Magentiq Eye

A deep learning model analyzing colonoscopy video in real time is the entire product. The 510(k) summary (K223473) describes MAGENTIQ-COLO acquiring the endoscope's digital video, processing frames through neural network based detection using deep machine learning with supporting algorithms, and returning the video with highlighted regions to the monitor fast enough to be useful during withdrawal (median marker latency 0.166 seconds).

There is no content library, workflow layer or services component on the company's pages; the home page presents the one product. Remove the model and nothing remains but a video passthrough. The company's account and the regulatory description match.

Concurrent read during a live procedure, with the endoscopist deciding. The cleared indication (K223473) is an adjunct that assists the endoscopist in identifying lesions by highlighting regions with visual characteristics of mucosal abnormalities, not a replacement for histopathological sampling, limited to white light imaging.

The model marks suspected polyps on the live video; the endoscopist decides whether to look closer, resect or move on, and nothing is diagnosed or removed without a person. Two risks follow. Concurrent marking during a live procedure is the configuration most exposed to automation bias, because the operator cannot later separate what they saw from what they were shown.

And false positives here cost procedure time and unnecessary resections rather than a second look; the summary's 0.0328 false positives per frame, and the trial's non inferior adenomas per extraction (0.31 against 0.27), speak to that burden. What is not published is how the system behaves with poor bowel preparation or outside white light, or how a mark the endoscopist disagrees with is meant to be handled.

Beyond the regulatory record, the model is closed. The 510(k) summary (K223473) describes neural network based detection using deep machine learning with supporting algorithms, processing each video frame and marking regions in real time, and reports the standalone test design and figures; the Lancet Digital Health paper describes the system and trial. Four clearances from 2023 to 2026 show the device carried forward in versions, including tool detection. The architecture beyond a neural network, the training data (size, sources, independence from the test set), the model version in each clearance, and how updates are delivered are not published.

The company's own cleared device carries the model, and the chain around it is short and unnamed. MAGENTIQ-COLO's detection model is Magentiq Eye's own, cleared four times as a device (K223473 through K260724), so it cannot be swapped without a regulatory consequence. The unit processes video locally, so no cloud inference provider sits in the procedure path.

What is not published is where the training data came from, whether any pretrained component sits inside the model, which operating system and third party software run on the unit, and what role PENTAX Medical plays beyond distribution.

A randomized, multicenter trial with a tandem design, published in a peer reviewed journal. The international trial ran for fourteen months across ten centers in Europe, the United States and Israel, with 31 endoscopists and about 950 randomized patients, was registered in advance, and was published in Lancet Digital Health in February 2024; the 510(k) summary (K223473) reports its results.

Adenomas per colonoscopy were 0.70 with the device against 0.51 with conventional colonoscopy, a 37 percent relative increase, with adenomas per extraction non inferior (0.31, 95 percent CI 0.27 to 0.35, against 0.27, 0.23 to 0.31); adenoma detection rate was significantly higher and adenoma miss rate significantly lower, reported as a 26 percent relative increase and a 48 percent relative reduction.

The tandem arm measures the miss rate directly, by reexamining what the initial colonoscopy left behind, which is uncommon. Two limits apply: 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 an established association rather than measured.

One fact about the data path is published, and nothing about stewardship. The 510(k) summary describes the device acquiring the endoscope's video and processing it in real time in a unit that sits in the video path of the procedure room, with the highlighted video shown on local monitors.

It does not say whether video is stored on the unit, whether recordings are transmitted after the procedure, for how long anything is retained, or whether recorded procedures are used to improve the model. That last question matters more than the architecture: a detection model improves by training on more procedures, so the incentive to retain video is strong, and the site links no privacy policy at all. Safety engineering, meaning guardrails and how a safety event is handled, is not published.

No statement of status and no privacy document of any kind is published. The site's footer links an accessibility statement and social media accounts, and no privacy policy, terms of use, HIPAA statement or business associate position is published.

One point cuts the other way: a device processing video locally in a procedure room has a smaller regulated data footprint than a hosted service, and may create a narrower business associate relationship or none, depending on whether it stores or transmits anything. That makes the absence less consequential than for a cloud vendor.

No security material of any kind is published. The site links no security page, trust center, vulnerability disclosure policy or certification, and the 510(k) summary as published includes no cybersecurity information. This is a networked computing device installed in a procedure room, in the video path of a colonoscope, and a hospital's device security team will ask what operating system it runs, how it is patched, what it talks to on the network and whether a software bill of materials exists; none of that is published. ISO 13485, likely for a device cleared in three jurisdictions, would be a quality system certification rather than a security one in any case.

