Magentiq Eye

A deep learning model analysing colonoscopy video frame by frame in real time is the entire product. There is no content library, no workflow layer and no services component: the device takes video in, runs inference, and returns video with bounding boxes drawn on it fast enough to be useful while an endoscopist is still withdrawing the scope. Remove the model and nothing remains but a passthrough.

The company has described the mechanism plainly and consistently since founding, and the trial literature describes the same thing, so there is no gap between the marketing account and the technical one.

Concurrent read, which is the opposite of the choice Ibex made in pathology and sits at the same point on the spectrum as Riverain. The model marks suspected polyps on the live video while the endoscopist is performing the procedure, and the endoscopist decides whether to look closer, resect or move on. Nothing is diagnosed and nothing is removed without a human. Two things a buyer should weigh, and neither is a criticism of the design.

Concurrent read during a live procedure is the highest automation bias configuration available, because the operator cannot afterwards separate what they saw from what they were shown, and the company's own description of drawing boxes on almost every frame in which a polyp appears is explicitly optimised for attention capture.

And false positives in this setting cost procedure time and unnecessary resections rather than a second opinion, so a false positive rate is a workflow and patient burden question, not only an accuracy one. The published specificity figure of 98 percent speaks to this and the real world false positive burden per procedure is the number to ask for.

Both routes to transparency are available here, which is rare. The FDA clearance puts a device description and intended use statement on a public regulatory record, and the Lancet Digital Health paper describes the system, the trial design and the outcome measures in a venue where reviewers and readers can interrogate them. Between the two, an outside party can characterise what the device claims to do and how well it did it under a randomised design.

Held at B rather than A because the model itself remains closed: no architecture, no training data description, no published performance breakdown by polyp morphology, size or location, and the headline figures of 99.6 percent polyp wise sensitivity and 98 percent specificity are stated without the dataset they were measured on.

The strongest trial design in this batch and one of the better ones in the index. An international multicentre randomised tandem trial ran for fourteen months across ten centres in Europe, the United States and Israel with 31 endoscopists and roughly 950 enrolled patients, registered in advance, and was published in Lancet Digital Health in February 2024.

It met its primary endpoint with a 7 percent absolute and 37 percent relative increase in adenomas detected per colonoscopy, alongside a reported 26 percent relative increase in adenoma detection rate and a 48 percent relative reduction in adenoma miss rate. The design is what earns the A. A tandem trial performs a second examination to establish what the first one left behind, which means the study measures the miss rate directly rather than inferring it.

Detection rate alone can rise simply because more of the easy lesions were found; miss rate is the number that says whether anything is still being walked past, and measuring both in one study is genuinely uncommon. Two honest limits: the trial is company sponsored, and adenoma detection is a surrogate for the outcomes that matter, so reduced cancer incidence and mortality are inferred from a well established association rather than measured here.

The architecture is favourable and the published detail is absent. Real time inference on live video effectively forces local processing, because the latency budget for drawing a box while a scope is being withdrawn does not permit a round trip to a remote service, so procedure video very likely never leaves the endoscopy suite. That is inferred from the product's nature rather than stated.

Two passes located no statement on whether video is retained after the procedure, for how long, or whether recorded procedures are used to improve the model. The last question matters here more than the architecture does: a detection model improves by being trained on more procedures, so the commercial incentive to retain video is strong, and no published position addresses it.

Two retrieval passes located no HIPAA statement, no Business Associate Agreement terms and no privacy policy covering patient data. Graded on published posture. One point of fairness belongs on the record: a device installed in a procedure room that processes video locally has a genuinely smaller regulated data footprint than a hosted service, and may create a narrower business associate relationship or arguably none at all depending on how it is configured. That cuts both ways.

It makes the absence less consequential than it would be for a cloud vendor, and it also makes the statement easier to write, since the company would be describing a favourable position rather than defending a complicated one.

Two passes located no SOC 2, no ISO 27001, no trust centre and no vulnerability disclosure policy. An ISO 13485 quality management certification is very likely given clearances in three jurisdictions but none was located, so none is asserted. The gap has a specific shape worth naming, and it connects this record to the healthcare cybersecurity lane in this same index.

