Public regulatory record · compiled August 2026
FDA Approved AI Medical Device List for Radiology: FDA Cleared Radiology AI Vendors Compared
The short answer
Radiology is by far the largest category on the FDA AI-enabled medical device list: 1,164 of 1,524 authorizations, about 76%, as of the snapshot dated March 31, 2026. Almost none of it is "FDA approved" in the strict sense. It is 510(k) cleared, which means substantially equivalent to an existing device. GE HealthCare holds the most authorizations at 130; among software-only vendors, Aidoc leads at 33. What matters more than any count is the product code on the clearance, because that is what decides whether the software is allowed to triage a study, mark a finding, or only measure one.
This is the regulatory homework a radiology group has to do before a purchase order, in one place: who holds what, what each clearance type permits, and how to check a vendor's claim yourself in about four minutes.
Last updated August 2026
Worklist
Structured report
DraftRun the assistant to draft this report for review.
Illustrative sample · not a real patient study, not a diagnosis
Decision support for qualified clinicians. Radiological.ai does not provide a diagnosis and is not a substitute for professional judgment.
Flag · prioritize · draft · you review and sign
1,524
AI-enabled medical devices authorized by the FDA since 1995, through March 31, 2026
1,164
Of those are radiology devices, about 76% of every AI authorization the agency has issued
69 of 92
Q1 2026 AI authorizations that were radiology, so the share is holding at three quarters
Source: the FDA Artificial Intelligence-Enabled Medical Device List, Q1 2026 snapshot, as reported by The Imaging Wire on June 17, 2026. The FDA states that the list is not a comprehensive resource, because it is compiled from devices whose public authorization summaries use AI terminology.
FDA cleared radiology AI vendors
Top radiology AI companies by FDA authorization count
Ranked by the number of entries attributed to each submitter on the FDA AI-enabled device list as of March 31, 2026. Read it as a measure of regulatory throughput, not of clinical quality.
| # | Vendor | FDA authorizations | What the portfolio covers |
|---|---|---|---|
| 1 | GE HealthCare | 130 | Imaging hardware plus reconstruction, quantification and workflow software across CT, MR, X-ray and ultrasound |
| 2 | Siemens Healthineers | 95 | Scanner-side AI, AI-Rad Companion reading support, cardiovascular and neuro quantification |
| 3 | Philips | 58 | Reconstruction, cardiac and vascular quantification, ultrasound automation |
| 4 | Canon | 48 | Deep-learning reconstruction and modality-attached analysis |
| 5 | United Imaging | 40 | Scanner-side acquisition and reconstruction AI |
| 6 | Aidoc | 33 | Acute triage and notification across CT, X-ray and CTA, plus a 2026 foundation-model authorization |
| 7 | DeepHealth | 29 | Mammography, chest and lung screening; RadNet-owned since the Aidence and Quantib acquisitions |
| 8 | Samsung | 21 | Ultrasound and mobile X-ray analysis |
| 9 | Rapid.ai | 20 | Stroke and vascular triage: LVO, perfusion, ICH, aneurysm, pulmonary embolism |
| 10 | Hyperfine | 13 | Portable low-field MRI acquisition and image processing |
The first five entries are scanner manufacturers, and that shapes the whole list. A large share of radiology AI authorizations are not reading assistants at all. They are reconstruction, denoising, segmentation and quantification features that ship attached to a CT or MR system, and each one is submitted separately. If you are shopping for software that reads a study, the list you actually care about starts at row six.
Among software-only vendors, the ordering is worth holding lightly. Aidoc at 33 and DeepHealth at 29 both got there partly through breadth of pathology coverage, and DeepHealth's number carries the history of Aidence and Quantib, which RadNet acquired and folded in. Rapid.ai at 20 is concentrated almost entirely in stroke and vascular work, which is a narrower portfolio serving a deeper use case. We work through what that concentration means for a stroke program in the RapidAI comparison, and the ownership question behind DeepHealth in who owns DeepHealth.
Read the fine print
Why a vendor's clearance count rarely matches the FDA list
Entries versus findings
The FDA list counts authorization entries. Vendors frequently count cleared findings or indications, and a single submission can carry several. Viz.ai, for example, publicly states more than 50 cleared algorithms while not appearing in the top ten by list entries. Both numbers can be true at once; they are answers to different questions.
Acquisitions split the history
Clearances are attributed to the submitter name on the paperwork. When a company is bought, renamed or merged, its earlier authorizations stay filed under the old name unless the record is updated. Any count of a company that has grown by acquisition is understated somewhere.
The list is not exhaustive
The FDA says so directly. The roster is built by finding AI terminology in public authorization summaries, so a device whose summary never uses the words can be legally marketed and still be missing from it. Absence from the list is not evidence of anything.
The practical consequence for a buyer is simple. A clearance count is a marketing number, whoever produced it. Ask instead for the K numbers that cover the specific modality, body region and patient population you intend to run the software on, and count those.
