Cerebriu

The models are the whole product and the capability does not exist without them. Deciding which MRI sequence should be acquired next, from what an abbreviated opening set has already shown, is not a workflow that a rules engine or a checklist could perform, because it requires interpreting images in seconds while a patient lies in the scanner.

There is no hardware, no service layer and no image archive underneath. Even the distribution routes make the point: the company reaches buyers by embedding its model inside another manufacturer's scanner and by listing on independent artificial intelligence platforms, both of which are ways of selling a model into infrastructure someone else owns.

The most consequential real time position in this index, and the company is careful about where it stops. Apollo suggests the next sequences and the technologist at the console decides, with the company describing the suite as a safeguard for radiographers and radiologists rather than a replacement for their judgement.

What makes the placement unusual is that the decision cannot be revisited. Almost every other artificial intelligence product here operates on data already captured, so an error can be corrected by looking again. Here an omitted sequence is simply never acquired, and once the patient has left the scanner the only remedy is to recall them. An error of omission is therefore irreversible in a way a misread image is not.

Held at B because nothing published describes what the technologist is shown when the model is uncertain, how a suggestion is declined, or whether the system records which recommended sequences were not taken. That audit trail is the thing a department would want when a finding is missed.

Performance is published with a denominator, which is the disclosure that matters and which most of this index omits. A validation study in a peer reviewed radiology journal reports roughly 89 percent sensitivity and 90 percent specificity for ischaemic lesions across 800 patient scans, and the company separately reports 90.1 percent sensitivity for infarct detection at the initial stage.

The target pathologies are named specifically rather than described in general terms: infarcts, intracranial haemorrhage and tumours. What is absent is any architecture description, any account of training data provenance or scale, and any breakdown of how performance differs across the three products in the suite, which perform quite different tasks.

The architecture implies part of an answer for one distribution route and says nothing about the other, and that split is the finding. Analysis has to complete in the seconds between one imaging sequence finishing and the next beginning, which means processing happens at or very near the scanner console rather than in a distant service, and that in turn narrows where images travel and who could hold them.

It is a structural constraint arising from the product's own timing requirement rather than a policy, which makes it more reliable than a promise. The same product is also distributed through third party imaging marketplaces, and that route necessarily involves platforms the vendor does not operate, with a different data path, different processing locations and a different set of parties. Nothing published addresses it.

So a buyer's exposure depends on which route they bought through, and only one route is described. Nothing else is enumerated either: no model or model family, no hosting arrangement for the marketplace path, no sub processor list, and no position on whether customer images contribute to model improvement. Ask which distribution route applies to your deployment, what the marketplace path involves, and for a training position on customer images.

Peer reviewed validation exists and is specific: a study in the European Journal of Radiology in 2023 covering 800 patient scans with reported sensitivity and specificity for ischaemic lesions. That is more than most records in this index can show and it is the right kind of evidence for a detection claim.

Regulatory assessment adds a second layer, since CE marking under the current European regulation requires a notified body to examine the technical documentation rather than accept a manufacturer's declaration, and the embedded version was assessed separately.

Held at B rather than A because there is no outcome study. Nothing published shows that adapting protocols in real time changes recall rates, time to treatment or diagnostic yield in practice, and those are the claims the product is sold on. Case studies are referenced without published measurement.

Graded on an honest basis. No published retention schedule, encryption detail or position on model improvement from customer images was located in this pass.

The architecture implies part of an answer. Analysis has to complete in the seconds between one sequence finishing and the next beginning, which means processing happens at or very near the scanner console rather than in a distant service, and that in turn narrows where images travel. Whether that holds for the marketplace distributed products, which route through third party platforms, is a separate question and is not addressed.

Graded on an honest basis. No compliance statement or agreement posture was located in this pass.

One fact simplifies the current position and will complicate it later. The company states its United States status is for investigation only, so it is not at present selling clinically into the American market and United States health privacy obligations are correspondingly limited. That changes the moment clearance arrives, and a buyer evaluating it in anticipation should ask what the posture will be rather than what it is.

Recorded honestly: the dedicated trust and security search this index requires was not run in this pass, so the grade is provisional and should not be quoted until it has been.

One factor points upward more strongly than usual. Software embedded inside a scanner sold by a major manufacturer must satisfy that manufacturer's own supplier and product security assessment before it can ship inside their device, and that is a considerably harder gate than a hospital procurement review. It has evidently been passed. Nothing about it is public.

