FDA Cleared Radiology AI: What 510(k) Really Tells You

The short answer: Most radiology AI sold in the United States is FDA cleared, not FDA approved, and the difference matters. Clearance almost always comes through the 510(k) pathway, which asks whether a device is substantially equivalent to one already on the market for a stated intended use. It does not certify accuracy, it does not promise the algorithm will behave the same way on your scanners, and it says nothing about whether the product is worth buying. As of March 30, 2026 the FDA AI-Enabled Medical Device List held 1,524 entries, and radiology accounts for roughly three quarters of them. So the useful question is not whether a vendor is cleared. It is what the clearance actually says.

Updated July 2026.

Every radiology AI evaluation reaches the same slide eventually: a logo grid with "FDA cleared" underneath. It ends the conversation when it should start one. Two products can both be cleared and mean completely different things by it, because the categories underneath cover very different claims. Here is how to read the claim properly, and what to do with it once you have.

What is the difference between FDA cleared and FDA approved?

FDA approved means the device went through premarket approval, the most demanding pathway, generally reserved for high-risk Class III devices and usually requiring clinical trial evidence of safety and effectiveness. FDA cleared means the device went through 510(k) premarket notification, where the manufacturer demonstrates substantial equivalence to a legally marketed predicate device. Roughly 96 percent of AI device authorizations take the 510(k) route. Almost no radiology AI is FDA approved, so a vendor using that phrase is either imprecise or overselling.

There is a third route worth knowing. De Novo exists for genuinely new device types with no predicate. When Viz.ai took its large vessel occlusion product through De Novo in February 2018, the FDA created a new classification for computer-aided triage and notification software. That single decision is why every stroke triage product cleared since then had a predicate to point at.

What do CADe, CADx and CADt mean?

This is the distinction that changes what a product is allowed to tell your radiologists, and most buyers never ask about it. The three letters at the end are not marketing. They describe the claim the FDA reviewed.

The practical consequence: a CADt product that moves a suspected pulmonary embolism to the top of the queue is doing exactly what it was cleared to do, and a vendor implying it also tells you whether the clot is there is describing a different device class than the one on the certificate. When a stroke or PE tool is on your shortlist, check which category each module sits in before you compare them. Our pages on AI triage in radiology and AI pulmonary embolism detection both walk through where the line falls in practice.

How do I check if a radiology AI product is FDA cleared?

Do not take the website at face value, and do not take the sales deck at all. The check takes about ten minutes per vendor.

Start with the FDA AI-Enabled Medical Device List, which the agency maintains and updates periodically. Search the manufacturer name rather than the product name, because products get renamed and acquired companies keep old registrations. Then pull the 510(k) summary itself from the FDA device database and read two things: the indications for use statement, and the predicate. The indications statement is the only sentence in the whole process that legally constrains what the device claims, and it is usually narrower than the brochure. A module cleared for adult non-contrast head CT is not cleared for pediatric studies, and a device cleared for use as a triage aid is not cleared to inform your interpretation.

Then check the clearance date against the product you are being sold. Algorithms get retrained. If the clearance is from 2021 and the vendor is demonstrating a model they describe as substantially improved, ask which version is the cleared one and whether the improvement went through its own submission.

Does FDA clearance mean the AI will be accurate at our site?

No, and this is the single most useful thing to understand before signing anything. Clearance is based on the manufacturer test data, which was collected somewhere that is not your department. Case mix, scanner generation, acquisition protocol and even the equipment manufacturer all move the numbers.

There is now published evidence for exactly this. An AJR multisite study in 2026 found that a commercial fracture detection algorithm performed differently depending on which radiography equipment vendor produced the image, at sites all using a cleared product as intended. That is not a scandal, it is how machine learning works, and it is the reason your contract needs a local validation period rather than a launch date. We go through what that looks like on the AI fracture detection page.

Sensible groups shadow the tool against signed reports for a defined window before it influences anything, then keep measuring. Somebody has to own that measurement after go-live, and it needs a record of where each number came from, which is easier when the pipeline feeding the quality dashboard has documented data lineage rather than a spreadsheet somebody rebuilds every quarter. This is the part of AI governance that gets promised in the evaluation and forgotten by month four.

