Milvue

Deep learning detection and measurement on radiographs is the whole product. The Milvue Suite reads X-rays and returns findings, measurements, and a draft report; there is no scanner, PACS, or services layer of its own. One useful piece of candour: the company labels its report generation module as subscription software and explicitly NOT a medical device, separating the regulated detection engine from the unregulated reporting convenience layer, which is a distinction most vendors blur.

Detection plus a pre-filled draft report that the radiologist validates, with the company stating plainly that the user validates the output. This sits a step further toward automation than pure detection overlay, since generating draft report text shapes what the reader writes, but it stops well short of autonomous reporting. The clearance basis is a reader study measuring physician performance with the tool, which is the correct framing for an assistive device.

Headline performance is published, ROC-AUC above 96 percent for fractures and a reader study across 24 physicians of varying expertise reporting sensitivity improvements of 20 percent for fractures and 18 percent for elbow joint effusions plus reduced interpretation time.

But the aggregate AUC figure is not broken out by pathology, anatomical region, or patient age, and the claim to be number one in adult and pediatric fractures is not substantiated against any named comparator or third party benchmark. Training data composition and architecture are not published.

The buyer's own deployment choice determines their exposure here, which is unusual enough to state as the finding. An on premises option exists, and where the software runs inside the hospital's own estate images never leave, so the entire question of vendor retention and secondary use becomes moot for that site.

This index treats that architectural choice as decisive elsewhere in imaging, where lanes split on whether the study leaves the building, and it is the cleanest answer a vendor can give. It is available here and it is not the default answer to this axis, because it only holds for sites that take it.

Under the cloud option none of it applies and nothing is published: no retention of radiographs or derived outputs, no de identification statement, no position on whether customer images inform model development, no deletion term at contract end, and no model, hosting arrangement or sub processor list.

Given reported analysis of more than twenty five million radiographs in a single year across hundreds of sites spanning five continents, the cloud path carries a very large volume of studies under terms a reader cannot see, and the scale makes the training question pointed rather than theoretical, since a corpus of that size is exactly what a detection vendor would want. Ask which deployment applies to you, and under cloud what is retained, whether images train models, and whether a site can decline.

Real evidence with real limits. The FDA clearance study involved 24 experienced physicians across musculoskeletal, radiology, and pediatric expertise levels, which tests the tool across the actual mix of readers who use it rather than radiologists alone. Peer reviewed work by Parpaleix and colleagues assessed the combined musculoskeletal and chest solution in an emergency setting. Deployment scale is substantial at a reported 25 million X-rays analysed in 2024 and 600 plus sites.

What is missing is outcome level evidence and independent head to head comparison against the several competing fracture detection vendors now indexed here, which matters given how crowded that specific niche has become.

Converted from Not Rated. No stewardship terms were located, and the architecture offers the strongest available mitigation.

Nothing was retrieved on retention of radiographs or derived outputs, de identification, whether customer images inform model development, or deletion at contract end.

The mitigation is the on premises deployment option, and it belongs on this axis as much as on deployment. Where the software runs inside the hospital's own estate, images never leave, and the entire question of vendor retention and secondary use becomes moot for that site. This index treats the same architectural choice as decisive elsewhere in imaging, where lanes split on whether the study leaves the building. It is the cleanest answer a vendor can give and it is available here.

Under the cloud option none of that applies and nothing is published. Given reported analysis of more than 25 million radiographs in a single year across 600 or more sites spanning five continents, the cloud path carries a very large volume of studies, and the terms governing it are not visible.

The scale also makes the training question pointed. A corpus of that size is valuable for model development, and whether any of it is used that way, and whether a site can decline, is not addressed.

Ask which deployment applies, and under cloud what is retained and whether images train models.

Converted from Not Rated. No business associate terms were located, and the deployment choice materially changes what the agreement has to cover.

No agreement, addendum, role statement, subcontractor flow down, breach notification timetable or review cadence was retrieved.

The role follows the deployment. Under the cloud option, radiographs leave the hospital and are processed by the vendor, which makes it a business associate holding protected health information, with direct liability under the rule. Under the on premises option the software runs inside the hospital's own estate, and the vendor's position narrows considerably, potentially to software supply and support access rather than data custody. Those warrant different agreements, and using one template for both is a common error worth avoiding.

The company also distributes through at least one third party imaging platform, which introduces the same chain question this index has raised for its French peers: whether the hospital contracts directly or through the platform determines who holds the instrument and who notifies on a breach.