What a clearance actually permits
The four FDA product codes behind radiology AI
This is the part of a clearance that decides what the software is allowed to do, and it is almost never on the vendor's homepage. Regulation numbers are from 21 CFR Part 892.
Radiological computer-assisted triage and notification (CADt)
- What it permits
- Flags a study as suspicious and moves it up the worklist, or sends a notification.
- What it does not
- Does not mark a region on the image and does not offer a diagnosis. Nothing about the pixels changes.
- Why a buyer cares
- This is what most "AI triage" platforms hold. It is a workflow claim, not a reading claim.
Radiological computer-assisted detection and diagnosis (CADe/x)
- What it permits
- Identifies and marks findings on the image, and may characterize them.
- What it does not
- Does not replace the interpreting physician; the radiologist still reads and signs.
- Why a buyer cares
- A stronger claim than triage, and cleared against a reader study rather than a timing study.
Radiological computer-assisted diagnostic software for lesions suspicious of cancer (CADx)
- What it permits
- Characterizes a lesion the reader has already found, for example scoring likelihood of malignancy.
- What it does not
- Does not find the lesion for you. Detection and characterization are separate clearances.
- Why a buyer cares
- Ask which one you are being sold, because the demo usually shows both happening at once.
Medical image analyzer
- What it permits
- Processes, quantifies or measures an image without making a detection claim.
- What it does not
- Does not assert that anything abnormal is present.
- Why a buyer cares
- Common for volumetrics, segmentation and quantification products.
The distinction that catches people out is between triage and detection. A CADt device under QAS or QFM reorders your worklist and can send a notification, and that is the entire claim. It is explicitly not permitted to direct a radiologist's attention to a specific part of the image. A detection device under QBS or QDQ does exactly that, and it had to clear a harder evidentiary bar to get there.
Both are useful. They solve different problems, they fail in different ways, and they should not be compared on price per study as though they were the same product. If the demo you were shown highlighted a region on a scan, ask which product code that behavior sits under, and ask to see it in the Indications for Use text.
A fifth code sits next to these four and is worth knowing because it governs software almost every practice already runs. Image management and processing systems, including the viewers radiologists read from, clear under 21 CFR 892.2050, most often as product code LLZ, which had 2,285 clearances as of an openFDA pull on August 31, 2026. That is where the answer lives to a question buyers ask constantly and rarely get straight: whether the DICOM viewer in front of them is cleared for diagnostic reading at all. We worked through it with verified K numbers in DICOM viewer software and which DICOM image viewers are FDA cleared.
Four minutes, before the purchase order
How to verify a radiology AI clearance yourself
Every step uses a free public database. None of it requires the vendor's cooperation beyond one number.
1
Ask for the K number
Not a press release, not a badge on a slide. A 510(k) number looks like K243210. A vendor that cannot produce one for the specific product you are buying has told you something important.
2
Look it up in the FDA database
Search the number in the FDA's public 510(k) database. Confirm the decision date, the applicant name and the device name match what you were told. Applicant mismatches are common after an acquisition and worth asking about.
3
Read the Indications for Use
This is the document that matters. It names the modality, the body region, the patient population and often the specific finding. Anything outside that text is off-label, however well the software performs.
4
Check the product code
Match it against the four codes above. If the clearance is QAS and the sales conversation was about detection accuracy, you are being sold a workflow product on reading-assistant terms.
Do this once per product on the shortlist and keep the printouts. It is the cheapest due diligence available in this category, and in our experience compiling vendor pages it resolves more disagreements in a procurement meeting than any demo does. The rest of the evaluation, the questions that clearance cannot answer, is covered in the vendor evaluation questions worth asking.
What changed in 2026
The FDA declined to make radiology AI easier to clear
In December 2025 the question of whether radiology CAD devices still need premarket notification was live. A petition, filed by Harrison.ai through Rubrum Advising, argued that manufacturers holding prior CAD clearances had demonstrated enough process competence that future devices of the same types could be exempted from 510(k).
On April 1, 2026 the FDA denied it. The agency's reasoning is quotable and worth carrying into a vendor conversation: holding a 510(k) clearance "may not reflect proficiency in processes for all future devices of subject types," and the "processes and methods necessary to evaluate a CADe device are not the same as a CADx device." The agency pointed vendors toward predetermined change control plans and Q-Submissions instead.
For buyers, three things follow. Every new radiology AI device still arrives with its own clearance, so the K number test above keeps working. A vendor's track record does not transfer to its next product in the FDA's eyes, and it should not transfer in yours either. And because model updates are the pressure point, the predetermined change control plan is now the document to ask about: it defines what the vendor may change in a deployed model without coming back to the agency.
Ask this at the next vendor call
- Do you have a PCCP on this device? If yes, what changes does it cover, and what would still require a new submission?
- When did the deployed model last change? And how were we told?
- Which scanners was it validated on? Named makes and models, not "major vendors."
- What was the test-set composition? Site count, demographics, prevalence of the target finding.
- What happens when it is wrong? Specifically, what the radiologist sees and what gets logged.
