Lunit

Detection models are the entire product. The INSIGHT suite spans chest X-ray, mammography, and digital breast tomosynthesis, with the mammography product returning lesion location plus an abnormality score reflecting model confidence. The company sells no imaging hardware and no services layer; where hardware is involved it is through OEM relationships that embed the algorithm into someone else's scanner.

Cleared as detection and triage support with the radiologist reading every study, but this vendor sits closer to the autonomy question than most because of what its own evidence argues. A prospective study of more than 50,000 screening cases published in The Lancet Digital Health found the AI paired with a single radiologist achieved a higher cancer detection rate than the traditional two radiologist double reading standard, which the company frames as AI replacing one human reader. That is a substitution claim rather than an assistance claim, and buyers in double reading programs should treat the staffing implication as the real decision.

Training data scale is disclosed specifically, with the mammography product trained on more than 240,000 cases including up to 50,000 cancer cases, and performance is published in venues that permit external scrutiny rather than in marketing material.

Most notably the company has submitted to third party head to head evaluation twice: a JAMA Oncology comparison against two other commercial mammography AI products, and a Radboud University led multi vendor validation on lung nodule detection. Inviting independent comparative benchmarking is the strongest transparency signal available in imaging AI and almost no competitor does it.

Nothing identifies any party in the chain: no hosting arrangement, no sub processor list, no retention or de identification statement, and no position on whether customer studies train models was located in two passes. The spread is what makes that consequential rather than routine.

Deployment is reported across more than three thousand institutions in over forty countries, so the applicable health data regime differs by site and no market specific terms are published for any of them, and distribution partly through imaging equipment manufacturer relationships adds further parties to the path, each with its own handling posture and none named. A buyer therefore cannot establish who is in their particular chain, which will differ from the next customer's.

The training question carries specific weight here because of what this company has published: more than a hundred peer reviewed studies, including the independent head to head validations credited on the other axis, rest on very substantial imaging datasets. Where those came from, whether commercial deployments contribute, and whether a site can decline, is unaddressed.

The answer may be entirely clean, since research datasets are normally assembled under arrangements separate from commercial deployment, and that distinction is precisely what a buyer needs stated rather than assumed. The modality raises the stakes, since screening mammography reaches large asymptomatic populations. Ask what is retained after a read, whether studies train models, and which parties are in the path for your deployment.

The deepest published evidence base of any imaging vendor in this index. More than 100 peer reviewed papers since 2018, split roughly 55 on lung abnormalities through the chest X-ray product and 45 on breast cancer, appearing in The Lancet Digital Health, JAMA Oncology, and Radiology.

Evidence includes a prospective Karolinska study of 55,581 real world screening cases, a European Radiology study finding no statistically significant accuracy difference versus radiologists at matched specificity, and independent third party validations. Deployment spans a reported 3,000 plus institutions in 40 plus countries. Buyers should read the European Radiology result carefully, since at matched sensitivity the AI showed slightly lower specificity and higher recall than first readers.

Converted from Not Rated. No stewardship terms were located, and the jurisdictional and distribution spread makes this a per deployment question rather than a single policy.

Nothing was retrieved on retention of studies or derived outputs, de identification, model training use, or deletion at contract end.

The spread is the complicating factor. Deployment is reported across more than three thousand institutions in over forty countries, which means the applicable health data regime differs by site, and no market specific terms are published for any of them. Distribution partly through imaging equipment manufacturer relationships adds further parties to the path, each with its own handling posture.

The training question carries specific weight given what this company has published. More than a hundred peer reviewed studies, including third party head to head validations, rest on very substantial imaging datasets. Where those came from, whether customer deployments contribute, and whether a site can decline, is unaddressed. As with its French peers graded alongside it, the answer may be entirely clean, since research datasets are normally assembled under separate arrangements from commercial deployment, and the distinction is precisely what a buyer needs stated.

The modality raises the sensitivity. Mammography and breast tomosynthesis studies are among the more personally sensitive images in routine care, and screening populations are large and asymptomatic.

Ask what is retained after a read, whether studies train models, and which parties are in the path for your deployment.

Converted from Not Rated. No business associate terms were located, and this record has the longest contracting chain of the four imaging vendors graded alongside it.

No agreement, addendum, role statement, subcontractor flow down, breach notification timetable or review cadence was retrieved.

The role is standard: the hospital or screening programme is the covered entity, and a vendor receiving studies and returning findings is a business associate.