Home jurisdiction adds the usual cross border layer, since a French manufacturer selling into United States emergency departments must establish which entity signs and where processing occurs.

Ask which agreement applies to your deployment model, what support access persists on premises, and whether the contract is direct or through a platform.

Converted from Not Rated. No independent security attestation was located across two differently phrased searches.

No SOC 2 of either type, no ISO 27001, no HITRUST, no trust centre and no penetration testing statement was retrieved.

What is published is the device dossier, and it is detailed: two United States clearances covering musculoskeletal detection and triage plus chest, Class IIa conformity assessed by a named notified body under the European medical device regulation, and a Canadian Class II licence. The existing note draws the right line and it holds. Notified body conformity assessment examines whether a device is safe and performs as claimed. It does not examine how a hospital's images are protected in transit, at rest or in the vendor's environment.

One architectural fact is more relevant to security here than for most peers, and it is credited on the deployment axis: an on premises option exists alongside cloud. Where a hospital runs the software inside its own estate, much of what an attestation would evidence becomes the hospital's own posture rather than the vendor's. That is a genuine risk reduction, and it is available rather than theoretical.

Under the cloud option none of it applies and nothing is published.

Ask what independent assessment exists for the cloud service specifically, and what the hospital retains responsibility for on premises.

Authorization across three jurisdictions with clearly delineated scope, which the company documents unusually precisely: the Milvue Suite is CE Class IIa under EU MDR 2017/745 via notified body BSI, TechCare Trauma is FDA cleared as a Class II device covering musculoskeletal detection and triage, a second FDA clearance covers chest, and Smart Trauma plus TechCare Metrics are Health Canada licensed Class II. Naming which specific product carries which authorization in which market, rather than claiming blanket global clearance, is exactly the disclosure discipline this axis rewards.

The clearance reader study spans physicians of differing expertise levels including pediatric specialists, and the product covers both adult and pediatric fractures, so age group generalization has been considered. No formal governance framework, post market monitoring commitment, or demographic subgroup performance analysis was located.

The reader study design is the right one and the reporting stops short of what a buyer needs. Testing across two dozen physicians of varying expertise is a better design than a study of experts alone, because the benefit of a detection aid is not uniform across skill levels and a tool that helps a trainee substantially may add little for a subspecialist, so a mixed panel produces a figure that means something for a real department.

Reported sensitivity improvements for fractures and joint effusions with reduced interpretation time follow from it. Held at C on three things. The aggregate discrimination figure is not broken out by pathology, anatomical region or patient age, and the age breakout matters more than usual here because paediatric bone is radiographically different from adult bone, growth plates are a well known source of false positives, and a product claiming both populations should show both.

The superiority claim to be first in adult and paediatric fractures is not substantiated against any named comparator or third party benchmark, so it invites a comparison nobody can check. And training data composition, architecture and any warranty, indemnity or remediation commitment are absent. Ask for the breakout by region and age, the false positive rate on growth plates, and the basis of the superiority claim.

Integration into the RIS and PACS is the delivery mechanism, with results and a pre-filled report returned inside 30 seconds into the systems radiologists already use. Availability through the Blackford imaging platform provides a second route for customers already running that marketplace, alongside the direct path. Support for 17 or more languages and deployment across five geographic regions indicates the integration generalizes beyond one national environment. Radiology workflow integration; no EHR integration claimed.

Better disclosed than most peers because the choice is explicit: the solution is available via secure cloud platform OR on-premises deployment. Offering a genuine on-premise option matters for institutions unwilling to send imaging outside their perimeter, and few competitors in this lane state it plainly. Specific tenancy and data residency terms for the cloud option are not published.

No pricing is published. Scale is disclosed rather than cost, 25 million X-rays analysed in 2024 across 600 plus sites, and a second procurement route exists via the Blackford platform with presumably different economics. The suite structure means a buyer should establish which modules are licensed, since detection, the 60 automatic measurements, and the report generation layer are distinct components with different regulatory status.

Broad within plain radiography and deliberately so: seven pathology families spanning trauma and chest, adult and pediatric, plus 60 automatic measurements covering full legs, hip and pelvis, foot, shoulder, spine, knee, vertebrae, EOS, and bone age. Serving radiologists, emergency physicians, and orthopedic specialists across private practices, university hospitals, and teleradiology groups. The measurement suite alongside detection is a genuine differentiator, since most competitors do one or the other. Confined to X-ray rather than spanning CT and MRI.