Expert-annotated medical imaging for a regulated pathway
A medical-imaging team required clinician-grade annotations with auditable provenance for a regulated submission.
The challenge and pain points
Labels needed defensible quality, documented reviewer credentials, and a chain of custody for audit.
- A regulated submission required labels with defensible quality, not just high accuracy on paper.
- Reviewer credentials had to be documented, because an auditor would ask who annotated each image and why they were qualified.
- Every label needed a chain of custody, so the dataset could be traced from source image to final annotation.
- General-purpose annotators could not sign off on clinical labels, so the work needed credentialed reviewers under a protocol.
- Any gap in provenance risked the whole dataset being rejected in review.
What Corpshore did
Credentialed reviewers annotated under a documented protocol with full provenance capture and tiered QA sign-off.
- 1Assemble a credentialed reviewer panel
Staffed reviewers with documented clinical credentials and recorded those credentials as part of the dataset record.
- 2Write the annotation protocol
Documented a protocol covering label definitions, edge-case handling, and sign-off, so every reviewer worked to the same defensible standard.
- 3Capture provenance per label
Recorded a chain of custody linking each annotation to its source image, its reviewer, and its QA sign-off.
- 4Apply tiered QA sign-off
Ran the three-tier QA cascade with an expert clinical audit tier, so labels were gated before entering the submission dataset.
- 5Package the audit trail
Delivered the dataset with its provenance and credential documentation assembled for regulatory review.
The solution
Corpshore treated provenance as a first-class deliverable, not an afterthought. Credentialed reviewers annotated under a written protocol, and each label carried a record of its source image, its reviewer, and its QA sign-off, so the dataset could be traced end to end.
The three-tier QA cascade included an expert clinical audit tier, which gated labels before they entered the submission set. This produced accuracy that was measured against gold references rather than asserted, with the credentials behind each sign-off on record.
The output was an auditable dataset that stood up to regulatory scrutiny: high-accuracy labels, documented reviewer qualifications, and a chain of custody an auditor could follow. Specific submission details remain confidential to the client.
Results
The engagement produced a clinician-grade dataset with 97%+ accuracy against gold references, documented reviewer credentials, and a full chain of custody from source image to final label. Because provenance was captured per label rather than reconstructed later, the dataset was defensible in regulatory review. The client and submission are confidential, so outcomes are described qualitatively and no exact submission metrics are shown.
| Metric | Result | Notes |
|---|---|---|
| Reviewers | Credentialed | Qualifications recorded per dataset |
| Accuracy | 97%+ | Measured against gold references |
| Provenance | Full audit trail | Source image to final label |
| Protocol | Documented | Single defensible standard |
| QA sign-off | Tiered clinical audit | Expert tier in the cascade |
| Regulatory readiness | Auditable | Stood up to review |
| Cost vs US-domestic | 50 to 70% lower | Methodology-level figure, not a client quote |
Illustrative cost index, Corpshore set to 100. The band reflects Corpshore's typical 50 to 70% cost advantage versus US-domestic providers. Representative of methodology, not exact client data.
Representative distribution of caught errors across the cascade, locking in 97%+ accuracy before delivery.
Client names and some figures are confidential. Where an exact client metric is not published, outcomes are described qualitatively and charts show representative or methodology-level data, including the three-tier QA cascade distribution and Corpshore's typical 50 to 70% cost advantage versus US-domestic providers.
Ready to scope a pilot?
Tell us your modality, volume, and languages. We'll return an indicative scope, timeline, and cost band.