--- title: "Medical Imaging Annotation by Clinical Readers" url: https://ypai.ai/solutions/healthcare/medical-imaging-annotation/ description: "Radiology, pathology, dermatology and ophthalmology annotation by credential-verified clinical readers, with consensus, adjudication and a QA record." source: "src/copy/routes (route /solutions/healthcare/medical-imaging-annotation/)" --- # Every label on the study is a clinical opinion. Ours come signed. > Radiology, pathology, dermatology and ophthalmology annotation by credential-verified clinical readers, with consensus, adjudication and a QA record. Radiology, pathology, dermatology and ophthalmology studies annotated by credential-verified clinical readers, under a task protocol agreed before the first study is opened. Two readers, a measured agreement, adjudication where they differ, and a QA record that ships with the labels. ## Three questions decide whether a label set can carry a model. A label is a clinical judgement drawn onto a study. What matters is who drew it, whether a second reader agreed, and whether the record of both survives to the audit. ### Who drew the line? Radiology reviewer · credential verified for this engagement reader 1 · CT chest · nodule - Labels for a clinical model are drawn by people qualified in the modality and the task. Reader credentials are verified for the engagement, matched to the domain the study belongs to, and written into the record with every label they sign. ### Did a second reader agree? Dice 0.53 · below threshold · to adjudication MR brain · lesion core · reader 1 accepted the proposal - Every study is read twice, the second reader blinded to the first and to any model proposal. Agreement is measured as the overlap between the two drawn structures (the Dice score), and a reading below the threshold goes to a sub-specialty adjudicator. ### Can we show the auditor? Task protocol · v1.2 · label set 1.2 QA record · one line per study - The task protocol, the acceptance threshold, both readers' roles, the agreement, the route and the adjudication are logged per study. That record is part of the delivery, in the form a device or SaMD documentation review asks for. ## Four annotation engagements, scoped to one modality and one task. Each is independently purchasable and configured per engagement. Four things are settled before the first study is opened. - The task protocol and the structure to be drawn - Who reads, and the credential verified for them - The acceptance threshold and the adjudication path ### Radiology annotation - CT, MR and X-ray studies segmented, detected and classified by radiology readers. - Segmentation, detection and classification on DICOM series, drawn by credential-verified radiology readers under your task protocol. Model proposals can be shown to the first reader and hidden from the second, so the label set measures the readers, not the model. - **Modality**: CT · MR · X-ray · DICOM in, DICOM or NIfTI out - **Reader**: confirm. Radiology reviewer · credential verified - **Proposal**: Shown to reader 1 only, if used Labels signed · agreement measured per study ### Digital pathology annotation - Whole-slide regions and cells outlined by pathology readers, at the magnification the task needs. - Tumour regions, tissue classes and cell-level marks on whole-slide images, read by credential-verified pathology readers. Region boundaries are the usual source of disagreement, so adjudication is designed in from the start. - **Material**: Whole-slide images · H&E and IHC - **Reader**: confirm. Pathology reviewer · credential verified - **Boundary**: flag. Disagreement routed to adjudication Region labels with the adjudication kept ### Dermatology and ophthalmology annotation - Skin lesions, fundus photographs and OCT scans labelled by readers in the specialty. - Lesion outlines and classifications on dermoscopy and clinical photographs, disc, cup and vessel structures on fundus images, layer boundaries on OCT, read by credential-verified dermatology and ophthalmology readers. - **Material**: Dermoscopy · fundus · OCT - **Reader**: confirm. Specialty reviewer · credential verified - **Classification**: Against the scheme in the protocol ### Protocol, consensus and QA design - The task protocol, the reader roles, the threshold and the record, written before a study is opened. - The engagement starts with the protocol: the structure to draw, the reference cases that calibrate readers, the acceptance threshold, the adjudication path and the QA record's fields. It is the document your regulatory reviewer reads first. - **Calibration**: Reference cases read before the set - **Threshold**: Agreement floor set per task - **Record**: confirm. One line per study, in the delivery Agreed before the first study is opened ## The modality decides the reader. A study is only labelled by someone qualified in its modality. Credentials are checked for each engagement. Who reads it · verified for the engagement Reader roles are matched per engagement from our clinical reviewer network, and the engagement names the readers it verifies. - Cross-sectional and plain-film studies read by a radiology reviewer, against the task protocol and the reference cases agreed for the engagement. - Slides read by a pathology reviewer at the magnification the task names, with region boundaries routed to