Pixel-grade segmentation for diagnostic medical AI

A diagnostic-AI startup serving teleradiology networks across APAC needed clinician-grade lesion segmentation and structural tagging across dermatology and chest-imaging modalities, with full audit trails for regulatory submission. Their in-house clinical team could only sustain a fraction of the throughput their model roadmap required.

Off-the-shelf labelling vendors lacked the clinical depth – previous batches had a 14% rejection rate on QA review, eating margin and timeline.

We paired domain-trained annotators with a panel of practising radiologists and dermatologists who acted as second-pass reviewers. All work ran inside a HIPAA-aligned pipeline with PHI stripped at ingest and per-image lineage retained for regulator review.

Every study went through a three-stage workflow – primary annotation, peer adjudication, and clinician sign-off – with discrepancy logs captured for the client's SaMD documentation.

Delivered 220K studies with 100% clinician sign-off and zero rejected batches over the 8-month engagement.

The client's lesion-detection model gained 6.8 mAP on the held-out validation set after the first 30K studies landed, and the audit trail satisfied a TGA pre-market review on first submission.