FDA clears updated Cleerly LABS and PLAQUE algorithm trained on CONFIRM2 data

Cleerly, a cardiovascular imaging analysis company, announced FDA clearance of an updated version of Cleerly LABS and the newest generation of its PLAQUE algorithm, which quantifies and classifies coronary plaque from CCTA scans. Axis Imaging News reported the updated algorithm is trained on data from the CONFIRM2 Registry: roughly 6,550 individuals from 46 countries, with close to even female representation. Both products rolled out automatically to existing customers in November, and clinics don't need to do anything on their end.

Michal Sofka, Cleerly's chief technology officer, said the clearances "reflect our commitment to building tools that perform well across the patients clinicians actually see." That framing matters more than it sounds. A plaque-quantification algorithm trained on a narrow population can under- or over-call risk in patients it wasn't validated against, and closing that generalization gap is exactly what regulators and clinicians have been pushing imaging-AI vendors to do.

Shipping a population-diversity retrain as a seamless, zero-action update to every existing customer is itself a non-trivial engineering problem for a cleared SaaS platform. It's the kind of silent, backward-compatible model swap that clinical software teams have to architect for from day one, not bolt on after a submission.

Why a silent model update needs its own clearance

Retraining an FDA-cleared algorithm on a new dataset is not a routine software patch — any change that could shift a cleared algorithm's outputs typically requires its own regulatory submission, even when the clinical workflow around it looks unchanged to the end user. That's why Cleerly's update shipped as a formal 510(k)-cleared release rather than a quiet model swap: the underlying math changed, so the clearance had to change with it.

For vendors building this kind of update, the architecture decision that matters most is keeping the cleared inference layer cleanly separated from the delivery and update mechanism. A platform that can push a newly cleared model version to every customer automatically — without each clinic re-installing software or re-validating a local deployment — has to be built that way from the start: versioned model artifacts, feature-flagged rollout, and a release pipeline that treats "new cleared algorithm" as a first-class deployment event, not a hotfix.

What this means for imaging-AI vendors building their next submission

Diverse-population training data is becoming table stakes for FDA clearance of plaque, stenosis, and other quantitative imaging algorithms, not a differentiator. Vendors scoping their next submission should plan the update-delivery architecture alongside the clinical validation plan — a cleared algorithm that can't be redeployed cleanly across an installed base turns every future retrain into a fresh integration project instead of a scheduled release.