AZmed

Deep learning detection on plain radiographs is the entire company. The Rayvolve suite comprises AZtrauma and AZchest modules that read X-rays and return findings to the interpretation console, with no hardware, PACS, or services layer sold alongside. The company was founded in 2018 around this single technical proposition and has expanded by adding anatomical coverage rather than adjacent business lines.

Detection presented into the clinician's existing interpretation console, with the reader retaining the diagnosis. The design emphasis is explicitly on requiring no change in work habits, which is a deliberate low friction posture rather than an autonomy claim. The reader studies underpinning clearance measure performance WITH and WITHOUT the tool, which is the correct framing for an assistive device and implies the human remains the decision maker by design rather than by disclaimer.

Performance is published with unusual specificity for a company this size, and importantly it is reported as reader improvement rather than standalone algorithm accuracy. The University Hospitals Cleveland study reports sensitivity rising from 0.865 to 0.955 and specificity from 0.826 to 0.831 with AI assistance, accuracy improving 5.6 percent, and interpretation time falling 27 percent.

The pediatric study reports 96 percent sensitivity, 86 percent specificity, and AUC of 94 percent across 3,000 radiographs. Publishing the specificity figure that barely moved, alongside the sensitivity figure that moved a lot, is a more honest presentation than most vendors offer. Model architecture and training data composition are not published.

The company's own evidence base raises the question it does not answer, and the point is worth stating carefully because it cuts both ways. The published validation covers hundreds of thousands of radiographs across a hundred clinical centres in dozens of countries, alongside studies at named institutions, and assembling evidence at that scale requires access to very large volumes of clinical imaging from many organisations.

That access is a credit on the evidence axis, and it makes the provenance question sharper here rather than softer: what governed the use of those images, whether they came from customer deployments or from separate research agreements, and whether ordinary customers are contributing to the next such study by default.

The answer may be entirely clean, since multi site academic validation is normally run under its own ethics and data sharing arrangements distinct from commercial deployment, and nothing published says so, which is exactly the distinction a buyer needs. Nothing was retrieved on retention of radiographs or derived outputs, de identification, training use or deletion at contract end, and no model, hosting arrangement or sub processor list was named.

Operating under a comprehensive data protection regime imposes purpose limitation and lawful basis obligations that are real and enforceable and are not published as customer terms. Ask whether deployment images are separable from research use, and what is retained after a read.

Validation is externally generated rather than vendor run, which is the distinguishing feature. The adult clearance rested on a study at University Hospitals in Cleveland where board certified musculoskeletal radiologists read and annotated 2,626 X-rays of patients at four UH sites. The pediatric clearance rested on an independent study with SimonMed Imaging, one of the largest US outpatient imaging providers, across 3,000 pediatric radiographs.

A 2026 study by Cohen and colleagues evaluated the suite across 258,373 X-rays from 100 clinical centers in 26 countries, which is a scale of multi-site evaluation few vendors in this index can match. Separate published work covers external validation in children and thoracic pathology detection. Deployment spans a reported 2,000 plus facilities.

Converted from Not Rated. No stewardship terms were located, and the company's own evidence base raises the question it does not answer.

Nothing was retrieved on retention of radiographs or derived outputs, de identification, model training use or deletion at contract end.

The training question is unusually concrete here because of how the validation was assembled. The company points to a 2026 multi site analysis covering more than 250,000 radiographs across 100 clinical centres in 26 countries, alongside studies run at named United States institutions. Assembling evidence at that scale requires access to very large volumes of clinical imaging from many organisations. That access is a credit on the evidence axis and it makes the provenance question sharper here rather than softer: what governed the use of those images, whether they came from customer deployments or from separate research agreements, and whether ordinary customers are contributing to the next such study by default.

The answer may be entirely clean. Multi site academic validation is normally run under its own ethics and data sharing arrangements distinct from commercial deployment. Nothing published says so, and the distinction is exactly what a buyer needs.

The company operates under the European data protection regulation, which imposes purpose limitation and lawful basis obligations that are real but not published as customer terms.

Ask whether deployment images are separable from research use, and what is retained after a read.

Converted from Not Rated. No business associate terms were located in public materials.

No agreement, addendum, role statement, subcontractor flow down, breach notification timetable or review cadence was retrieved. A privacy policy exists and is framed under the European data protection regulation, which governs the company at home and is a different instrument from what a United States covered entity needs.

The role is not ambiguous. A hospital, emergency department or imaging centre is the covered entity. A vendor receiving radiographs and returning detections processes protected health information on its behalf and is a business associate with direct liability.