The chain is not standard. Distribution runs partly through imaging hardware manufacturer relationships, meaning the software can reach a customer embedded in or alongside equipment supplied by a third party, across a reported three thousand plus institutions in more than forty countries. Where an algorithm arrives as part of a scanner vendor's offering, the hospital's contract may be with the equipment manufacturer rather than with the model developer, and the question of who holds the business associate agreement, who is accountable for a breach and who has visibility of the data path becomes genuinely hard to answer from outside.

Cross border adds the final layer. A Korea headquartered manufacturer selling into United States screening and diagnostic settings must establish which entity contracts and whether any processing, support access or model operations occur outside the country of care.

Ask who your counterparty actually is, whether the agreement is direct or through an equipment partner, and where studies are processed.

Converted from Not Rated. No independent security attestation was located across two differently phrased searches.

No SOC 2 of either type, no ISO 27001, no HITRUST, no trust centre and no penetration testing statement was retrieved.

The absence sits oddly against the rest of this record, which is the most externally validated in the imaging category. More than a hundred peer reviewed publications across major journals, including third party head to head comparisons against competing commercial products, is a posture of submitting to outside scrutiny that very few vendors anywhere in this index match. A company willing to be measured against its competitors by independent investigators is not a company avoiding examination.

That contrast is the point rather than a criticism. Clinical validation and information security assurance are separate disciplines with separate audiences, and excellence at one does not imply the other. A radiology department can be fully satisfied that the model reads mammograms accurately and still have no evidence about how its patients' studies are protected.

One structural note for a United States buyer: the company is listed on a public market in its home country, so it carries securities disclosure obligations there. Those filings may describe risk management in ways the marketing does not, and are worth reading directly.

Ask what independent security assessment exists, and check the home market filings for cybersecurity disclosure.

FDA 510(k) cleared for both the mammography detection product and the chest X-ray triage product, making both commercially available in the United States, alongside CE marking and approval for commercial sale in more than 35 countries. The company is publicly listed on the Korean exchange, which adds a layer of financial disclosure obligation absent from most private vendors here. Buyers should note the US cleared scope differs from the fuller product line available under CE, and should confirm which specific indications their deployment covers.

Stronger than most, though expressed through published research rather than a governance program. Studies specifically examine performance in dense breast tissue, which is the best documented failure mode in mammography screening and a population level equity issue, and third party validation across independent institutions tests generalization beyond the developer's own data. The multi vendor Radboud validation places its performance in comparative context. What is absent is a formal governance framework, monitoring commitment, or subgroup analysis across race and age.

This company has done the thing almost nobody in imaging does: submitted its products for independent head to head evaluation against named competitors, twice, in a journal comparison against two other commercial mammography products and in a university led multi vendor validation on lung nodule detection.

That is the strongest transparency signal available in this field, and the reason is structural rather than reputational: in a comparison the vendor did not design and cannot withdraw from, it does not control the outcome, so a favourable result carries information a self run benchmark never can. This index has recorded a direct counterexample in the same cluster, where a competitor was unable to provide the analysis required for inclusion in exactly that kind of study.

Training scale is disclosed at more than two hundred and forty thousand cases, and performance appears in venues permitting external scrutiny. One construction deserves flagging even on a strong record, because the treatment should not soften for vendors this index otherwise praises: the cancer case count is given as up to fifty thousand, which sets a ceiling rather than reporting a figure, and an up to number is an absence of a number. Held below the top grade because no warranty, indemnity or remediation commitment attaches. Ask for the cancer case count as a figure, and for performance in the screening population you serve.

Distribution strategy substitutes for integration engineering here, and effectively. The company embeds its algorithms directly into imaging hardware through OEM relationships with major scanner manufacturers and distributes via major cloud infrastructure, meaning a customer can acquire the AI as part of equipment already being purchased rather than as a separate integration project. That is a materially different adoption path than PACS side deployment. No EHR integration applies or is claimed.

Multiple delivery paths are evidenced, spanning embedded OEM deployment inside imaging hardware and cloud distribution through a major provider, which gives buyers real choice about where inference runs. Specific tenancy, hosting, and data residency terms are not published, which matters given the global footprint and varying localization requirements.

No published product pricing. As a publicly listed company it discloses revenue and financial performance under exchange requirements, which gives a buyer more visibility into vendor stability than any private competitor in this category offers, but nothing about per study, per site, or licensing cost. OEM embedded distribution likely carries different economics than direct sale, and neither is disclosed.

Deliberately concentrated on cancer detection rather than general radiology, spanning chest X-ray for lung abnormalities and mammography plus tomosynthesis for breast, with a separate oncology business line addressing therapy response. Settings center on organized screening programs and breast imaging centers, which is where the double reading economics the company targets actually exist. Emergency triage, neuro, musculoskeletal, and general radiology workflows are outside scope, distinguishing it from broad platform competitors.