Honest limits
What FDA clearance does not tell you
Nothing about your population
Clearance rests on the submission's datasets. Disease prevalence, scanner mix and protocol differences at your sites can move real-world performance in either direction, and nobody finds that out until the software is running.
Nothing about the current model
A clearance is dated. Under a predetermined change control plan the model can be retrained and updated inside the boundaries of the original submission, so what you evaluate may not be the version that was cleared.
Nothing about the workflow
Whether a flag reaches the right radiologist in time, and whether they trust it enough to act, is an integration and change-management question. It decides most of the value, and no clearance speaks to it.
Nothing about reimbursement
Clearance and payment are separate systems. Very few imaging AI products have a dedicated payment pathway, which is worked through in radiology AI CPT codes and reimbursement.
Nothing about how many you need
Narrow clearances push groups toward buying several products. Whether that is the right shape is a separate decision, covered in how many radiology AI vendors a practice needs.
Nothing about liability
Every one of these devices is decision support. The interpreting radiologist reads, edits and signs, and that responsibility does not move because a cleared algorithm looked at the study first.
Questions buyers actually ask
FDA approved radiology AI, answered
Is AI in radiology FDA approved?
Mostly cleared rather than approved, and the difference matters. Almost all radiology AI reaches the US market through 510(k) clearance, which shows substantial equivalence to an existing device. "FDA approved" describes the far stricter premarket approval pathway that few imaging AI products use. As of the FDA list snapshot dated March 31, 2026, 1,164 of the 1,524 AI-enabled devices the agency has authorized since 1995 are radiology devices, about 76% of the total.
How many FDA approved AI medical devices are there?
The FDA AI-Enabled Medical Device List recorded 1,524 authorizations through the end of the first quarter of 2026, of which 1,164 were radiology devices. In that quarter alone the agency authorized 92 AI-enabled devices, 69 of them radiology. The FDA states plainly that the list is not comprehensive, because it is assembled from devices whose public authorization summaries use AI terminology.
Which radiology AI company has the most FDA clearances?
By count of entries on the FDA list, GE HealthCare leads with 130 authorizations, followed by Siemens Healthineers at 95 and Philips at 58. Among the software-only vendors a radiology group actually shortlists, Aidoc leads with 33, then DeepHealth at 29 and Rapid.ai at 20. Counting entries rewards vendors that ship many narrow device submissions, so treat it as a measure of regulatory throughput rather than of clinical value.
How do I check if a radiology AI product is FDA cleared?
Ask the vendor for the K number, then look it up yourself in the FDA 510(k) database. Read the Indications for Use statement rather than the marketing page, confirm the product code, and check the modality and body region the clearance actually covers. A vendor with fifteen clearances may hold none for the scanner or the population you plan to run it on.
What is the difference between FDA cleared and FDA approved?
Cleared means the FDA agreed the device is substantially equivalent to a legally marketed predicate, through the 510(k) pathway. Approved means the FDA reviewed clinical evidence of safety and effectiveness directly, through premarket approval, which is reserved for higher-risk devices. Radiology AI is almost always cleared. A vendor writing "FDA approved" on a 510(k) product is being loose with a regulated term.
Does FDA clearance mean the AI will be accurate on my patients?
No. Clearance shows the device performed acceptably on the datasets in the submission against a predicate, not that it will hold up on your scanners, protocols and patient mix. Ask for the test-set composition, the scanner makes it was validated on, and whether the vendor has a predetermined change control plan that lets it update the model without a new submission.
Do radiology AI devices still need a 510(k) in 2026?
Yes. On April 1, 2026 the FDA denied a petition, filed by Harrison.ai through Rubrum Advising, to exempt radiology CADe, CADx and CADt devices from premarket notification. The agency wrote that holding a 510(k) clearance "may not reflect proficiency in processes for all future devices of subject types" and that CADe and CADx cannot be evaluated interchangeably. Every new device still needs its own submission.
Why does a vendor claim more clearances than the FDA list shows?
Usually because the two are counting different things. The FDA list counts authorization entries attributed to a submitter name, while vendors often count cleared findings, indications or algorithm modules inside those entries, and acquisitions and name changes split a history across several submitter names. Neither figure is dishonest on its own. Ask for the K numbers and count those.
Next steps in the evaluation
Best AI radiology software compared
The vendor shortlist ranked by capability and fit rather than by regulatory paperwork.
Is radiology AI FDA cleared?
The shorter definitional answer, including what clearance means for a radiologist's signature.
Radiology software pricing
The five pricing models and real US federal contract award figures by software category.
AI triage for radiology
What a CADt product does to a worklist, and where the time actually gets saved.
Radiology AI vendors compared
Side-by-side pages for the vendors a US radiology group is most likely to shortlist.
How we research these pages
Our sourcing table, editorial standards and the conflict of interest we have to declare.
Clearance is the floor, not the decision
Radiological.ai is decision support built around the radiologist: it flags suspected findings for a second look, prioritizes the worklist and drafts the structured report. You review, edit and sign every study. We are pre-launch and we say so on every page rather than implying otherwise.