Infervision

Infervision's products are built on its own models, and the line now carries more than the models. The InferRead detection products (lung nodules on chest CT, chest radiographs, coronary CTA, hemorrhage triage on head CT, chest fractures and pneumonia on CT) are the models themselves, and the lung nodule product's 510(k) summary (K240554, May 2025) states that it uses deep learning to perform nodule detection.

Alongside them the company sells the InferOperate suite of 3D reconstructions for thoracic, liver and urologic surgical planning, InferCare for patient management and image follow up, including RECIST tumor assessment, and two platforms, InferScholar for medical research and InferMatrix for medical applications.

The surgical planning, follow up and platform products carry workflow value beyond any single detection model, so the models are the engine of core modules rather than the whole of what is sold. What InferMatrix and InferScholar run besides Infervision's own models is not described on the home page.

The cleared lung nodule product is an aid, not a reader. Its 510(k) summary (K240554, May 2025) indicates InferRead Lung CT.AI for detecting pulmonary nodules of 4 mm and larger on chest CT in asymptomatic people aged 55 and over, with both lungs in the field of view; it places detection marks as an overlay on the original CT, provides adjunctive information, and is not intended to be used without the original series, so the radiologist reads every study.

The stroke product (K211179, 2021) is cleared under the FDA's triage and notification product code, which reorders a worklist without removing studies from it, and the surgical planning products produce reconstructions a surgeon reviews. That is an oversight structure with the limits of use published.

The missing part is the operating point: the K240554 summary does not state the detection threshold, the false positive rate per scan it runs at, or what happens after a missed or false mark, and it does not say whether the marks are meant for concurrent or second reading.

The approach is named for the lead product and little else is. The InferRead Lung CT.AI 510(k) summary (K240554, May 2025) states that the product uses deep learning to perform nodule detection, with nodule segmentation, measurement, registration between prior and current scans, lobe localization and lung lobe segmentation alongside it, and describes the 2025 version as adding follow up comparison and lung analysis to the version cleared in 2020 (K192880), so a buyer can see that the product changed through a cleared update.

It does not characterize the training data (size, sources, sites, equipment makers or period), and no model versioning or update practice is published for any product in the line. The other products are described by their task only.

The model layer is short. InferRead Lung CT.AI is Infervision's own deep learning detection model, cleared by the FDA in 2020 (K192880) and again in 2025 (K240554), so a change to the model goes back through the company's own regulatory process rather than to an outside supplier; the stroke triage product, the InferOperate surgical planning suite and InferCare RECIST are Infervision's own filings too.

The K240554 summary says the software runs in the healthcare facility or on a cloud platform. For the cloud option no hosting provider, region or subprocessor is named, and the privacy policy on the site (last modified 1 August 2021) names only website analytics and social media parties.

Three Infervision entities appear as FDA applicants (Beijing Infervision Technology, Infervision Medical Technology and Beijing Infervision Healthcare Medical Technology), and which of them processes a customer's studies is not stated.

Evidence for the products in use is thin on the company's pages. The InferRead Lung CT.AI 510(k) summary (K240554, May 2025) reports standalone tests of its follow up functions on lung cancer screening scans: nodule matching across prior and current scans in 98 cases with 206 nodule pairs (0.970), lobe localization in 94 scans with 188 nodules (0.957) and lung lobe segmentation in 22 cases (Dice 0.966), with no reference standard stated, and says detection and segmentation were consistent with the 2020 version; it reports no reader study.

The company's research page lists more than 330 peer reviewed publications and conference abstracts, and the home page names none of them for a product sold. Deployment is reported at scale, including a US partnership with SimonMed and tuberculosis screening programs procured through international agencies in Eastern Europe and Central Asia and in Zimbabwe; the home page reports coverage of more than 30 countries. Breadth of use on that order is not, by itself, evidence of benefit.

Little is published on what happens to the studies Infervision's products process. The InferRead Lung CT.AI 510(k) summary (K240554) says the software can run in the healthcare facility or on a cloud platform, and that cybersecurity and vulnerability analyses found it conforms to the cybersecurity requirements, a general assurance about the product.

No retention period for studies or outputs, no statement on whether customer studies are used to train or improve models, and no de identification practice is published for any product. The privacy policy on the site (last modified 1 August 2021) covers visitors' contact and technical data and says nothing about medical images. Safety engineering, meaning guardrails and how a safety event is handled, is not published either.

The question has two settings: commercial deployments in the United States and Europe, and tuberculosis screening programs procured through international agencies, where the program rather than the individual decides and participants may have little practical ability to refuse a scan.

No business associate status and no privacy document reaching the product are published. The only privacy document on the site is a privacy policy last modified 1 August 2021, which covers visitors' names, contact details, IP addresses and browsing data and names website analytics and social media parties; it says nothing about medical images, patient data or the products, and its service terms are governed by Chinese law.

Infervision sells into the United States under six FDA clearances, a US imaging partnership with SimonMed is reported, and the lung product can run on a cloud platform, so a US customer using that option would send protected health information to the company. Three Infervision entities appear as FDA applicants, and which one would sign an agreement is not stated. Two legal frameworks with cross border provisions apply to a China headquartered processor of US imaging, so the agreement has to answer to both.

Infervision publishes no information security certification and no outside check of its security. The site footer links a privacy policy, service terms and an accessibility statement, and no security or trust page, and the privacy policy names no security measures or certification.