This is a networked computing device installed in a procedure room, sitting in the video path of a colonoscope, which is exactly the class of connected clinical device that vendors such as Asimily, Armis and MedCrypt exist to inventory, patch and monitor. A hospital's device security team will ask what operating system it runs, how it is patched and what it talks to on the network, and none of that is published.

Authorised in three jurisdictions and maintained across product versions, which is what separates an A from a single clearance. A CE mark and Israeli approval arrived in mid 2021, an FDA clearance through the 510(k) route followed in July 2023 for the automatic polyp detection system, and an updated cleared version was released in January 2026 adding endoscopic tool detection, so the clearance has been carried forward rather than left to age.

The split a buyer must see, following the same disclosure this index applies to Heartflow, Caristo and Artrya: the feature sets differ by market. The computer aided diagnosis capabilities covering polyp size estimation and characterisation, and the procedural quality metrics, are described as CE marked, while the United States version carries detection plus tool detection. A European and an American buyer are therefore not evaluating the same product, and a vendor claim demonstrated under one mark does not automatically travel to the other.

The trial design addresses the variation that actually matters most in this procedure, and the company does not present it as a governance disclosure. Adenoma detection rate varies enormously between endoscopists, so the operator is the dominant source of variation in colonoscopy, and running the trial across 31 endoscopists at ten centres in three regions is meaningful evidence that the effect is not an artefact of a few unusually receptive operators. That is credited.

What is missing is everything else: no performance breakdown by patient demographics, no analysis by polyp morphology, size, location or bowel preparation quality, no model card, no bias testing methodology and no drift monitoring concept.

Bowel preparation quality in particular is worth asking about, because it varies systematically with factors that track deprivation and comorbidity, and a detection model's performance on a poorly prepared colon is a fair question nobody has published an answer to.

Almost none required, and for this product class that is mostly a feature. The device sits between the colonoscope and the display and preserves the existing procedure workflow, so adoption needs no record system project, no interface build and no vendor negotiation with the hospital's software supplier, which is the same low friction adoption path this index credited to FDB for different reasons.

Graded C rather than higher because the axis measures depth and there is none: no named integration with any endoscopy reporting system, no interoperability standard and no documented path into the procedure report. The newer computer aided quality metrics create an integration requirement the older product did not have, since quality measures are only useful if they land somewhere durable, and no published material describes where they go.

An on premise appliance in the procedure room performing inference locally in real time. That is the favourable answer to the question this index asks across categories, and here it is close to forced by physics rather than chosen as a policy, since the latency budget for marking a moving video frame does not allow a remote round trip. Residency therefore resolves well by default: the video is analysed where it is produced. Held at B rather than A because nothing is published.

No architecture description, no statement of whether the unit connects outward for updates, telemetry or model improvement, and no statement about what leaves the room. A device that never phones home and a device that uploads flagged clips are very different propositions and the material does not distinguish them.

Nothing published on price, pricing mechanism, whether the unit is sold or leased, or whether charges are per procedure or per site. The reimbursement position is the more consequential gap and it is unresolved across the whole segment rather than specific to this vendor: no dedicated payment mechanism was located for computer aided detection during colonoscopy in the United States, so a purchasing endoscopy unit is buying an improvement in adenoma detection that generates no additional payment while adding procedure time and resections. That is a genuinely difficult economic case and the vendor publishes no analysis of it, which is a missed opportunity given how strong the clinical result is.

One procedure, one specialty, one organ. Colonoscopy in gastroenterology, deployed in hospital endoscopy units and ambulatory surgery centres, with geographic reach across the United States, Europe and Israel. Expansion so far has been depth rather than breadth: detection first, then characterisation and size estimation, then procedural quality metrics, with combined upper and lower gastrointestinal endoscopy raised in company material as a direction.

The narrowness is appropriate to the product, since a detection model is trained on the appearance of one thing in one kind of video, but it does mean an endoscopy unit is buying a single purpose appliance rather than a platform.