adjudication when readers differ. - Echocardiography frames and rhythm strips read by a cardiology reviewer, against the clinical criteria written into the protocol. - Fundus photographs and OCT scans read by an ophthalmology reviewer, disc, cup, vessel and layer structures drawn to the protocol. - Dermoscopy and clinical photographs read by a dermatology reviewer, lesions outlined and classified against the scheme in the protocol. ## Two readers, one label. The agreement is measured, not assumed. Try the two-reader demonstration. You are reader 1: a model proposal is on the study, and you rule on it. Reader 2 has already read the same study blind. The overlap between your contours, the Dice score, is computed in the page and decides whether the label locks or goes to adjudication. The studies are synthetic and the geometry is simplified, so every part of the mechanism is visible. Synthetic studies and a toy geometry, so the mechanics are visible. Not patient data. Nothing on this bench is a result of ours. ### Task protocol - **Readers**: Two per study, blinded to each other - **Proposal**: Model contour shown to reader 1 only - **Acceptance**: Dice at or above 0.80 between readers - **Below threshold**: Sub-specialty adjudicator rules; both contours kept - **Record**: One line per study, in the delivery Read the study blind, before you opened it Rules when readers differ, on both contours - Axial · 1 mm · lung window - Whole-slide image · 20x · H&E - Invasive tumour region on the slide ### The second reader is blind. - Reader 2 never sees reader 1's contour or the model's proposal, so the agreement measures two clinical judgements, not one judgement copied. ### The threshold is written down. - The protocol names the agreement floor for the task. A reading below it goes to a sub-specialty adjudicator, and both contours stay in the record. ### The record is per study. - Reader roles, action, agreement, route, adjudication and label version are logged for every study, so a label can be traced to the people and the protocol behind it. ## The QA record an engagement delivers. Every study with its task, its protocol version, both readers, the agreement between them, the route the label took and the version it landed in. The example below fills in as you use the demonstration. Synthetic QA record · filled in from the demonstration ## Start with one modality, one task, one bounded set. A pilot is the same engagement at its smallest useful size. The protocol is written and the readers calibrated on reference cases before the first study is opened, a bounded set is read twice, and the pilot ends with a label set, its QA record and a decision. You judge it against your own model programme and your documentation review. - The modality, the task and the structure to be drawn - The reference cases that calibrate the readers - What the QA record must contain - The task, the structure, the threshold and the record's fields, written and agreed. - Readers read the reference cases first, so the set starts from a measured baseline. - Every study read twice, blind, with any model proposal shown to reader 1 only. - Readings below the threshold ruled on by a sub-specialty adjudicator, both contours kept. - The label set, the QA record and a decision on your own criteria, ready for the next set. The pilot ends with a label set you can trace to its readers and its protocol. ## Studies are special-category data. They are handled that way. A Norwegian company under GDPR. Imaging studies are stored in Europe by default and processed in the EEA by default, under Article 9 controls and a privacy-preserving workflow, with the paperwork to prove it. US healthcare work is handled HIPAA-compliant. - **Jurisdiction**: Norway · GDPR-native - **Category**: Article 9 special-category handling - **Storage**: European by default - **Processing**: EEA by default - **Studies**: DICOM de-identified before reading, where the protocol requires - **Readers**: Credential-verified per engagement - **US engagements**: HIPAA-compliant handling - **Contracts**: Standard DPA terms · SCCs available - **Erasure**: 30-day end-of-contract SLA - [The healthcare hub](https://ypai.ai/solutions/healthcare/) - [Clinical data collection](https://ypai.ai/solutions/healthcare/custom-data-collection/) - [Clinical AI systems](https://ypai.ai/solutions/healthcare/clinical-ai-systems/) - [Clinical model evaluation](https://ypai.ai/solutions/healthcare/clinical-model-evaluation/) - [Clinical text and coding](https://ypai.ai/solutions/healthcare/clinical-text-and-coding/) - Medical imaging annotation services for AI training ## Start with one modality, one task, one protocol. We scope against the model the labels will train or test, write the protocol before a study is opened, and test the method on a bounded first set. Scope · Protocol · Calibration · First set · Delivery A bounded first set carries its own acceptance criteria and its own QA record. You judge it against your own programme. Briefs are treated as confidential. We are used to studies that cannot leave the EEA and programmes that cannot be named. Tell us the modality, the task and what the labels will train or test. We reply with a feasibility read. We come back with a feasibility read on the modality, the task and the reader requirements. Enquiry details are treated as confidential.