contextflow

Deep learning on three dimensional imaging is the whole company, and the founding technology was a model rather than a product.

The origin makes the point. contextflow began as a three dimensional image based search engine capable of retrieving visually similar cases from a reference corpus, which is a representation learning problem before it is a clinical one: the system must encode what a region of lung tissue looks like well enough that similarity in that encoding corresponds to similarity in pathology. The clinical product was built on top of that capability rather than the reverse.

What ships now runs several distinct model tasks. Nodule detection and quantification, longitudinal tracking of the same nodule across studies, quantitative tissue analysis, and qualitative classification across 19 named image patterns are separate inference problems, and the last of those is unusually demanding since interstitial lung disease patterns are subtle and disputed among human readers.

No workflow layer exists to carry independent value. The software integrates into a viewer the hospital already owns, so there is not even a portal to fall back on.

Graded A. The reference case retrieval that distinguishes the product commercially is itself a model output, so even the explainability feature is inference rather than presentation.

Explainability built into the output as comparable cases rather than asserted as a principle, which is the strongest oversight design in this imaging lane.

Most imaging artificial intelligence hands a radiologist a finding and a confidence figure, and asks them to accept or reject it with no visibility into why. This product surfaces reference cases and differential diagnosis information alongside its findings, so the reader sees several visually similar prior examples and can judge the comparison directly. Case based reasoning is inspectable in a way a probability is not: a radiologist can look at the retrieved cases and disagree with the resemblance, which is a form of scrutiny a score does not permit.

The positioning is explicit rather than incidental. Company material and a named academic user both frame the product against black box artificial intelligence, and that user describes it as supporting his workflow while leaving the final decision to him. Integration into the radiologist's own native viewer reinforces the same posture, since the reader never leaves the environment where they exercise judgement.

That design matters most for the hardest condition in scope. Interstitial lung disease classification is genuinely difficult and contested among specialists, and comparable cases are how radiologists actually reason about it.

What is missing is calibration. No sensitivity, specificity or operating point is published for any of the detection or classification tasks, so a reader cannot know the miss rate they are working alongside.

Ask for detection and classification performance at deployed thresholds, and how many reference cases are returned.

Capabilities enumerated with checkable specificity, and no performance figures anywhere.

The functional disclosure is precise in a way that invites verification. Four capabilities are named individually: nodule detection with quantification, nodule tracking across successive studies, quantitative lung tissue analysis, and qualitative analysis of 19 image patterns. That last figure is the kind of detail this index rewards, because a stated count of 19 patterns is checkable against what the product actually returns, where a claim to analyse lung patterns would not be.

The underlying method is also identified rather than obscured. The core technology is described as a three dimensional image based search engine performing similarity retrieval over volumetric imaging, and its origin in a named funded European research project is stated openly, so a technical reader can trace the intellectual lineage to published work.

The reference case output is itself a transparency mechanism, since the radiologist sees the evidence the system considered similar rather than only its conclusion.

What is absent is every number. No sensitivity, specificity, detection rate, classification accuracy or operating point appears in any material located, no model card exists, and the reference database size and composition are not characterised despite being fundamental to how the retrieval behaves.

Graded B because the method and feature set are documented with genuine specificity, and held there because nothing quantifies how well any of it performs.

Ask for performance per capability at deployed thresholds, and the size and composition of the reference corpus.

The intellectual origin is disclosed unusually openly, and the operational chain is not disclosed at all.

The provenance disclosure is genuine and rarer than it looks. The company states that its core technology derives from a named European research project and from work at a named university medical school and technical university, with four founders identified. That lets a reader trace the underlying method to published academic work and to a funding programme with its own public record, which is a stronger origin story than most vendors provide and is verifiable independently of the company.

Models appear to be developed in house on that foundation rather than assembled from third party components.

What is undisclosed is everything operational. No cloud or hosting provider, no sub processor register, no machine learning framework and no third party component was named.

The reference corpus is the dependency that matters most and it is entirely undescribed. Case retrieval is only as good as the database behind it, so that corpus is simultaneously the product's differentiator and its largest undisclosed input. Where the imaging came from, which institutions contributed, under what consent and data sharing arrangements, how it is curated and labelled, and whether it grows from customer deployments are all unanswered. A corpus assembled during a research project operates under research governance that does not automatically extend to commercial use.

