RadiusHealth

Radiologists needed to compare current scans against historical cases with similar presentations, but the existing PACS system only supported DICOM metadata search. Finding a comparable prior case required manually scrolling through hundreds of studies. This slowed diagnostic workflows by an average of 22 minutes per case and reduced the number of second-opinion consultations by 40%, as the effort to find comparison cases was simply too high.

Mixpeek indexes diagnostic imaging studies by extracting visual embeddings from scan slices alongside structured metadata from radiology reports. A retriever pipeline enables radiologists to select a region of interest on a current scan and instantly retrieve the most visually similar historical cases, complete with their associated reports and outcomes. The system respects patient privacy through namespace-level isolation per hospital system.

Deployment followed a rigorous validation process required by RadiusHealth's clinical governance board. A pilot with 50,000 de-identified studies ran for 12 weeks to validate retrieval relevance against radiologist ground truth. After achieving 94% precision at k=5, the system was extended to all participating hospital systems. Batch ingestion of the full 2.6M study archive completed in 18 days.