European rather than American, and the asymmetry is the exact inverse of several records in this index. Apollo is CE marked under both the older directive and the current regulation, meaning a notified body has examined it, while clearance from the Food and Drug Administration is described as pending and the United States status is stated plainly as for investigation only. That plainness is to the company's credit: a buyer cannot mistake the position.

The distinguishing achievement is the embedded clearance. In 2025 the company secured CE marking for its protocol product as software embedded inside a Siemens Healthineers MRI system, and describes it as the first clearance anywhere for clinical artificial intelligence embedded in a scanner from a major manufacturer. That is a different regulatory object from a standalone application, because the artificial intelligence becomes part of the device rather than something running beside it, and the assessment has to cover the combination.

Partial, which is better than the usual nothing. Published sensitivity and specificity with a stated cohort size give an outside reader a starting point, and the target conditions are named.

What is missing is any breakdown. No performance by age, sex, scanner manufacturer, field strength or site, and for a product distributed across Denmark, the United States, Norway, India, Israel and Brazil that spread is not incidental. The failure mode also deserves its own analysis and has none: a protocol suggestion that systematically under recommends a sequence for a particular presentation would produce incomplete studies for that group, and because the missing images were never acquired the shortfall is invisible in any later review of the scan.

Performance is published with a denominator and decomposed into the statistics that matter, which is what this axis rewards and what most of this index omits. A validation study in a peer reviewed radiology journal reports roughly 89 percent sensitivity and 90 percent specificity for ischaemic lesions across 800 patient scans, and the company separately reports 90.1 percent sensitivity for infarct detection at the initial stage.

Reporting sensitivity and specificity separately rather than a single accuracy figure is the correct form for a detection product, and doing it in a venue that applied external review is what makes the numbers something a buyer can rely on and a competitor can contest. The target pathologies are also named specifically rather than described in general terms, so a radiologist knows what the system claims to find and, by omission, what it does not.

Held below the top grade for reasons of scope and commitment rather than quality. No breakdown is published showing how performance differs across the three products in the suite, which perform quite different tasks, so a buyer purchasing one cannot assume the published figure describes it. No architecture description or training data account exists. And no warranty, indemnity or remediation commitment attaches to any of it. Ask for performance per product and for the population the validation cohort was drawn from.

Three distinct routes to the point of use, which is unusual and commercially deliberate. The product runs at the MRI console during acquisition, it is embedded inside a scanner sold by a major manufacturer, and it is distributed through two independent radiology artificial intelligence platforms that already integrate with hospital imaging infrastructure.

The console position is the technically demanding one, since it requires operating inside the acquisition workflow rather than reading from an archive afterwards. Held at B because no interface standard, image archive certification or specific integration detail is published, and because the capability available may differ between the embedded, platform and standalone routes without that being stated.

Better than most because the customer has genuine alternatives. The product can arrive embedded in the scanner, through an independent imaging artificial intelligence platform, or presumably standalone, and those are different data paths as well as different commercial arrangements.

The real time constraint pushes processing close to the machine, which is the strongest available answer to a residency question. Held at B because no hosting arrangement, region or retention position is published for any of the three routes, and because a buyer choosing the platform route inherits that platform's data handling rather than this company's.

Nothing is published: no price, no mechanism, no unit of sale.

The distribution structure gives a partial explanation rather than an excuse. Where the software is embedded inside a manufacturer's scanner, the price a hospital pays may sit inside that manufacturer's contract and may not be this company's to disclose, which is the same structure already recorded on a diagnostics vendor distributing through a large manufacturer's platform. The same applies to the independent platform route. A buyer should establish which of the three routes they are purchasing through before asking anyone for a number, because the answer determines who sets it.

Deliberately narrow and coherent. The scope is brain MRI, and within it the named conditions are infarcts, intracranial haemorrhage and tumours, which is the acute neurological set where speed changes management.

The settings follow from that: neurology, emergency imaging where stroke is the driver, private imaging centres wanting throughput, and teleradiology providers reading remotely. Geographic spread is wide for a company of this size, with deployments named in six countries across four regions, but everything remains inside one organ and one modality. Nothing addresses body imaging, other modalities or paediatrics.