The setting sharpens why the terms matter. The products are sold into emergency departments, where studies are read under time pressure and often out of hours, and where the population includes children, since paediatric fracture detection is separately cleared. Paediatric imaging carries heightened sensitivity in most privacy frameworks, and nothing published addresses whether handling differs for it.

Cross border is the second layer. A French headquartered manufacturer selling into United States emergency care must establish which legal entity contracts, whether any processing or support access occurs outside the country of care, and how the two regimes are reconciled in one agreement.

Ask which entity signs, where images are processed, and whether paediatric studies are handled differently.

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. The company does publish a privacy policy framed under the European data protection regulation, which is its home obligation and a real framework, and which is not an information security attestation.

What is published instead is regulatory and evidential, and on the evidence side this company is unusually strong for its size. Clearances covering adult fracture detection since 2022 and paediatric since 2024, plus two chest clearances. Validation drawing on studies run by external institutions rather than the company itself, including a 2026 multi site analysis spanning more than 250,000 radiographs across 100 clinical centres in 26 countries. Independent external validation at that scale is rare in this index and it is credited on the evidence axis.

The distinction still holds and it is worth stating precisely because the record is otherwise so strong. Clinical validation establishes that the model reads images correctly. It says nothing about whether the infrastructure processing those images is secured, monitored or independently assessed. A vendor can be excellent at the first and unevidenced on the second, and this one is.

Ask what independent security assessment exists and whether one is planned for US procurement.

Multiple current US clearances across two anatomical domains: adult fracture detection cleared 2022, pediatric fracture detection cleared September 2024, and two further clearances for AZchest covering lung nodule interpretation plus pneumothorax and pleural effusion triage. The company states AZtrauma is the only FDA cleared solution in the United States covering fractures, joint effusions, and dislocations together, a scope claim worth verifying against a buyer's specific indications. CE marked since 2021. Both fracture clearances were supported by named external institutional studies rather than vendor internal data, which strengthens the regulatory record.

Stronger than most, expressed through validation breadth rather than a governance programme. The 2026 Cohen study spanning 258,373 X-rays across 100 centers in 26 countries directly tests generalization across geography, equipment, and practice patterns, which is the failure mode most imaging AI never examines publicly. Separate validation in a pediatric population addresses a distinct physiological group rather than assuming adult performance transfers. What is absent is a formal governance framework, post market monitoring commitment, or demographic subgroup breakdown.

Performance is published with unusual specificity for a company this size, and two choices in how it is reported lift it above the lane. The first is the form: results are given as reader improvement rather than standalone algorithm accuracy, with a study reporting sensitivity rising substantially with assistance alongside interpretation time falling, and a separate paediatric study reporting sensitivity and specificity across several thousand radiographs.

Reader improvement is the clinically relevant measure, because the product is used by a radiologist rather than instead of one, and a standalone figure describes a configuration nobody deploys. The second is what is included. Specificity barely moved while sensitivity moved a lot, and both are published.

A vendor selecting what to report would show the metric that improved and omit the one that did not, and publishing the flat figure alongside the moving one lets a reader see the actual shape of the benefit: the tool helps radiologists find more without helping them rule out more, which is exactly what a department needs to know when planning downstream workup capacity.

Held below the top grade because model architecture and training data composition are unpublished and no warranty, indemnity or remediation commitment attaches. Ask for performance broken out by fracture site and by patient age, and the false positive rate per study in your own population.

Results are delivered directly into the clinician's existing interpretation console with the explicit design goal of requiring no workflow change, and the suite is integrated into third party cloud radiology workflow platforms including RADIN. Deployment across a reported 2,000 plus facilities in more than 20 countries implies the integration is genuinely portable across varied PACS environments rather than bespoke. This is radiology workflow integration; no EHR integration is claimed or applicable.

Multiple delivery paths exist, direct deployment and distribution through cloud based radiology workflow platforms, which gives buyers routes but also means residency depends on which path is chosen. Specific tenancy, hosting, and data residency terms are not published, a live question for a European vendor processing US patient imaging.

No pricing is published. The company reports scale, more than 2,000 facilities and 2,500 plus radiologists and emergency physicians using the software in clinical routine across 21 or more countries, but no per exam, per site, or licensing rates. Availability through partner radiology platforms creates a second procurement route with likely different economics, also undisclosed.

Concentrated on plain radiography across two domains: musculoskeletal trauma covering fractures, joint effusions, and dislocations for both adults and children, and thoracic covering lung nodules plus pneumothorax and pleural effusion triage. Read by radiologists and emergency physicians in emergency, trauma, and outpatient imaging settings.

Deliberately X-ray centric rather than spanning CT and MRI, which distinguishes it from multimodality platforms, though the pediatric plus adult clearance combination gives it unusual population breadth within that scope.