AZmed

Deep learning detection is the product. AZmed sells two suites: Rayvolve for X-ray, with AZtrauma for fractures, dislocations and joint effusions, AZchest for chest findings, AZmeasure for osteoarticular measurements and AZboneage for bone age; and Rayscan for CT, with AZnod for lung nodules. Each module is a model, and there is no hardware, archive or services layer sold alongside; remove the models and nothing is left to sell. Rayvolve's FDA 510(k) summary (K260378, May 2026) describes supervised deep learning that detects and localizes findings on osteoarticular X-rays.

Rayvolve is an aided diagnosis device and the reader keeps the diagnosis. Its FDA 510(k) summary (K260378, May 2026) states that it does not operate autonomously, that users must not use its output as the primary interpretation, and that only the users' diagnosis remains valid for the patient. It marks suspected fractures, dislocations and effusions with bounding zones and, when it finds nothing, places a sentence at the top of the image saying so.

That negative statement is where a busy emergency reader could lean on the software, and the summary does not publish the operating threshold behind it or how often a study marked as clear held a finding. The structure is described and that part is missing, as is any route for reporting a missed finding. The home page's 99.7 percent negative predictive value and 67 percent reduction in false negatives for AZtrauma are performance claims without method.

The approach is named and little else. Rayvolve's FDA 510(k) summary (K260378, May 2026) describes supervised deep learning that detects and localizes findings on CR and DX radiographs received from DICOM nodes, and it lists what changed from the previous version (K240845): spine and ribs added, and dislocations and joint effusions added to fractures.

It does not characterize the training data, give a version number, or describe how models are updated, and it carries no predetermined change control plan. Reader improvement figures from published studies describe the product's effect, not what produces the output.

The model layer is AZmed's own. Six FDA 510(k) clearances in AZmed's name cover Rayvolve's fracture, dislocation and effusion detection and its chest modules (K220164, K240845, K260378, K243808, K243831 and K261378), and the K260378 summary describes a supervised deep learning model built by the company, so the models cannot be swapped for another supplier's without a new submission. That keeps the chain short at its core, but the rest of it is not named.

The summary says Rayvolve is designed to run on a cloud platform connected to the radiology center's local network, and the home page describes SOC 2, GDPR and HIPAA compliant infrastructure, but no hosting provider, region or subprocessor is named, and the privacy policy covers only the website.

The evidence behind the latest clearance is retrospective. Rayvolve's FDA 510(k) summary (K260378, May 2026) states that no clinical studies were conducted for the submission; it reports a standalone test on 6,246 radiographs and two retrospective multiple reader studies of dislocation detection (178 and 225 radiographs, 20 readers, half musculoskeletal radiologists), in which aided readers' area under the curve rose from 0.899 to 0.945 and sensitivity from 0.849 to 0.915, with reading time down 14 percent.

Earlier clearances rested on retrospective studies at named US institutions: University Hospitals in Cleveland for adults, where aided sensitivity rose from 0.865 to 0.955 with specificity nearly unchanged (0.826 to 0.831), and SimonMed Imaging for children. A 2026 multicenter analysis by Cohen and colleagues covers 258,373 X-rays from 100 centers in 26 countries.

The home page's headline claims are in use outcomes, a 67 percent reduction in false negatives, 83 percent faster turnaround and 99.7 percent negative predictive value for AZtrauma, with no method, population or design beside them. The company describes no prospective evaluation in live use behind those claims.

AZmed offers general assurances and nothing on what happens to the radiographs it processes. The home page describes SOC 2, GDPR and HIPAA compliant infrastructure. The privacy policy, approved in December 2025, covers the website, billing, prospecting and recruitment, with retention periods for those records only; it does not address images or outputs processed by Rayvolve or Rayscan.

No retention period for studies, no statement on whether customer images are used to train or improve the models, and no de identification practice is published. The question is concrete here because the company's published validation draws on hundreds of thousands of radiographs from many institutions; whether routine customer studies feed model development, or research runs under separate agreements, is not stated. Safety engineering, meaning guardrails and how a safety event is handled, is not published; the aided diagnosis labeling is an oversight control, described under autonomy.

Compliance is claimed and nothing supports it. The home page describes AZmed's infrastructure as HIPAA compliant, but no business associate status, agreement, tier or subcontractor terms are published, and the only privacy document, approved in December 2025, covers the website and business records, not patient data processed by the products. A compliance claim with no underlying document grades here.

AZmed sells into United States emergency and outpatient imaging under six FDA clearances, including pediatric fracture detection, so a hospital or imaging center will need a business associate agreement.

AZmed's home page describes its infrastructure as SOC 2 compliant, which points to an outside audit, but it does not say whose report it is, its type or its date, and a SOC 2 attached to the hosting infrastructure is the host's rather than an audit of how AZmed operates security. No SOC 2 report, ISO 27001 certificate or trust center is presented as AZmed's own, and the privacy policy states only that appropriate technical and organizational measures are taken. The ISO 13485 and MDSAP certifications the company names cover its quality management system for devices, not information security, and the FDA summaries carry no cybersecurity statement.