Acquisition adds a further open question about whether corpus and infrastructure migrate to the parent.

Ask for the reference corpus provenance, consent basis and curation process, and the hosting provider.

An academic foundation and clinical partnerships, without the published results or deployment figures the stronger records in this lane carry.

The research lineage is genuine rather than decorative. The company emerged from a university medical school and a funded European research project, states that peer reviewed research underpins its development, and maintains a publications and abstracts page. Clinical validation is described as achieved through partnerships across Europe, and a named senior academic radiologist at a German university medical centre speaks publicly about the product, which is a checkable reference rather than an anonymous testimonial. An advisory board drawing on European and international radiology society leadership adds professional scrutiny.

What is missing is measurement. No specific study, sensitivity figure, reader performance comparison or outcome result was located. Nothing states how many hospitals run the product, how many studies have been analysed, or what changed at any site.

That gap is more visible here than it would be in isolation, because three other vendors in this same imaging lane publish journal studies with patient numbers, treatment rate effects, follow up rate improvements and health economic analyses. Against that standard, research heritage is a starting position rather than evidence.

The 2026 acquisition implies commercial diligence was performed and describes an established customer base, which is corroboration of a kind rather than clinical evidence.

Ask for reader performance data against unassisted reading, the current deployment count, and any study measuring reporting time or diagnostic agreement.

A dedicated pass located no encryption statement, no retention schedule, no access control description, no deletion process, no anonymisation position and no statement on training data use.

The absence is more consequential for this product than for a straightforward detector, because two of its features depend on holding imaging rather than transiently processing it. Nodule tracking across time requires prior studies or derived measurements to persist so that the same nodule can be matched between examinations. Reference case retrieval requires a database of other patients' scans to exist and be searchable at the moment of reading.

That second dependency is unusual in this index. Most imaging vendors process the study in front of them; this one is architecturally committed to a repository of third party patient images, and nothing published describes what is in it, whether it is identifiable, how it was assembled or whether it grows from customer deployments.

A radiologist looking at a reference case is looking at some other patient's chest. Whether that patient consented to that use, and whether the image is anonymised beyond the removal of header data, is not stated anywhere.

Medical device certification under the current European framework requires documented risk management, so internal controls plainly exist. None is published.

One pre emptive note: restating regulatory certification cannot move this grade, since device certification governs clinical safety rather than information handling. Only a published data handling position will.

Ask what the reference database contains and how it was assembled, what is retained for longitudinal tracking, and whether customer imaging enters either.

A dedicated pass located no health privacy position of any kind: no compliance statement, no data protection statement, no agreement template and no processing terms.

The applicable regime here is European data protection law rather than United States health privacy law, since certification and deployment are European and no United States presence was found. That does not excuse the silence; it relocates it. European hospitals require a data processing agreement, a documented lawful basis and a transfer position before any patient imaging moves, and none of that is published.

The product architecture raises a question sharper than the usual template gap. Reference case retrieval compares a patient's scan against a database of prior cases in order to surface visually similar examples, which means the system depends on a corpus of other patients' imaging being held and queried. Where that corpus came from, under what consent, whether it remains identifiable, and whether customer scans are added to it are all unaddressed, and the last of those is the question a hospital's data protection officer would raise first.

A product whose value proposition is showing you other patients' scans has a consent story to tell and has not told it publicly.

Ownership by a listed parent in a third jurisdiction adds a transfer question that did not previously arise.

Ask for the data processing agreement, the provenance and consent basis of the reference case database, and whether customer studies are added to it.

A dedicated pass located no security page, no information security certification, no trust centre, no penetration testing statement, no vulnerability disclosure policy and no documentation available under agreement.

One distinction needs stating because it is routinely blurred and this vendor's regulatory record invites it. Certification under the European medical device regulation requires a quality management system governing design, development and post market surveillance, and that system concerns clinical safety and manufacturing discipline rather than information security. A device certificate says nothing about encryption, access control, secure development practice or breach response. This record does not credit one as the other.

The comparison within this lane is unfavourable and shows the credential is obtainable. Two other imaging vendors indexed here hold and publish a named information security certification, one of them naming the standard revision and the certified scope, and both operate at comparable scale.

