Milvue

Deep learning detection and measurement on radiographs is the product. The Milvue Suite comprises TechCare Alert, which detects fractures, dislocations, joint effusions, pneumothorax and nodules; TechCare Metrics, about 60 automated bone measurements including lower limbs, EOS, bone age and Cobb angles; and TechCare Report, which pre-fills the radiology report and which the company labels as not a medical device.

The report layer is built on the detection and measurement output, and there is no scanner, archive or services layer of Milvue's own; remove the models and nothing is left to sell. The US cleared devices, TechCare Trauma and SmartChest, are built on supervised deep learning according to Milvue's FDA filings.

Detection with the clinician deciding, plus a pre-filled report. The FDA 510(k) summary for TechCare Trauma (K242171, January 2025) describes a concurrent aid whose results are not intended for stand alone clinical decisions. It runs at two fixed thresholds, a high sensitivity DOUBT point and a high specificity YES point, and returns a study level result (suspected, doubt, not suspected or not processed) with solid boxes for YES findings and dotted boxes for DOUBT findings. The thresholds' sensitivity and specificity are published in the summary. TechCare Report then pre-fills the report for the radiologist, and the company reports 64 percent more exams read per session with it. A drafted report shapes what the reader writes, which moves this a step toward automation without making it autonomous.

The missing part is what happens after an error and how the pre-filled report is checked. Nothing published describes how a missed or false finding is reported and handled, or what review step sits between TechCare Report's draft and a signed report. The home page's "up to 90 percent fewer errors in the emergency department" is a performance claim without method, not a control.

The FDA 510(k) summary for TechCare Trauma (K242171, January 2025) names the approach and characterizes the training data: supervised deep learning trained on 95,266 radiographs covering all anatomical areas from multiple equipment manufacturers, patients aged 0 to 102 (mean 53.3). It describes two fixed operating thresholds and states that the test sites were independent of the training data.

What is not published is versioning and update practice: the summary carries no predetermined change control plan, and nothing states how models are retrained or how a customer learns that a new version is running. The home page's area under the curve above 96 percent for fractures is broken out in the summary by anatomical region and age band, with the lowest values at the pelvis (0.893 in adults) and in children under two (0.948).

The model layer is Milvue's own. TechCare Trauma (K242171) and SmartChest (K232410) are FDA cleared devices in Milvue's name, and the TechCare Trauma summary describes a supervised deep learning model trained by the company, so the models cannot be swapped for another supplier's without a new submission. The chain from model to product is short by construction.

The summary also states that the software is deployed on premises or in the cloud; where a hospital runs it on premises, studies need not leave the hospital. For the cloud option nothing is published about the hosting provider, region or subprocessors: the privacy policy on the site covers only the website and names Cloudflare, Google Analytics and HubSpot for that purpose.

The TechCare Trauma 510(k) summary (K242171, January 2025) reports four retrospective, fully crossed reader studies on sites independent of the training data, read with and without the aid after a washout of at least a month: adult fractures (304 cases), pediatric fractures (256), adult elbow joint effusions (109) and pediatric effusions (100). Aided readers' area under the curve rose in every study, from 0.865 to 0.955 for adult fractures and from 0.857 to 0.931 for pediatric fractures, with the gains statistically significant. The number of readers is not stated. Standalone tests used 4,109 adult and 2,872 pediatric radiographs from at least four US states and fourteen equipment makers, against a reference of three board certified radiologists per case.

These are retrospective test results, which show what the aid does on curated cases. The home page's headline outcome claim, up to 90 percent fewer errors in the emergency department, has no method, population or denominator beside it, and published work on the combined musculoskeletal and chest solution in an emergency setting (Parpaleix and colleagues) and a 2026 single center external validation of the TechCare Kids bone age model in the journal Diagnostics are the independent studies on record. A prospective evaluation in live emergency use is what the headline claim calls for.

Milvue publishes one fact about how studies move and nothing on what happens to them. The TechCare Trauma 510(k) summary (K242171) states that the software is deployed on premises or in the cloud; a hospital that runs it on premises keeps its studies inside its own systems, which removes most of the retention and secondary use question for that site.

For the cloud option nothing is published: no retention period for radiographs or outputs, no statement on whether customer studies are used to train or improve the models, and no de identification practice. The privacy policy on the site, updated in June 2026, covers only the website.

Safety engineering, meaning guardrails and how a safety event is handled, is not published either; the operating thresholds and the not for stand alone use labeling are oversight controls, described under autonomy.

No business associate status and no privacy document reaching the product are published. The privacy policy linked from the site footer, updated in June 2026, states that it covers the processing of personal data on the milvue.com website; it does not mention patient data, medical images or the products.

The legal notices page, in French and English, names Cloudflare as the website's host, states the company's ISO 13485:2016 certification and the products' Class IIa status under the EU Medical Device Regulation (notified body BSI, 2797), gives site users a route to their own personal data, and refers to the same privacy policy for how personal data is processed; it says nothing about patient data, health data hosting or the company's role for the service.

Milvue sells into the United States under two FDA clearances, and the role it takes depends on deployment: under the cloud option studies leave the hospital and Milvue would hold protected health information, while on premises its role may narrow to software supply and support access. Neither position is stated. A service privacy notice, a data processing agreement or a business associate agreement would carry this to C or above.

Milvue publishes no information security certification and no outside check of its security. The site footer carries a legal notice, the privacy policy and declarations of conformity, and no security or trust page; the legal notices page names Cloudflare as the website's host and states ISO 13485:2016 certification, and says nothing else on security.

