AIRS Medical

The models are the product. SwiftMR is a deep learning algorithm that takes MRI images and returns them with less noise and more sharpness, so a site can scan faster or get better images from the same scan. SwiftSight produces quantitative brain MRI measurements. AIRS Medical sells no scanner or hardware; both products run behind scanners made by others.

Remove the models and nothing is left to sell. The gateway, routing and archive connections exist only to carry images to and from the models.

SwiftMR runs automatically on each MRI series and sends the enhanced images to PACS, where the radiologist reads them. The K253775 filing says the site sets the strength of processing: denoising from level 0 to 8 and sharpening from 0 to 5, used by radiology technologists. Clinicians remain responsible for all patient decisions. The original and processed series can be sent to the archive separately, so the unprocessed image stays available to check.

What is missing is the check on what the model changes. The radiologist reads an altered image rather than reviewing a finding, and nothing published says when the original should be compared or whether both series are kept by default. SwiftSight, a separate brain measurement product, promises on the homepage reproducible measurements across scanner vendors and field strengths, with no results published by vendor or field strength and nothing on whether follow ups should stay on the same scanner.

The method is described concretely. K253775 says SwiftMR uses a deep learning algorithm for noise reduction and a separate sharpening filter, each with set levels. It was validated on scans from seven scanner makers (Siemens, GE, Philips, Canon, Esaote, Fonar, Fujifilm) and field strengths from 0.25 to 3.0 tesla. The validation also covered body, cardiac, neuro and musculoskeletal regions and seven sequence types, including sites not used in training. The company says it retrained SwiftMR on images already processed by scanner makers' own deep learning reconstruction, and publishes an AI Transparency Statement.

There is nothing on the network architecture, the size of the training and validation sets, or how model updates are released and announced.

The chain is structurally short. SwiftMR's denoising model is AIRS Medical's own and FDA cleared, with no outside model provider in the path. In the cloud deployment only deidentified images leave the hospital, and the reidentification key stays with the customer. An on premises option removes the transfer entirely.

The cloud provider is referred to but not named, and no subprocessor list is published. Training data sources are not described beyond retraining on images processed by scanner makers' own reconstruction.

The evidence is operational and specific, with a clinical evidence library on the site. AIRS Medical reports more than 1,700 institutions in over 40 countries and more than 7 million MRI exams a year. It names site results: routine brain scans on a GE 3T scanner already running the maker's own deep learning reconstruction cut from 15 to 9 minutes, and 3T Radiology & Research in Miami reporting higher throughput. The company notes that results vary by scanner, field strength, protocol and site.

The FDA summary for K253775 reports non clinical testing only: signal to noise and sharpness measured against acceptance criteria. No reader study showing that diagnostic accuracy holds at shorter scan times is in that filing.

A specific commitment is published on AIRS Medical's SwiftMR information security page. For cloud deployments, identifiable PHI is removed from the DICOM data inside the hospital before images go to the SwiftMR cloud. Re-identification happens on premises, and the mapping needed to re-identify stays with the customer. The cloud service is designed not to store, transmit or display PHI or personal information. Data is encrypted with AES-256 at rest and TLS 1.3 in transit, and an on premises deployment keeps images inside the hospital altogether.

The company does not say how long deidentified images stay in the cloud, or whether customer images are used to train or improve models. The full whitepaper is available on request.

The design keeps identified data in the hospital, but no status is stated. The SwiftMR security page says identifiable PHI is removed inside the hospital before images reach the cloud and reidentification happens on premises, so the cloud service is designed not to hold PHI. The compliance page links privacy policies for the US, UK, EU and globally without saying which of them covers the service, and the US policy may say. No business associate agreement or HIPAA status is stated on the security or compliance pages.

The gateway that strips identity still runs in the hospital and handles PHI, and AIRS Medical staff may support it. Nothing says whether AIRS Medical signs business associate agreements or what support access reaches identified data.

Controls are described with an outside check behind them. The SwiftMR information security page says the whitepaper covers penetration testing and vulnerability management, 24/7 security monitoring and incident response, AES-256 and TLS 1.3 encryption, and MFA and enterprise single sign on. AIRS Medical's release notes document security fixes.