Cleared in the US and carried forward across versions, with the feature set differing by market. FDA records list four 510(k)s for MAGENTIQ-COLO (ME-APDS) under 21 CFR 876.1520, product code QNP (gastrointestinal lesion software detection system): K223473 (25 July 2023; predicate Cosmo's GI Genius, K211951), K244023 (24 January 2025), K252178 (3 October 2025) and K260724 (3 April 2026).

The cleared indication is an adjunct to white light screening and surveillance colonoscopy that highlights regions consistent with mucosal abnormalities, not a replacement for histopathological sampling. The home page states the product is CE compliant and that its CADe version is FDA cleared; CE marking and Israeli approval date from mid 2021.

The features differ by market: polyp size estimation and characterization and the procedural quality metrics are described as CE marked, while the US version carries detection and tool detection, so a European and an American buyer are not evaluating the same product.

The clearance names a broad set of subgroups and states that performance held across them. The 510(k) summary (K223473) reports results stratified by polyp size (small, medium, large), morphology (sessile, flat), location (proximal colon), histology, patient sex, age bands (under 50, 50 to 60, over 60), BMI and geography (US against outside the US), and states that performance remained consistent across subgroups; the per group figures are not in the summary.

The randomized trial ran across 31 endoscopists at ten centers in three regions, which speaks to the dominant source of variation in colonoscopy, the operator. Flat and small lesions are both the hardest to detect and among the most consequential to miss, so the missing figures for the morphology and size strata matter most. Bowel preparation quality, which varies with factors that track deprivation and comorbidity, is not among the groups named. No model card, bias testing method or drift monitoring is published.

A published error rate with its method and denominator is on the record, and nothing contractual stands behind it. The 510(k) summary for MAGENTIQ-COLO (K223473, July 2023) is Magentiq Eye's own filing for its deep learning detector: tested on 172 colonoscopy videos containing 449 polyps (330 of 5 mm or less, 76 of 5 to 10 mm, 43 of 10 mm or more), it reports per polyp sensitivity of 99.6 to 100 percent depending on the detection threshold (one to seven frames), and 0.0328 false positives per frame normalized to a 15 minute procedure.

It is a test set result for the cleared indication (an adjunct to white light screening and surveillance colonoscopy), not performance in use, and per polyp sensitivity on video counts a polyp as found if any frame is marked, which is a generous definition beside the flat and small lesions that matter most. No warranty covering model output, accuracy commitment, indemnity or remediation route is published.

Almost no integration is required, and none is documented. The device sits between the colonoscope and the display and preserves the existing procedure workflow, so adoption needs no record system project and no interface build; the company's partnership with PENTAX Medical suggests hardware compatibility with that maker's endoscopes, which is not the same as record integration.

No integration with an endoscopy reporting system, interface standard or documented path into the procedure report is published. The newer quality metrics (adenomas per colonoscopy, adenoma miss rate and detection rate among the metrics the site lists) create an integration need the detection product did not have, since quality measures are useful only if they land somewhere durable.

An on premises unit in the procedure room, performing inference where the video is produced. The 510(k) summary describes the device acquiring digital video from the endoscope camera, processing it in real time and outputting highlighted video to a touchscreen monitor, with an HDMI splitter for additional displays; the median marker latency of 0.166 seconds is consistent with local processing. Residency therefore resolves by design: the video is analyzed in the room.

What is left open is the processing path beyond the procedure: whether the unit connects outward for updates, telemetry or model improvement, and whether clips are uploaded later. A device that never connects out and a device that uploads flagged clips are very different propositions, and the published material does not distinguish them.

How the product reaches buyers is published; its price is not. The company's page on the economic case argues for additional income from more colonoscopies performed and more adenomas resected, and savings on colorectal cancer treatment, and notes that in the US and some European countries centers are reimbursed further when polyps are resected; that is the payer's side of the equation, not the vendor's price.

The site describes an expanded partnership with PENTAX Medical covering EMEA, JAPAC, Latin America and North America, which is a named distribution route. What is not published is any price, whether the unit is sold or leased, or whether charges are per procedure or per site. For computer aided detection in colonoscopy no dedicated US payment mechanism is described, so a unit buying it gains adenoma detection without additional payment while adding procedure time and resections.

One procedure, stated clearly, with validation behind it. Coverage is white light screening and surveillance colonoscopy in gastroenterology (K223473), deployed in hospital endoscopy units and ambulatory centers across the US, Europe and Israel, with a PENTAX Medical partnership covering EMEA, JAPAC, Latin America and North America.

The randomized trial validated the detection claim across ten centers in three regions and 31 endoscopists, and the clearance reports consistency across US and non US sites. Expansion has been in depth rather than breadth: detection, then tool detection, with characterization, size estimation and quality metrics under the CE mark, and combined upper and lower endoscopy raised as a direction. The narrowness is appropriate to the product, and it is validated; what is not published is performance by setting (hospital units against ambulatory centers) or outside white light.