What is medical image annotation?
Medical image annotation is the labeling of scans and clinical pictures so that an AI model can learn from them. Annotators mark organs, structures and findings, and the model learns to detect, measure or classify the same things in new images.
The work is close to other image annotation, with two differences: the right answer comes from clinical knowledge, and the data is sensitive. Both shape how a project is run.
Images we work with
- X-ray and mammography
- CT and MRI series, usually delivered in the DICOM format
- Ultrasound
- Pathology slides
- Retinal, dental and skin images
- Endoscopy and surgical video
Types of labels
- Classification: a label for the whole image, such as normal or abnormal, or the presence of a finding.
- Bounding boxes: a rectangle around a finding, such as a nodule or a fracture.
- Segmentation: the exact outline of an organ, lesion or tumour, pixel by pixel.
- Landmarks: anatomical points used for measurement and alignment.
- Series labeling: the same structure followed slice by slice through a CT or MRI volume.
Accuracy needs a clinical standard
In most annotation, a careful person can see the right answer. In medicine it often takes training to see it. So the standard is set first:
- Your guideline is the reference. We label to your written definitions and example images.
- Clinical questions go to clinicians. Where a task needs clinical judgement, your clinicians, or the reviewers you appoint, set the reference standard and settle the difficult cases.
- Annotators are trained before production. They work on the guideline and a test set first, and they escalate doubts instead of guessing.
- Work is reviewed. Labels are checked in a second pass before delivery.
Privacy and confidentiality
- De-identified data. We ask for images with names, patient numbers and dates removed before they reach us.
- Confidential by contract. Our people work under non-disclosure agreements and strict data protection protocols, with data security and physical security measures to keep client information confidential.
- Your data, your project. Images are used only for the project they were sent for.
Where it is used
- Detecting findings to support diagnosis and triage
- Measuring organs, lesions and change over time
- Planning surgery and radiotherapy
- Counting and classifying cells on pathology slides
- Building research datasets
Annotation produces training data for software. It is not a diagnosis and it is not medical advice.
Quality you can measure
Pripton's AWS Marketplace listings report 98.5% average accuracy across our annotation work. If a project falls below 96.5% accuracy, we rework it.
Start a project
Tell us the image type, the labels you need and how the reference standard is set, through the contact page. A small pilot batch is the quickest way to agree the guideline.