No information security certification is named on AIRS Medical's own pages; the security page refers to "certifications and attestations" inside a whitepaper available on request. ISO 27001, 27017 and 27018 have been reported in a distributor's material, which is not the vendor's own statement. The only standard the vendor names itself is ISO/IEC 42001, which covers AI governance, not security.

SwiftMR is FDA 510(k) cleared. The first clearance came in November 2021. The latest, K253775 (summary prepared March 2026; predicate K230854), is under 21 CFR 892.2050, product code QIH, Class II, and covers noise reduction and sharpening for MR images of all body parts. It covers field strengths of 0.25, 0.31, 0.4, 0.55, 0.6, 1.5 and 3.0 tesla. Its validation included images already processed by scanner makers' deep learning reconstruction, which AIRS Medical announced in April 2026 as clearance to run alongside those pipelines.

The indication is image enhancement. The filing validated on images acquired with up to 50 percent shorter scan times and measured signal to noise and sharpness, not diagnostic accuracy. It is not a detection or diagnostic clearance, and it is not for use on mobile devices.

AIRS Medical names AI governance under ISO/IEC 42001, the AI management system standard, on its SwiftMR information security page, and publishes an AI Transparency Statement on its compliance page. The K253775 filing describes the validation set's make up by scanner maker, field strength, body region, sequence, age (8.9 percent pediatric) and sex. It also says sites outside the training data were included.

That is composition, not results: no performance is given by group, scanner, field strength or body size. The risk specific to denoising is that the model smooths toward typical anatomy, so a small or unusual finding is what it is most likely to soften.

A reader can check each output. The original and processed series can be sent to the archive separately, so the unmodified image is available beside the enhanced one, and the site sets the strength of processing.

Nothing stands behind the output, and no error rate is published. The K253775 filing reports that processed images met targets for signal to noise and sharpness on at least 90 percent of the dataset. Those are image quality measures, not a rate at which findings are lost or altered, which is the failure that matters for a product that changes the diagnostic image. No warranty, indemnity or remedy is attached.

Integration is through the imaging stack, and the systems it works with are named. AIRS Medical says SwiftMR is compatible with MRI scanners from Siemens Healthineers, GE HealthCare, Philips and Canon Medical. The K253775 validation also covered Esaote, Fonar and Fujifilm scanners. Enhanced images return to the site's PACS as DICOM, and the original and processed series can be routed to separate archive destinations. It now runs alongside scanner makers' own deep learning reconstruction.

No PACS product or record system integration is named, and no chart integration applies to an enhancement tool.

Both options are documented, with the data path for the cloud one. AIRS Medical's security page describes cloud and on premises deployments. In the cloud path, identifiable data is removed inside the hospital, only deidentified images go to the SwiftMR cloud, and reidentification happens back on premises with the mapping held by the customer. It describes shared responsibility between AIRS Medical, the cloud provider and the customer, plus backup and disaster recovery objectives.

The cloud provider and region are not named, and tenant isolation is not described, though the security whitepaper available on request may cover it. For a Korea based company serving more than 40 countries, where deidentified images are processed matters for European and other customers.

No price is published, but the economics are laid out. The site has an ROI calculator and case studies framed on throughput: shorter scans, appointment slots moving from 30 to 20 minutes, and capacity from scanners a site already owns instead of a new purchase. SwiftMR and SwiftSight are separate products, and access starts with a demo request.

The unit of charge is not published, which matters for a product sold on added volume. A per study fee grows with the extra scans the product creates; a per scanner or per site fee does not.

Coverage is MRI only and broad within it. SwiftMR is cleared for all body parts at field strengths from 0.25 to 3.0 tesla, validated across body, cardiac, neuro and musculoskeletal imaging and seven sequence types. The validation data included some pediatric scans (8.9 percent). It is sold to imaging centers and hospitals in more than 40 countries. SwiftSight adds quantitative brain measurement.

Validation stands behind part of that: the filing measures image quality, not diagnostic performance, and does not break results down by region, field strength or age. Supported body parts and field strengths may vary by country.