The InferRead Lung CT.AI 510(k) summary (K240554) states that cybersecurity and vulnerability analyses were conducted and that the device conforms to the cybersecurity requirements; that is the manufacturer's own premarket testing, reviewed as part of the device, not an audit of how the company operates security. Device authorizations examine whether the software is safe and performs as labeled, not how customer studies are protected. Controls asserted with nothing independent behind them grade here.

In the FDA database, the company's US position can be traced product by product. InferRead Lung CT.AI was cleared on 2 July 2020 (K192880, 21 CFR 892.2050, product code OEB) and again on 16 May 2025 (K240554): computer assisted reading tools that help the radiologist detect pulmonary nodules of 4 mm and larger on chest CT in asymptomatic people aged 55 and over, adjunctive to the original series, with follow up comparison, lobe localization and lung analysis added in 2025.

InferRead CT Stroke.AI was cleared on 12 August 2021 (K211179, 892.2080, product code QAS) for triage and notification. The InferOperate Suite was cleared on 15 September 2025 (K250237) and again on 7 July 2026 (K260973), and InferCare RECIST on 13 March 2026 (K252261), each under 892.2050, product code QIH. The applicants are Infervision group entities in Beijing (Beijing Infervision Technology, Infervision Medical Technology and Beijing Infervision Healthcare Medical Technology).

InferRead DR Chest, CT Coronary, CT Bone and CT Pneumonia are sold on the company's site and do not appear among its FDA clearances. Authorizations in Europe, the UK, Japan and China are reported for parts of the line and are not listed on the company's home page.

No results by patient group are published for any product. The InferRead Lung CT.AI 510(k) summary (K240554) breaks one result out, nodule matching by the interval between scans (0.976 at up to six months, 1.000 at six to twelve, 0.938 at twelve to twenty four), which is a technical condition rather than a patient group, and reports nothing by sex, age, smoking history, scanner maker or slice thickness. No governance framework or monitoring commitment is published.

The question matters because the products run in populations that differ in disease prevalence, body habitus, equipment and protocol, from US and European health systems to tuberculosis screening programs procured through international agencies.

Prevalence changes a detection model's positive predictive value even when sensitivity and specificity hold, so in a screening program where most people scanned are well, a shift in prevalence changes how many healthy people are pulled into follow up, and the screening settings are those with the least local capacity to notice a problem.

Mechanisms exist that let a reader challenge an output, and nothing stands behind it. InferRead Lung CT.AI places detection marks as an overlay on the original CT, provides adjunctive information and is not intended to be used without the original series, so the radiologist can check each mark against the images.

The 2025 510(k) summary (K240554) states that the product uses deep learning for nodule detection and reports standalone figures for other functions: nodule matching between prior and current scans (0.970, 98 cases, 206 nodule pairs), lobe localization (0.957, 94 scans, 188 nodules) and lung lobe segmentation (Dice 0.966, 22 cases, 110 lobes), with no reference standard stated; for detection it says only that results were consistent with the 2020 version.

So no error rate for the learned detection function is published on the company's pages or in that summary; the 2020 summary for the first version (K192880) is the other filing a buyer can read for detection figures. No warranty, indemnity or remediation commitment is published. Detection at 4 mm sits where benign findings outnumber malignant ones, so the false positive rate per scan is the figure a department needs.

Infervision's products work in the imaging workflow. The InferRead Lung CT.AI 510(k) summary (K240554) lists DICOM image input and reporting of results, with detection marks delivered as an overlay on the original CT. The product is reported in use at SimonMed, a US outpatient imaging provider, and in tuberculosis screening programs. That is standards support with a named deployment behind it. No PACS, viewer, RIS or record system is named, no HL7 or FHIR path for results is published, and no EHR integration is claimed.

Two options are stated and the location of neither is committed. The InferRead Lung CT.AI 510(k) summary (K240554) says the software can be installed in a healthcare facility or on a cloud based platform. No cloud provider, region, tenancy arrangement or residency commitment is published, and the privacy policy on the site does not address where patient data is processed.

The cross border question runs both ways for a China headquartered vendor: imaging from US and European health systems engages those buyers' own transfer rules, and China's data security and personal information protection laws impose their own requirements, including on transfers out. The on premises option resolves most of it for sites that take it.

No price, tier or unit of charge is published. The company is reported to offer several pricing options rather than a single model, and its tuberculosis screening deployments were procured through international agencies by public tender, where the price is disclosed to the purchaser rather than published. The products are sold directly and through public procurement, but what drives the cost is not stated.

Coverage is named product by product, with validation behind part of it. The lung nodule product is cleared for asymptomatic people aged 55 and over with nodules of 4 mm and larger on chest CT, and its 2025 summary tests its follow up functions on lung cancer screening scans; hemorrhage triage on head CT, surgical planning for thoracic, liver and urologic surgery, and RECIST tumor assessment carry their own US clearances.

The company also sells chest radiograph, coronary CTA, chest fracture and pneumonia products without a US clearance, and tuberculosis screening deployments in Central Asia, Eastern Europe and Zimbabwe, procured through international agencies, are reported. The research page lists more than 330 publications and abstracts, and which of them validate which setting is not stated on the home page; no validation is described for tuberculosis screening or for the uncleared products.