How many AI medical devices has the FDA authorized?

As of March 30, 2026, the FDA AI-Enabled Medical Device List contained 1,524 entries. Radiology accounts for about 76 percent of the total, which works out to roughly 1,100 imaging devices. The list has grown steadily rather than explosively, and the concentration in radiology reflects a simple fact: imaging produces standardized digital data with a ground truth that can be labeled, which is a much easier machine learning problem than most of medicine.

The number is often quoted as though it were a measure of maturity. It is closer to a measure of activity. Many of those entries are modules from the same handful of vendors, and a clearance count on a vendor slide tells you the company files a lot of submissions, not that any individual module is good. When you see a vendor advertising a large clearance portfolio, ask how many of those modules you would actually turn on.

Is the FDA changing how it regulates AI CAD devices?

It is consulting on it. On January 8, 2026 the ACR reported that the FDA is seeking comment on a proposal to offer optional premarket flexibility for CAD devices across all four categories, CADe, CADx, CADe/x and CADt. Under the proposal, a manufacturer that already holds at least one cleared CAD device could launch a similar device without submitting a new 510(k). Comments closed on February 27, 2026.

The trade is more post-market responsibility. Manufacturers taking the flexible route would have to develop monitoring plans, make instructions accessible inside the device interface, enable export of AI results, and restrict distribution to facilities with qualified training programs. There is also flexibility on testing methodology, including alternatives to the multi-reader multi-case studies usually recommended for showing improved reader performance.

Whatever the final form, the direction is clear enough to plan around: less weight on the premarket gate, more weight on what the buying institution monitors after installation. Groups that already run local validation are ahead. Groups treating the clearance certificate as the end of diligence are about to be further behind than they think.

Is FDA clearance required to sell radiology AI in the US?

It depends entirely on what the software claims. Software that detects, characterizes or triages findings in medical images is a medical device and needs an authorization. Software that manages workflow, transcribes dictation, drafts report text for a radiologist to edit, or arranges a queue on non-clinical criteria such as order time and modality generally is not making a device claim. The distinction sits in the intended use, not in whether the product happens to contain a model.

This is where we should be direct about our own position. Radiological.ai is decision support. It flags suspected findings for review, prioritizes the worklist and drafts the structured report into your template, and the radiologist reviews, edits and signs every study. We publish no accuracy figures and we make no regulatory-status claims on this site. If a vendor tells you their status, including us, verify it yourself in the FDA database rather than in a slide, and put whatever you are told into the contract.

What should we ask a vendor about clearance?

Five questions, in this order, and get the answers in writing.

1. Which exact 510(k) or De Novo number covers the module we are buying, and what does the indications for use statement say word for word?
2. Which CAD category is it, and what is the product therefore not cleared to tell our readers?
3. Is the model you demonstrated the cleared version, and if it has been retrained since clearance, what happened regulatorily?
4. What did performance look like at sites with our scanner mix and our patient population, and will you support a shadow period before the tool influences anything?
5. Who owns performance monitoring after go-live, what is measured, and what report do we get?

None of that is adversarial. A vendor with a strong product answers all five without hesitating, and the ones who cannot are telling you something useful. The rest of the diligence, the parts about integration, pricing and renewal, is covered in our radiology AI vendor evaluation questions, and if you are still assembling a shortlist, the category overview at best AI radiology software groups the main vendors by the job they actually do rather than scoring them. For the acute pathway specifically, the RapidAI alternative and competitor comparison covers how the neurovascular platforms differ, including the published head-to-head evidence.

The short version

FDA clearance is a floor, not a recommendation. It tells you a device was reviewed for a specific intended use and found substantially equivalent to something already sold. It does not tell you the product works on your scanners, that your radiologists will trust it, or that the alert rate will be tolerable at two in the morning. Read the indications statement, find out which CAD category you are buying, insist on a local validation window, and decide who owns the monitoring before the contract is signed rather than after.