FDA Approved AI Medical Devices: 1,614 Listed, 76% Radiology

The US Food and Drug Administration (FDA) keeps a public list of the medical devices it has authorized that use artificial intelligence. The version we downloaded on 23 October 2026 holds 1,614 entries, from a Pap smear screener cleared in September 1995 to devices cleared on 29 June 2026.¹

Of those 1,614 devices, 1,230 were reviewed by the FDA's radiology panel.¹ That is 76%. Every other area of medicine shares the remaining quarter.

That proportion says what an "FDA-approved AI" usually is today: software that reads, sorts or improves medical images. This piece counts the list itself, then checks what three independent studies found about the evidence behind the devices on it.

The list, counted by specialty

The FDA files each device under a "lead panel", the group of specialist reviewers that assessed it. We counted the downloadable file by that field.¹

Radiology holds 1,230 devices. The cardiovascular panel is a distant second with 154, about 10% of the list, and neurology follows with 73, about 4.5%.¹

After that the numbers are small enough to name one by one. Anesthesiology has 30 devices, gastroenterology and urology 27, and hematology 22. Pathology and ophthalmology have 10 each, dental has 7, obstetrics and gynecology has 5, and ear, nose and throat has one.¹

Skin cancer makes a useful test, because consumer apps promise to check moles. Only three devices on the list are built to assess skin lesions: MelaFind in 2011, Nevisense in 2017 and DermaSensor in 2024.¹ That is three out of 1,614.

Two independent counts agree on the radiology share

The 76% figure is not a quirk of one download. Two independent counts, taken at different dates, land on the same share.

Daniel Windecker and colleagues analyzed the list as it stood on 31 August 2024, for a study in JAMA Network Open. They found 692 radiology devices out of 903, or 76.6%.² The Imaging Wire, a radiology trade outlet, counted the March 2026 edition and reported 1,104 radiology devices out of 1,451, or 76%.³

So three counts across two years put radiology between 76% and 77%. The share is the most stable fact about the list.

Other specialties are growing faster, from far smaller bases

The list grew slowly for two decades, then quickly. It gained six authorizations in 2015 and 114 in 2020.¹ The current file counts 226 for 2023, 235 for 2024 and 335 for 2025.¹

Radiology's share of each year's new devices has drifted down. It took 99 of 114 in 2020, or 87%. It took 254 of 335 in 2025, or 76%, and 135 of 181 in the first half of 2026, or 75%.¹

The other two large panels grew faster in relative terms. Cardiovascular authorizations rose from 8 in 2020 to 30 in 2025. Neurology went from one to 19.¹ Radiology still added more devices in 2025 than every other panel combined.

The FDA's own count has shifted between editions

MedTech Dive, which tracks the list, reported 331 devices for 2025.⁴ Our count from the newer file is 335. MedTech Dive notes that the FDA sometimes adds older devices to the list after the fact, which would explain a small rise.⁴

A gap in the other direction needs flagging. The Imaging Wire reported 1,451 devices on the list through the end of 2025.³ The file we downloaded holds 1,433 devices dated 2025 or earlier.¹

That means 18 entries from that period appear to have left the list, or had their dates changed, between the March edition and the current one. We could not find an FDA note explaining the change. Anyone quoting a total should give the edition date with it.

Most were cleared by comparison, not by trial

Most devices on the list did not go through a new clinical trial. They went through the 510(k) route, where a company shows the FDA that its device is "substantially equivalent" to one already sold.

That is 96% of the list.¹ Forty devices went through De Novo review, the route for new low-to-moderate risk devices that have no match on the market. Twenty-one went through premarket approval, the strictest route.¹

The Windecker team found the same pattern independently: 877 of its 903 devices, or 97%, were cleared under 510(k).² The FDA says only that listed devices "have met the FDA's applicable premarket requirements".⁵ For most of them, that requirement was the comparison rather than a trial.

What "radiology AI" includes

The radiology count mixes two different kinds of product. Some are software that reads images, and others are the scanners themselves, with AI built into how they take pictures.

We matched the radiology entries to the FDA's product codes. They include 103 CT scanners, 100 MRI systems and 124 ultrasound imaging systems.¹⁶ Together that is 327 of the 1,230, or about 27%.

Those are machines that use AI to sharpen an image or shorten a scan, not tools that tell a doctor what the image shows. The largest makers on the list reflect this: Siemens Medical Solutions USA has 56 entries and Canon Medical Systems has 42.¹

Image-reading software is still the bigger group. The most common product code, with 305 devices, is "automated radiological image processing software".¹⁶ Another 86 are triage tools that flag urgent scans, such as a suspected stroke, so a radiologist reads them sooner.¹⁶

What the evidence behind them shows

Three peer-reviewed studies have checked the public records behind these devices. They used different samples and asked slightly different questions, so their numbers differ, but they point the same way.

About four in ten had no published clinical test. A team led by Sammy Chouffani El Fassi at the University of North Carolina reviewed 521 devices. As reported by MedTech Dive, they found 226 with no published clinical validation data, and only 22 tested in a randomized trial.⁷

About half reported a performance study. The Windecker team found clinical performance studies for 505 of 903 devices, or 56%.² For 218 devices, or 24%, the FDA documents said no such study had been done. Of the studies that did exist, 12 were randomized.²

Recalls cluster early. Branden Lee and colleagues matched 950 AI devices against FDA recall records and found 60 devices tied to 182 recalls.⁸ Of those recalls, 79 came within 12 months of clearance.⁸

Devices with no clinical validation were more likely to be recalled, with an odds ratio of 2.8.⁸ The most common cause was a diagnostic or measurement error, behind 109 of the recalls.⁸

The 43% and 56% figures are not in conflict. The North Carolina team asked whether any clinical validation had been published. The Windecker team asked whether the FDA's own summary reported a clinical performance study. They asked different questions of overlapping samples, and both found roughly half the devices without one.

What the headline numbers leave out

The panel is not the patient. The lead panel records who reviewed a device, not where it is used. HeartFlow's heart-artery analysis sits under the cardiovascular panel, for example, yet it works from CT scans.¹ Treat the specialty split as a rough guide to where AI is used, not a precise one.

The list is not complete. The FDA says the list "is not a comprehensive resource". It finds devices mainly by searching for AI-related terms in their public summaries.⁵ A device whose summary never used those words can be missing from it.

No chatbot-style AI appears on it. None of the 1,614 device names mentions a language model or generative AI.¹ The FDA says it "will explore methods to identify and tag" devices that use foundation models, the large general-purpose models behind chatbots.⁵

The agency opened a public consultation on regulating generative AI devices on 18 August 2026.⁹ As of October 2026, that paper is a discussion document, not guidance.

Authorized does not mean paid for. A study of Medicare claims from 2018 to 2023 counted 83,392 AI services billed by radiologists, as reported by AuntMinnie.¹⁰ Only 47% were paid, for about $8 million in total. One heart-scan analysis accounted for 91% of the paid services.¹⁰

How to use the list

When a clinic, app or advert says its AI is "FDA approved", search the FDA list for the product name. The list is free to browse and to download.⁵ If the product is not there, ask the seller for its FDA submission number.

If it is there, check the submission number first. A number starting with K means 510(k) clearance by comparison, one starting with DEN means De Novo review, and one starting with P means premarket approval. Then look for a published clinical study. In the samples checked so far, roughly half of listed devices have one.²⁷

Then look at the specialty. A claim that AI diagnoses something other than an image belongs to the quarter of the list that is still small, and deserves a closer look.