AIRS Medical

Deep learning is the whole product. The company sells no scanner and no hardware; SwiftMR applies a model after image acquisition to remove noise and blur, which is what allows either shorter scans or better images from the same acquisition. The vendor neutral design reinforces this, since the value proposition is explicitly the algorithm rather than any imaging system it runs behind.

The system operates automatically in the background with images reaching PACS typically in under two minutes, and the radiologist reads the enhanced image rather than reviewing an AI finding. That places the oversight question in an unusual and important spot: this AI does not detect anything, it alters the image the diagnosis is made from. The FDA clearance is what governs whether that alteration preserves diagnostic quality, which makes the regulatory pathway the substantive safety control here rather than clinician review.

The method is described concretely, including that enhancement is applied post acquisition without altering the scan itself, and that the newest capability was achieved by training specifically on images already processed by scanner manufacturers' own deep learning pipelines and testing across multiple vendors and field strengths. That is a specific and checkable technical account.

The company also publishes candid qualifiers that scan time results vary by scanner model, field strength, protocol, and site configuration. Model architecture and quantitative image quality metrics are not published.

One credential here is the material item rather than a general security badge, and the distinction is worth carrying to other records. The company holds the standard specifically addressing protection of personally identifiable information processed in public cloud environments, which covers consent, purpose limitation, transparency about sub processors and restrictions on using customer data for the provider's own purposes.

For a service that moves patient imaging out of the scanner environment into a cloud pipeline and back, that is closer to the question this axis asks than a general security certification is, because a security standard establishes that data is protected while this one bears on what may be done with it. Two architectural mitigations sit alongside. An on premises deployment option means studies never leave the estate for sites that take it.

And original and processed images can be routed independently to the archive, so the unmodified series survives rather than being replaced, which matters because a radiologist who doubts an enhanced image can go back to the source rather than having only the vendor's rendering of the examination.

Held below the top grade because no retention period for images or derived data in the cloud path, no de identification statement, and no position on whether customer studies contribute to model development was located, and operating across more than forty countries means the applicable regime varies with no market specific terms published. Ask what the cloud path retains, and whether studies train models.

Evidence is operational and reasonably specific, with deployment scale as the strongest signal: a reported 1,700 plus imaging centers and hospitals across more than 40 countries, in clinical use since 2023. Named site results include routine brain scans falling from 15 to 9 minutes on a scanner already running the manufacturer's own deep learning reconstruction, and a site reporting appointments cut from 30 to 20 minutes.

The company appropriately labels these as individual site results that may not be typical. What is absent is published reader study evidence that diagnostic accuracy is preserved at reduced scan times, which is the clinical question underneath the operational one.

Corrected from Not Rated. A relevant credential is held and one architectural control is genuinely protective.

ISO 27018 is the material item here rather than on the security axis alone. It is the standard specifically addressing protection of personally identifiable information processed in public cloud environments, covering consent, purpose limitation, transparency about subprocessors and restrictions on using customer data for the provider's own purposes. For a service that moves patient imaging out of the scanner environment into a cloud pipeline and back, that is closer to the question this axis asks than a general security certification is. Held alongside ISO 27017 for cloud security.

The on premise deployment option is the second mitigation, and it is the strongest available: where a site runs the software locally, studies do not leave the estate at all.

The third is retention of the unmodified series. Because original and processed images can be routed independently to the archive, the source study need not be discarded when the enhanced one is produced.

What is still unpublished, and holds this at B: no retention period for images or derived data in the cloud path, no de identification statement, and nothing on whether customer studies contribute to model development. The company operates in more than forty countries with a Korean base, so the applicable regime varies and no market specific terms exist.

Ask what the cloud path retains, and whether studies train models.

Converted from Not Rated. Compliance is asserted in customer material and the instrument is not published.

No business associate agreement, addendum, role statement, subcontractor flow down, breach notification timetable or review cadence was retrieved. What exists is a reference in a health system case study to a centrally managed, health privacy compliant deployment across a multi vendor scanner fleet. A customer describing a deployment as compliant is evidence that the question was addressed in contracting, and it is not a published posture.

The role depends on deployment, and a buyer should be explicit about which applies. Under the cloud path, studies leave the imaging estate and are processed by the vendor, which makes it a business associate holding protected health information with direct liability. Under an on premise installation the vendor's position narrows considerably, potentially to software supply and support access rather than data custody. Those merit different agreements.

Cross border is the second layer and it is not incidental here. The company is Korea headquartered with United States operations, deployed across more than forty countries, and its own release notes indicate feature availability varies by market. Which entity contracts, whether support or engineering access to live systems occurs from outside the country of care, and how that access is logged are the questions to settle.

Ask which entity signs, which agreement applies to your deployment model, and what standing vendor access exists.

Corrected from Not Rated. Certifications are held that the earlier review did not locate, and the sourcing needs stating.

The company is reported to hold ISO 27001, ISO 27017 and ISO 27018. That trio is well matched to what this product actually is: 27001 covers the information security management system, 27017 covers cloud service security specifically, and 27018 covers protection of personally identifiable information in public cloud environments. For a cloud processed imaging service, the second and third are more relevant than a generic attestation would be, and very few vendors in this index hold either.

The sourcing qualification is real and consistent with how this index treats such claims. The certifications appear in a distribution partner's product material rather than on the vendor's own site. A reseller is closer to a party with access to the documentation than an aggregator directory is, so this is credited with attribution, but the vendor does not publish it directly and a buyer should ask for the certificates.

Supporting evidence is better than most: published release notes document multi factor authentication, strengthened password policy, session management and remediation of known vulnerabilities in core components. Continuous, dated, public evidence of patching is unusual disclosure.

