The FDA's June 4, 2026 510(k) batch carried a clearance worth reading closely: K260497, DEEPVESSEL Plaque, from Keya Medical Technology Co., Ltd., a Beijing-headquartered company that has been building a cardiovascular AI portfolio for several years. The classification fields tell the regulatory story in shorthand. Product code QIH, regulation 21 CFR 892.2050, device class II, advisory committee Radiology, clearance type Traditional, decision Substantially Equivalent. Each of those is a constraint, and together they place the device precisely on the regulatory map.
892.2050 is the FDA's classification for a “picture archiving and communications system” and, more broadly, for medical image management and processing software — the post-acquisition layer that stores, displays, manipulates, and analyzes images after a scanner produces them. The generic device name attached to this product code in the openFDA record is “Automated Radiological Image Processing Software.” That phrasing matters. It signals that the device operates on already-acquired images and produces a processed or quantified output, rather than controlling acquisition or rendering an autonomous diagnosis. For an AI plaque tool, that is exactly the right bucket: the software takes a coronary CT angiography dataset and computes something about the vessel wall.
What the device does, and where the line is drawn
The clinical problem DEEPVESSEL Plaque addresses is one of the most consequential in cardiology: characterizing atherosclerotic plaque in the coronary arteries from CT angiography. Plaque is not a single thing. It ranges from dense calcified deposits to lower-attenuation, lipid-rich material, and the composition and burden of that plaque carry prognostic weight. Quantifying it by hand on a CT volume is slow, operator-dependent, and hard to reproduce. Automated software that segments the vessel, identifies plaque, and reports its volume and composition is a natural target for machine learning, because the task is fundamentally one of pixel-level segmentation and classification on a 3D dataset — the kind of problem deep convolutional networks do well.
The regulatory frame, though, draws a careful line around what the cleared software is allowed to claim. Under 892.2050, the device is image-processing and analysis software. It computes and presents quantities; it does not, by virtue of this classification, make the diagnostic call. The interpreting physician reads the output as additional information alongside the underlying images. This is the same adjunctive posture that governs most of the radiological AI clearing through 510(k) today, and it is what keeps a tool like this inside Class II rather than pushing it toward the higher-risk classifications that demand premarket approval. The software's job is to be accurate and reproducible in what it measures; the diagnostic responsibility stays with the clinician.
The substantial-equivalence argument
The clearance type here is Traditional, not Special or Abbreviated, which tells us this was a full 510(k) submission rather than an iteration on Keya's own prior device or a reliance on FDA-recognized consensus standards alone. In a Traditional 510(k), the sponsor must identify a legally marketed predicate, demonstrate the same intended use, and show that any differing technological characteristics do not raise new questions of safety and effectiveness. For automated image-processing software, the predicate is typically another cleared device in the same 892.2050 class that performs an analogous quantification — often an earlier-generation plaque or vessel-analysis tool.
The intellectually honest part of that argument, and the part FDA scrutinizes, is the technology comparison. The predicate and the new device share an intended use: quantitative analysis of coronary CT images. But if DEEPVESSEL Plaque uses a deep-learning segmentation approach where an older predicate used conventional thresholding or model-based methods, the technological characteristics differ. The sponsor then has to show that the different method does not introduce new risk — which, for a measurement tool, means demonstrating that the AI's outputs are accurate against a reference standard and reproducible across cases, scanners, and operators. The standard performance evidence for this category is agreement with expert manual annotations or with an established reference measurement, reported across a representative dataset. That is the evidentiary core of the substantial-equivalence story, and it is where a plaque-quantification clearance lives or dies.
Why the 892.2050 placement is the interesting choice
For a strategy-minded reader, the notable decision is the classification itself. Coronary plaque analysis sits near a boundary. Pushed one way — toward software that estimates a patient's physiology or risk and influences treatment decisions — a tool can drift toward classifications with heavier evidentiary demands. Kept in the image-processing-and-management lane of 892.2050, with output framed as quantitative measurements presented to a reading physician, the device stays in the well-trodden 510(k) channel where dozens of radiological-analysis tools have cleared. The QIH product code and the Traditional pathway together signal that Keya threaded that needle: the device is positioned as a measurement and visualization aid on coronary CT, not as an autonomous risk-stratification engine.
That positioning is consistent with how the broader field of cardiac CT AI has navigated the FDA. The agency has been willing to clear automated plaque and vessel-analysis software as Class II image-processing tools provided the labeling keeps the physician in the diagnostic loop and the performance data substantiate the measurements. Each such clearance extends the predicate chain, making the next one easier to argue. K260497 adds Keya's tool to that lineage, and for an international AI vendor, a U.S. Class II clearance is the credential that opens the door to the American imaging market without the multi-year burden of a premarket approval.
The substance to verify against the cleared summary is the reference standard and the validation cohort: what the plaque measurements were compared against, how the dataset was constructed, and whether the agreement metrics hold across the range of plaque types. The classification settles the regulatory posture; the summary's methods section is where the equivalence claim earns its keep. A measurement tool is only as good as the ground truth it was tested against, and that is the document worth reading line by line.