Six FDA 510(k) clearances are listed under the applicant AZmed, all Class II radiology software. Rayvolve for fractures: K220164 (2 June 2022, 21 CFR 892.2090, product code QBS) and K240845 (17 July 2024). Rayvolve K260378 (12 May 2026) extends detection to dislocations (shoulder, elbow, fingers) and joint effusions (elbow, forearm, arm) and adds the spine and ribs, for adults and children aged two and over, used by radiologists and emergency physicians as an aided diagnosis device.

Rayvolve PTX-PE, K243808 (21 March 2025, 892.2080, QFM), triages pneumothorax and pleural effusion; Rayvolve LN, K243831 (26 March 2025, 892.2070, MYN), detects lung nodules; and Rayvolve AZchest, K261378 (28 August 2026), detects and localizes chest findings. The company counts eight FDA cleared findings. Its quality system is ISO 13485 and MDSAP certified, and its products are CE marked. Sold but not cleared in the US: AZmeasure and AZboneage on X-ray and the Rayscan AZnod CT product, which carry CE marking only.

Rayvolve's FDA 510(k) summary (K260378, May 2026) publishes standalone results by anatomical region and finding with numbers: sensitivity from 96.1 percent at the thoracic spine to 97.8 percent at the ribs, and specificity from 85.1 percent at the lumbar spine to 89.3 percent for arm effusions.

It states that performance was evaluated across patient age, position, sex and X-ray machine and was similar in adults and children aged two and over, but publishes no figures for those groups, so where the model underperforms by age, sex or equipment cannot be seen. The groups are disclosed and only the conclusion is published. Ethnicity is not addressed. Multicenter validation across many countries, such as the 2026 analysis of 258,373 X-rays, tests performance where the product happens to be used rather than within defined patient groups.

Rayvolve's error rate is on the record in AZmed's own filing. The FDA 510(k) summary for Rayvolve (K260378, May 2026), which AZmed prepared, reports standalone performance of the supervised deep learning model on 6,246 radiographs: case level sensitivity of 96.88 percent (Wilson interval 96.10 to 97.51) and specificity of 87.29 percent (86.21 to 88.30), with results by region and finding for the spine, ribs, dislocations and effusions.

The split between positive and negative cases and the reference standard are not stated. These are test set results for the cleared indications, not performance in use, and they describe the fracture, dislocation and effusion product only. No warranty, indemnity or remediation commitment attaches to the output.

Rayvolve receives CR and DX radiographs from DICOM nodes and returns annotated results into the reader's existing console, running on a cloud platform connected to the radiology center's local network, according to its FDA 510(k) summary (K260378).

The company reports use at more than 2,500 sites in more than 55 countries and 15.4 million X-rays analyzed in the last twelve months, and its clearances rested on studies at named US institutions, University Hospitals in Cleveland and SimonMed Imaging. It has also been distributed through third party cloud radiology workflow platforms, including RADIN. That is standards support with named deployments behind it. No integration with a named record system, RIS or archive vendor is published, and no HL7 or FHIR path is described.

Rayvolve is designed to run on a cloud platform connected to the radiology center's local network, and its FDA 510(k) summary (K260378) also describes on premises, cloud and secure local processing of DICOM images. Where the cloud runs, who hosts it and where studies rest are not published: no provider, region, residency commitment or tenant isolation is named, and the privacy policy covers only the website.

The home page's SOC 2, GDPR and HIPAA compliant infrastructure says nothing about location. AZmed is a French company processing US and European studies, so the region decides which transfer rules apply.

No price, unit of charge or tier is published. What is public is the module structure: the Rayvolve X-ray suite (AZtrauma, AZchest, AZmeasure, AZboneage) and the Rayscan CT suite (AZnod), which carry different regulatory status in the US. The software has been sold both directly and through cloud radiology platforms. The company's scale figures, more than 2,500 sites in more than 55 countries, describe use, not price.

Coverage is named, with validation behind the cleared X-ray products. Rayvolve is cleared for fractures across the limbs, pelvis, hip, spine and ribs, dislocations at the shoulder, elbow and fingers, and effusions at the elbow, forearm and arm, in adults and children aged two and over, for radiologists and emergency physicians; its FDA 510(k) summary (K260378) validates the newest indications by region on 6,246 radiographs.

Chest coverage spans pneumothorax and pleural effusion triage, lung nodules and AZchest detection under separate clearances. Beyond X-ray, the Rayscan suite adds lung nodule detection on CT (AZnod, CE marked). Validation by care setting, emergency against outpatient imaging, is not shown, and the similar performance in children is stated without figures.