No SOC 2 of either type and no trust centre were located, which United States procurement will ask for.

Ask for the certificates and scope, and whether a SOC 2 is planned.

FDA 510(k) cleared with a clearly stated scope: image enhancement supporting scan time reduction of up to 50 percent, covering all pulse sequences and body parts, in clinical use since 2023. The company subsequently obtained an additional clearance in April 2026 specifically permitting operation in conjunction with scanner manufacturers' own deep learning reconstruction pipelines, which is a meaningful and unusual regulatory step because it addresses the stacked processing case rather than assuming it. Buyers should note the clearance covers image enhancement and scan time reduction; it is not a detection or diagnostic claim.

Converted from Not Rated. The prior analysis identified exactly the right concern for this product class and it deserves stating at full strength.

No governance framework, subgroup performance analysis, bias evaluation or model update policy was located.

The risk is specific to reconstruction rather than detection. A model trained toward typical anatomy learns what an image usually looks like, and denoising works by suppressing what does not fit that expectation. A rare finding, an unusual anatomical variant, a small lesion or an artefact that happens to carry diagnostic meaning are all, statistically, the unexpected. The concern is not that the model errs at random but that it smooths in a direction, and that direction runs against the atypical.

Body habitus is the other axis. Signal to noise degrades with body size in magnetic resonance imaging as it does in other modalities, so the reconstruction is asked to do more work in larger patients, which is where this index has already found the evidence thinnest for a comparable product.

The stated coverage widens the exposure rather than narrowing it. The software is offered across all body parts, all ages, all pulse sequences and field strengths from 0.25 to 3.0 tesla, including legacy and open scanners. That is an enormous validation surface and no per segment performance is published against any of it.

Ask for performance by body habitus, field strength and sequence, and what evidence covers rare findings.

The technical account is concrete and one published qualifier is more useful than most vendors' headline claims. The method is described as enhancement applied post acquisition without altering the scan itself, and the newest capability is described as achieved by training specifically on images already processed by scanner manufacturers' own pipelines and testing across multiple vendors and field strengths.

That last detail is specific and checkable, and it names a real difficulty: enhancing an image that has already been enhanced by another system is a harder problem than enhancing a raw one, and saying that is what was trained for tells a physicist what to expect. The qualifier is the second useful disclosure, since the company publishes that scan time results vary by scanner model, field strength, protocol and site configuration.

A vendor stating that its headline benefit depends on local factors is giving up the clean number in exchange for an accurate expectation, and a department that reads it will design its own measurement rather than assume the brochure figure. Held at C because nothing quantifies image quality.

No architecture and no quantitative metrics were located, which for an enhancement product is the central omission: the trade being made is scan time against image fidelity, and only one side of it is described. No warranty, indemnity or remediation commitment attaches. Ask for quantitative fidelity metrics against conventional acquisition, and for any reader study on diagnostic equivalence.

Interoperability here is the imaging stack rather than the chart, and vendor neutrality is the design commitment: the platform works across scanner manufacturers, field strengths, and body parts, returns enhanced images into the customer's PACS typically within two minutes, and now holds clearance to operate alongside OEM deep learning reconstruction rather than requiring the customer to choose between them. For a radiology department running mixed scanner fleets that is the difference between one deployment and several. No EHR integration applies or is claimed.

Corrected from Not Rated. Both deployment models exist and are documented in technical detail, which the earlier review did not find.

On premise deployment is evidenced concretely rather than claimed: published release notes describe a launcher supporting installation and DICOM input and output directories on a secondary drive, giving flexible storage configuration for on premise installations, alongside gateway server setup steps covering network, scanner and archive connectivity. Cloud managed deployment exists in parallel and is positioned for enterprise networks standardising across sites.

One routing capability deserves specific attention because it answers a question this index raised about this whole product class. Original and processed series can be routed to separate archive destinations independently. For reconstruction software that modifies the primary diagnostic image, the central concern is that the radiologist reads the model's output with no unmodified comparator available, and the failure mode is invisible because a study that lost a subtle finding looks cleaner rather than worse. Preserving the original series alongside the enhanced one restores the ability to check, and that is a meaningful safeguard rather than a workflow preference.

What remains unpublished: cloud region, tenancy model, whether processing stays within the country of care, subprocessor list and retention terms.

Ask which model applies, whether both series are retained by default, and in which region cloud inference runs.

Converted from Not Rated. No price is published, and the return is quantified in a way that makes the missing denominator conspicuous.

Nothing is disclosed on rate, basis, term or minimum. Whether pricing is per scanner, per site, per study or subscription is unstated, and that determines everything for a buyer, because the value proposition is throughput. If the fee scales per study, it scales with exactly the additional volume the product exists to create, and the buyer's return is a share rather than the whole. If it is per scanner or per site, the return accrues to the buyer as volume grows. Two buyers with identical scan volumes could reach opposite conclusions.

The published return side is specific: scan times reduced by up to half, appointment slots cut from thirty minutes to twenty, and a cited site generating more than one hundred thousand dollars in additional monthly patient revenue from the added capacity. The capital argument is the sharper one, that software throughput gains defer a scanner replacement costing millions.

That framing is legitimate and it is precisely why the price matters. A buyer being asked to compare this against a capital purchase needs both numbers, and only one is available.

Ask for the pricing unit first, specifically whether it is metered per study, and what happens to the fee as volume grows.

Single modality by design, which is MRI, but broad within it: the cleared scope spans all pulse sequences and all body parts across vendors and field strengths, which is wider than most imaging AI that is cleared organ by organ. Settings span imaging centers and hospitals, with a second product referenced for the imaging portfolio. Buyers outside MRI are outside scope.