The privacy policy states only that necessary technical and organizational measures are taken, and the TechCare Trauma 510(k) summary carries no cybersecurity statement. The device certifications, CE Class IIa under the EU Medical Device Regulation through the notified body BSI (2797), the FDA clearances and a Health Canada Class II license, and the ISO 13485 quality system behind them review the device, not how the company operates information security. Controls asserted with nothing independent behind them grade here. The on premises option shifts much of the security posture to the hospital for sites that take it.

Two FDA 510(k) clearances are listed under the applicant Milvue. SmartChest, K232410 (10 May 2024, 21 CFR 892.2080, product code QFM), is chest X-ray triage and notification software. TechCare Trauma, K242171 (17 January 2025, 21 CFR 892.2090, QBS, predicate Gleamer BoneView K222176), aids detection of fractures on 2D radiographs across 13 anatomical regions from birth onward, and of elbow joint effusions from age two, for clinicians in primary care, emergency, urgent care and specialty settings, and is not for stand alone decisions.

Outside the US the Milvue Suite is CE Class IIa under the EU Medical Device Regulation through the notified body BSI (2797), and detection and measurements hold Health Canada Class II licenses. What is sold but not cleared in the US: dislocation detection in TechCare Alert, which K242171 does not cover, and any chest findings beyond SmartChest's cleared triage indications; TechCare Metrics, for which no FDA clearance under Milvue is listed; and TechCare Report, which the company labels as not a medical device.

The TechCare Trauma 510(k) summary (K242171, January 2025) publishes standalone results compared across the groups the product serves, with numbers. Fracture detection area under the curve by age band: 0.948 under two years, 0.971 from two to under twelve, 0.955 from twelve to under eighteen, 0.965 from eighteen to under twenty two, 0.972 from twenty two to under sixty five, and 0.950 at sixty five and over.

Results are also given by sex and by imaging manufacturer, and by anatomical region, where fracture detection ranged from 0.893 at the adult pelvis to 0.998 at the clavicle. A buyer can see where the model is weakest. Results by ethnicity are not reported for the standalone tests, and the reader studies' subgroup analyses by age, sex, ethnicity, state, fracture displacement and manufacturer are described without their figures. The home page's claim of an area under the curve above 96 percent across all age groups is a parity claim that the summary's own figures for infants and patients 65 and over contradict.

TechCare Trauma's error rate is on the record in Milvue's own filing. The FDA 510(k) summary (K242171, January 2025), which Milvue prepared, reports standalone performance of the supervised deep learning model at its two operating points, against a reference of three board certified radiologists per case (musculoskeletal radiologists for adults, pediatric radiologists for children) by majority vote.

On 4,109 adult radiographs, fracture sensitivity was 0.906 (0.891 to 0.919) and specificity 0.900 (0.887 to 0.912) at the high specificity point, and 0.940 and 0.845 at the high sensitivity point. On 2,872 pediatric radiographs the figures were 0.900 and 0.927, and 0.930 and 0.875. Elbow effusion detection was weaker in children, with sensitivity of 0.669 at the high specificity point on 483 radiographs. The split between positive and negative cases is not stated. These are test set results for the cleared indications, not performance in use, and they cover TechCare Trauma only. No warranty, indemnity or remediation commitment attaches to the output.

TechCare Trauma connects to DICOM sources, the archive or the imaging device, and returns annotated secondary capture images and optional DICOM structured reports listing findings, according to its FDA 510(k) summary (K242171); the company states results and a pre-filled report reach the reading workflow in under 30 seconds and that its products run in more than 600 sites.

Milvue is also listed in the Blackford Analysis imaging AI portfolio, a named orchestration platform that gives a second route into an existing workflow. That is standards support with a named platform behind it. No integration with a named record system, RIS or archive vendor is published, and no HL7 or FHIR path is described.

Two deployment options are stated and nothing about where the cloud option runs. The TechCare Trauma 510(k) summary (K242171) describes the software as deployed on premises or in the cloud. For the cloud option no hosting provider, region, residency commitment or tenant isolation is published, and the site's privacy policy covers only the website.

An on premises option matters for institutions that keep imaging inside their own perimeter, and for those sites residency is the hospital's own. For everyone else the location of processing is implied rather than committed. Milvue is a French company selling in Europe and the United States, so where a US or EU study is processed decides which rules apply.

No price, unit of charge or tier is published. The module structure is public, though. TechCare Alert handles detection, TechCare Metrics handles measurements and TechCare Report produces the prefilled report, and the three carry different regulatory status. There is also a second route to purchase through the Blackford Analysis platform.

The company's figures, more than 600 sites in more than 25 countries and 40 million analyses a year, describe use, not price, and its 100 percent client recommendation rate is a survey claim with no method stated.

Coverage is named, with validation behind the fracture product. TechCare Trauma is cleared for fractures across 13 anatomical regions from birth onward and for elbow effusions from age two, for clinicians in primary care, emergency, urgent care and specialty settings, and its 510(k) summary (K242171) validates it on US data from at least four states by age band, sex, manufacturer and region.

TechCare Alert also sells dislocation, pneumothorax and nodule detection, TechCare Metrics covers about 60 measurements including bone age and EOS, and the company's site lists a TechCare CT-Scan product, so the suite reaches beyond what the fracture summary validates. The home page's "fracture detection validated with no age limit" matches the cleared indication from birth. Validation by care setting is not shown.