Challenge
A healthcare AI startup developing a diagnostic support tool for chest X-ray analysis (pneumonia, tuberculosis, pleural effusion, cardiomegaly) faced a critical annotation quality problem. Public datasets had two fatal flaws: annotation inconsistency creating noisy labels, and population gap lacking South Asian pathology variants most common in their target Indian hospitals. Without accurate annotation the model would have unacceptably high false negative rates - a life-or-death quality requirement. HIPAA-compliant PII redaction from DICOM metadata was required before annotation could begin.
Solution
eQOURSE assembled a specialist medical annotation team with 15 medical annotators with radiology or clinical backgrounds and 3 consulting radiologists providing senior supervision. Annotation Methodology: Semantic segmentation of lung fields, cardiac silhouette, and pathological regions on 25,000 chest X-rays. Instance segmentation for multi-lesion cases. HIPAA-aware PII redaction from all DICOM metadata before annotation. Gold-standard validation with every 5th image double-annotated independently. Annotation protocol fully documented for FDA submission purposes. All annotation performed using medical-grade tooling with full audit trail.
Results
Key outcomes: The diagnostic model achieved 94.7% sensitivity and 96.1% specificity - exceeding FDA submission thresholds. South Asian pathology variants improved model generalisation to Indian hospital deployments. The fully documented annotation protocol gave the client evidentiary records for regulatory submissions. The client is pursuing FDA 510(k) clearance. eQOURSE demonstrated that radiologist-supervised annotation produces measurably superior model outcomes.