On June 4, 2026, the FDA cleared Sonio Suspect, a software device from the Paris-based company Sonio, under 510(k) premarket notification number K261519. In the agency's records it is a Class II radiology device, product code POK, regulation 892.2060 — "computer-assisted diagnostic software for lesions suspicious for cancer." That product-code description is, unusually, a near-complete spec: this is AI that looks at ultrasound images and flags spots that may be cancer, handing the radiologist a second set of eyes that never tires.
The clearance is notable for the pathway it took. Where most of the consumer-health devices clearing the FDA each week use the Traditional 510(k), Sonio Suspect cleared via the Special 510(k) — the route reserved for modifications to a manufacturer's own legally marketed device where the change can be reviewed against well-established methods. A Special 510(k) is a signal in itself: it usually means the company already had a cleared predicate and is iterating on it, which is exactly the cadence that defines medical AI. The model is not a one-time approval; it is a product line that clears, improves, and re-clears.
Computer-assisted detection, decoded
The category Sonio Suspect belongs to is computer-assisted detection and diagnosis — CADe/CADx in the trade. The idea predates the current AI wave by decades: software that highlights regions of a medical image for a clinician's attention. What changed is the engine. Modern systems are trained on large image sets to recognize the visual patterns associated with malignancy, and they have grown good enough that the FDA now clears a steady stream of them. But the regulatory framing has stayed deliberately conservative, and that framing is the most important thing to understand about this clearance.
The product-code language says the software flags lesions "suspicious for cancer." It does not say the software diagnoses cancer. That word choice is not marketing softness; it is the legal boundary of the device. Sonio Suspect is an aid that draws a radiologist's attention to a region and offers an assessment; the human reads the image, weighs the flag, and makes the call. The FDA's substantial-equivalence finding — decision code SESE on the K261519 record — is a determination that the software is comparable in intended use and safety to a predicate computer-assisted diagnostic device, not a blessing of autonomous diagnosis. For any AI medical product, the distance between "flags suspicious lesions" and "diagnoses disease" is the distance between a cleared aid and a product that would require a far heavier evidence burden.
Novel, or just renamed?
It is fair to ask whether an AI cancer-detection tool clearing in 2026 is genuinely new or a familiar technique wearing fresh branding. The honest answer is that the technique — pattern recognition on medical images — is mature, and the novelty lives in the specifics: the imaging modality, the lesion types, the training data, and the measured performance against a predicate. Sonio Suspect's modality is ultrasound, a notably operator-dependent imaging method where the quality of the scan varies with the hands holding the probe. Software that adds consistency to an inconsistent modality is a real contribution even if the underlying machine-learning approach is well-trodden. The Special 510(k) pathway reinforces this reading: this is iterative refinement of a cleared capability, not a leap into uncharted regulatory territory.
That Sonio is a French company clearing a device with the U.S. FDA is also part of the story. Medical AI is a global build, and the FDA's 510(k) database has become a worldwide leaderboard for clinical software. A clearance is a passport into the U.S. market, and the steady appearance of European and Asian AI developers in the records reflects how internationally distributed this category has become. The country code on K261519 is FR, and the device still had to satisfy the same substantial-equivalence standard as any domestic applicant.
The Special 510(k) detail deserves one more look, because it reveals something about the lifecycle of a medical-AI product that the Traditional pathway obscures. An AI model is not a finished object the day it clears; it is a system that benefits from retraining, expanded data, and tuned thresholds over time. The regulatory question that haunts the category is how to let a model improve without forcing a full new submission for every adjustment. The Special 510(k) is part of the FDA's answer — a faster lane for changes that can be evaluated against established methods. When you see a CADe/CADx product clear via this route, you are watching the regulatory machinery accommodate software that is meant to keep changing, which is a genuinely different problem from the static instruments the 510(k) system was originally designed for. Sonio Suspect clearing as a Special is a small data point in a much larger experiment: building a regulatory regime that can keep pace with software that learns.
The consumer-tech angle
This is not a device a consumer buys, but it is squarely a consumer-tech story, because it is where the AI hype cycle meets a hard accountability bar. The same year that chatbots were being asked to diagnose rashes from phone photos with no oversight, the FDA cleared a cancer-detection AI under a regime that forces it to be an aid, names it precisely, and holds it to a predicate. The contrast is the lesson. Consumer AI ships on a tweet; medical AI ships on a K-number. Sonio Suspect carries a product code, a regulation citation, a Special 510(k) designation, and a substantial-equivalence finding — the full apparatus that separates a regulated diagnostic aid from an app that guesses.
For anyone trying to read the medical-AI market without getting swept up in the marketing, the move is to follow the clearance database, not the demo reel. The flashy claim is "AI detects cancer." The filed reality is "computer-assisted diagnostic software for lesions suspicious for cancer, cleared as substantially equivalent to a predicate, intended to assist a radiologist." K261519 is the filed reality, and the filed reality is consistently the more accurate — and more durable — version of the story.