The holding of a reference imaging corpus raises the stakes above the ordinary. A vendor that maintains a searchable database of patient chest scans is holding a concentrated repository of clinical imaging, and no external party has evidently examined the controls around it.

European hospital procurement increasingly requires an information security certification as an entry condition, so the gap will surface at the first serious vendor review regardless.

One pre emptive note: restating device certification cannot move this grade. Only an information security attestation, or documentation available under agreement, will.

Ask whether an information security certification is held or in progress, and what security documentation can be shared under agreement.

One jurisdiction, entered through the demanding door, with no United States position at all.

ADVANCE Chest CT is CE certified under the current European medical device regulation. That distinction carries real weight and this record credits it: the regulation replaced a lighter directive with substantially heavier clinical evidence, quality management and post market surveillance obligations, and a meaningful number of legacy products were withdrawn rather than recertified. The company has publicly described the certification as requiring substantial investment in quality management and compliance, and has discussed incorporating regulatory requirements from the start of development rather than retrofitting them, which is the correct posture and an unusually candid account of the cost.

What holds the grade down is reach. No United States clearance was located and no United States commercial presence was found, so this is a single jurisdiction product. Three other vendors in this imaging lane hold clearances across two to four jurisdictions, and one holds ten United States clearances alone.

The acquisition changes the outlook rather than the current position. The acquiring company holds United States clearances for its own respiratory imaging products and describes the combination as complementary across the detection to monitoring pathway, so a route to United States submission now exists that did not before. Nothing has been filed or announced.

Nothing enumerates the certified indication for use or the classification assigned.

Ask for the certified indication for use, the device class, and whether United States submission is planned under new ownership.

A transparency posture expressed in the product itself, with no measurement of any kind behind it.

What earns the grade above the floor is that the governance position is architectural rather than rhetorical. Choosing to return comparable cases instead of an unexplained score is an interpretability decision embedded in the output, and interpretability is the one governance property that survives contact with a busy radiologist. Public commentary from the company and from a named academic user both frame this as an explicit choice about using artificial intelligence safely and making ethical choices, and an advisory board including radiology society leadership provides professional oversight of a kind.

What is entirely absent is evidence. No fairness testing, no subgroup performance, no calibration data, no drift monitoring and no external audit was located.

Two bias mechanisms are specific to this product. The technical one is acquisition variation, since computed tomography output differs by scanner manufacturer, slice thickness and reconstruction kernel, and interstitial pattern classification is unusually sensitive to reconstruction because the patterns being classified are fine textural features.

The second is particular to case retrieval and is not addressed anywhere. The reference database defines what the system can show a radiologist, so its composition determines whose disease presentations are represented. A corpus drawn from a small number of European academic centres will underrepresent presentations that differ by ancestry, occupational exposure or comorbidity, and the radiologist sees only what the database contains.

Ask for the reference database composition, and classification performance across scanner types and reconstruction protocols.

Nothing contractual is published, for software that supports cancer detection decisions.

A dedicated pass located no service level agreement, no accuracy warranty, no uptime commitment, no indemnity and no remediation position.

The mitigating structure is real and does not move the grade. European medical device certification brings post market surveillance obligations, vigilance reporting and change control, so algorithm changes cannot be made silently and a competent authority exists to receive reports. Under the current regulation those duties are materially heavier than under its predecessor.

What that does not address is the specific failure mode. This is detection support for suspected lung cancer, and a missed nodule that a radiologist relying on the tool also misses is the harm that matters. No sensitivity figure is published, so the residual risk cannot be characterised by the department accepting it, and the tool's own strength works against it here: a product praised for transparency and comparable cases earns a trust that makes automation bias more likely rather than less.

The reference retrieval feature introduces a second exposure absent elsewhere in this lane. If the system surfaces visually similar cases that carry a benign diagnosis, and the case in front of the radiologist is malignant, the comparison itself becomes a source of false reassurance. Nothing describes how retrieval confidence is conveyed.

One pre emptive note: further certifications cannot move this grade. Only contractual terms or published performance will.

Ask what is warranted on detection performance, and how reference case similarity confidence is presented.

An integration philosophy that is right for the user, described without a single specific.

The stated approach is delivery of findings directly inside the radiologist's native viewer rather than in a separate application. That is the correct design goal and it addresses the most common failure of imaging artificial intelligence in practice, which is that results arriving in a second window get looked at during evaluation and ignored during a busy list. A vendor that puts its output where the reader already works has understood the adoption problem.

