the adjudicated label is what ships.
Agreement reported per batch
Inter-annotator agreement, Dice or kappa as appropriate, is reported with every batch, so label quality is a number you can track rather than a promise.
Calibration and drift checks
Readers are calibrated against a gold set before production and re-checked on seeded cases through the project, so quality does not drift as volume rises.
How a project runs
Pilot first, then volume
Scope the schema
We define the labels, the clinical rules and the output format with your team, and annotate a pilot set to prove the guideline works before volume starts.
Pilot and calibrate
A pilot batch of fifty to two hundred studies is read, adjudicated and scored. You review the labels in your own tooling and we adjust the guideline together.
Production in batches
Labels ship on an agreed cadence with agreement metrics, reader counts and adjudication rates attached, in the format your pipeline ingests.
Iterate on failure modes
As your model’s weaknesses surface, the schema and the sampling change with them. We target the cases your model gets wrong, not the ones it already handles.
Formats
Delivered the way your pipeline reads it
Labels are produced in standard medical formats, linked to the de-identified study and series UIDs they describe, so they load against the data without a crosswalk.
Compliance
Handled as health data, start to finish
- Annotators work only on de-identified studies, in segregated project environments with audited access
- Business Associate Agreements or equivalent where identified data must be handled first
- NDA, confidentiality and protocol training for every reader before production
- Annotation provenance recorded per label: reader, adjudicator, guideline version and timestamp
- Documentation suitable for EU AI Act Article 10 data-governance files
FAQ
Frequently asked questions
Who performs the annotation?
Trained medical annotators working under radiologist supervision, with practising radiologists performing specialty reads, adjudication and final review. Every reader signs an NDA and completes confidentiality and protocol training before touching a study.
Can you annotate our data rather than yours?
Yes. We work on your de-identified studies in a segregated project environment, or on our catalogue if you are licensing data from us. If your data is not yet de-identified we can take that step first under a Business Associate Agreement or equivalent.
Which tools do you use?
We work in established medical annotation platforms and can work inside yours if you prefer the labels to be created where you will use them. Output is tool-independent: standard DICOM, NIfTI and JSON formats.
How do you report quality?
Each batch carries inter-annotator agreement, the adjudication rate and the number of readers per study. For segmentation we report Dice against the adjudicated mask; for classification, Cohen’s or Fleiss’ kappa. You see quality as numbers you can plot over the project.
What does it cost?
Priced per study or per label depending on the task, with the schema and pilot work as a fixed fee. A single-label classification pass and an adjudicated multi-organ segmentation set are different projects at different prices, and we quote them separately after the scoping call.
Tell us what the model needs to learn to see.
Bring a label schema or just the clinical question. We will propose the schema, the readers and the quality bar, and annotate a pilot before you commit to volume.
Or email [email protected] · Mohali, India