Clinical AI now ships as a reproducibility engine

Clinical AI in radiology is no longer sold as a second opinion. It is sold as a reproducibility engine: software that converts a variable human read into a standardized one, and gets paid when the workflow actually uses it. DeepHealth's new breast ultrasound clearance shows what that mechanism looks like at scale.

The headline number, 37% less interpretation time, invites the usual read: AI helps radiologists work faster. The pattern underneath is different. Handheld breast ultrasound has always drifted between readers: one radiologist scores a lesion BI-RADS 3, the next scores it 4, and the referring clinic gets two different stories. By automating lesion detection and ACR BI-RADS characterization across 700,000 annual studies inside RadNet's outpatient network, the platform compresses a known clinical variance into a single reading convention. The 8% sensitivity lift rides on top of that, and so does the Category III CPT reimbursement that makes the whole thing financially legible to the operators running it.

The pattern is repeatable: when an imaging AI ships with a working reimbursement code, a multi-reader validation, and a distribution channel already wired to use it, the regulatory clearance is the receipt, not the news. The news is that operator-variability economics now have a price tag.