What is missing is everything that would let a buyer confirm it. No viewer or picture archiving system vendor is named, no interface standard is described, no output format is specified, no marketplace or partner listing was located, and nothing states whether findings return as annotated series, structured report content or an overlay within a specific vendor's application. Native viewer integration is a claim that depends entirely on which viewers, since achieving it typically requires a partnership with each vendor.

No record system integration is mentioned at all, so whether findings reach the patient chart, and whether longitudinal nodule tracking is visible outside radiology, is unaddressed.

Graded C rather than higher on the same standard applied elsewhere in this lane, where a comparable vendor names the standard, the direction of flow and the returned format in one sentence.

Ask which viewers and archiving systems are integrated in production, through what mechanism, and whether findings reach the record system.

A dedicated pass located no hosting provider, no deployment options, no region statement, no residency commitment, no tenancy model and no statement of whether imaging is processed locally or transmitted.

The silence is harder to accept for this product than for most, because the architecture implies a dependency that has to live somewhere. Reference case retrieval requires a searchable corpus of prior imaging to be available at read time, and nothing states whether that corpus sits inside the hospital, in vendor infrastructure, or is queried remotely during reporting. Those are three very different deployments with three different latency, privacy and residency profiles.

European deployment makes the omission consequential rather than academic. Hospitals operating under European data protection rules require a stated processing location and transfer position, and the customer base is described as spanning multiple European markets, several of which apply national requirements above the European baseline.

Ownership introduces a new dimension that did not previously exist. Acquisition by a company listed and headquartered outside Europe, which delivers its own portfolio as a hosted service, raises the possibility that infrastructure consolidates into the parent, and any such move would be a cross border transfer question for every existing European customer.

Ask where processing and the reference corpus reside, whether an on premise option exists, and what the acquisition changes for European deployments.

Cost is absent from every published surface. A dedicated pass located no pricing page, no unit of charge, no range, no tiering, no implementation fee position and no minimum commitment.

One commercial development is worth recording precisely because it must not be overstated. The company has been reported as positioned to secure what would be a first radiology artificial intelligence reimbursement arrangement in Europe. That is forward looking language describing an expected outcome rather than a concluded agreement, no terms or rates accompany it, and no confirmation of completion was located. It is a signal about direction and not a disclosure a buyer can act on, and this record does not treat it as one.

If such an arrangement exists it would be genuinely significant, since European radiology artificial intelligence has largely lacked the reimbursement pathways that exist in the United States, and the absence of a payment route is the most common reason European hospitals decline these products.

Nothing indicates whether the software is licensed per study, per site, per module or per year, and the product spans three disease areas with several distinct capabilities, so whether a hospital can buy nodule detection without the pattern analysis is unstated.

Ownership change adds a further unknown, since pricing may migrate into an listed parent's commercial structure and be bundled with its existing respiratory imaging portfolio.

Ask whether the reimbursement arrangement completed and on what terms, the unit of charge, and whether pricing changes under new ownership.

One modality, one body region, three diseases, one continent.

The scope is deliberately concentrated on chest computed tomography, covering suspected lung cancer, interstitial lung disease and chronic obstructive pulmonary disease, with incidental pulmonary embolism also referenced. Within that boundary the coverage is coherent rather than thin, because those conditions frequently appear on the same scan and a radiologist reading a chest study is assessing all of them at once. Handling three findings on one image is more useful in practice than three separate tools would be.

Interstitial lung disease deserves specific mention as the harder problem. Pattern recognition in interstitial disease is subtle, inter reader agreement among specialists is imperfect, and it is exactly where a reference case comparison helps a general radiologist most.

The limits are geographic and regulatory rather than clinical. Certification covers Europe only, no United States clearance or presence was located, and the customer base is described as growing across European markets without enumeration. Nothing addresses which countries, how many sites, or whether use concentrates in academic centres.

No other modality or body region is addressed, despite the founding search technology being described as applicable across modalities and organs, so the deployed scope is narrower than the underlying capability.

The acquisition may change the geographic picture, since the acquirer holds United States clearances and describes the transaction as complementary.

Ask which European markets carry deployment, the site count, and whether United States entry is planned.