FDA's AI-enabled medical-device list showed 1,164 radiology entries out of 1,524 by Q1 2026; those are listed/authorized entries, not necessarily 1,164 specifically 510(k)-cleared products.
Last updated: July 6, 2026.
This is regulatory and procurement analysis, not medical or legal advice.
Radiology owns the FDA AI list. By the end of Q1 2026, trade-press trackers of the FDA list counted radiology at 76.3% of all listed AI/ML-enabled devices, according to The Imaging Wire's June 2026 tracker and the FDA's AI-enabled medical devices page.
That count is useful. It is also easy to misuse.
For a radiology AI product that is specifically 510(k)-cleared, clearance means marketing authorization for a defined intended use through substantial equivalence. The FDA AI-enabled device list is broader: it identifies AI-enabled medical devices authorized for marketing, including 510(k), De Novo, and PMA pathways.
Neither a list entry nor a clearance proves outcome benefit, site-level generalization, low alert burden, reimbursement fit, or safe long-term model updates.
TL;DR
FDA authorization is a regulatory floor. For radiology AI buyers, the live question is whether the exact authorized claim and pathway match your patient population, scanners, workflow, cybersecurity requirements, and monitoring capacity.
The FDA list reached 1,524 AI/ML-enabled device entries by Q1 2026, and radiology supplied 1,164 of them, according to trade-press trackers of the FDA page. Those are dated list-entry counts drawn from FDA's public authorization list; they are not a claim that every radiology entry is a 510(k)-cleared product.
The FDA itself says it updates the list periodically based on publicly available information, so exact dated counts should be treated as tracker snapshots of the primary list, not as a procurement verdict.
Key takeaways
- Radiology had 1,164 of 1,524 entries on FDA's AI-enabled medical-device list by Q1 2026, or 76.3%.
- That 1,164 figure is a list-entry count across authorization pathways, not a claim that every entry is specifically 510(k)-cleared.
- Tracker data in the research dossier put 510(k) at roughly 96-97% of AI/ML device authorizations in mid-2025, while De Novo and PMA were much smaller shares.
- FDA authorization does not establish ROI, local scanner performance, alert-fatigue tolerance, or outcome benefit.
- PCCP matters because it defines which model changes can happen without a fresh submission, but the research dossier found PCCP references on only roughly 10% of listed devices as of April 2026.
- QMSR became effective on February 2, 2026, so buyers should ask for ISO 13485:2016-aligned quality evidence.
FDA-Listed Radiology AI Is Everywhere
The FDA AI/ML device list grew from 343 entries at launch in September 2021 to 1,524 by March 2026, according to IQVIA's 2021 coverage, a May 2025 tracker, and The Imaging Wire's Q1 2026 update.
Radiology dominates because imaging has the right substrate for regulated software: DICOM images, PACS archives, modality metadata, and measurable endpoints. That makes retrospective testing easier than it is in many other clinical domains.
The vendor market reflects that pull. GE HealthCare led with 130 cumulative AI authorizations in Q1 2026, followed by Siemens at 95 and Philips at 58, according to The Imaging Wire's vendor ranking.
But parent-company counts hide the product-level decision. Procurement should decompose every platform pitch into K-number, intended use, modality, workflow role, and deployment site.
What Does FDA Cleared Mean vs Approved?
FDA authorization is not one pathway. For radiology AI, a 510(k) clearance means substantial equivalence to a legally marketed predicate, while De Novo grants and PMA approvals follow different standards. The authorized claim applies only to the labelled intended use, such as triage or detection in a specific modality, and does not imply autonomous diagnosis.
Most radiology AI products reach market through 510(k) clearance, with smaller numbers granted through De Novo classification or approved through PMA. In FDA usage, approval usually refers to PMA, the high-risk Class III pathway described in 21 CFR Part 814.
The 510(k) route is different. FDA asks whether the new device has the same intended use as a predicate and does not raise different questions of safety or effectiveness. The FDA's Artificial Intelligence in Software as a Medical Device page places these products inside the broader SaMD framework.
For buyers, the intended-use statement is the contract anchor. A tool cleared for triage notification of suspected intracranial hemorrhage on non-contrast head CT has not been cleared for general neurodiagnosis, radiologist replacement, or broad emergency-department prioritization.
| Pathway | What it usually means | AI buyer implication |
|---|---|---|
| 510(k) clearance | Substantial equivalence to a predicate | Most common route; inspect intended use and predicate carefully |
| De Novo classification | Novel low-to-moderate-risk device without a predicate | Often creates the category future 510(k)s cite |
| PMA approval | High-risk Class III approval | Rare for standalone imaging AI software |
De Novo still matters because a novel low-to-moderate-risk device can create a classification category that later 510(k) submissions may cite. For this article, treat named De Novo examples as starting points for FDA-record lookup, not as a substitute for checking the current device file.
How Many FDA-Listed AI Radiology Products Are There In 2026?
There were 1,164 radiology entries on FDA's AI/ML-enabled medical-device list by the end of Q1 2026. That was 76.3% of 1,524 total listed entries, but it was a list-entry count across authorization pathways, not a count of 1,164 specifically 510(k)-cleared products.
Searchers often say "approved" when they mean cleared, authorized, or listed. The distinction matters because most radiology AI products are not PMA-approved, and some FDA-listed AI-enabled devices reached market through De Novo or PMA rather than 510(k).
The FDA's public AI-enabled medical devices list is the right starting point, but it is not a deployment guide. FDA says it updates the list periodically based on publicly available information.
Two pathway numbers are useful for diligence, but should be treated as tracker estimates rather than FDA legal categories. A tracker cited in the research dossier estimated that roughly 96-97% of AI/ML device authorizations were 510(k) clearances in mid-2025, while a PCCP implementation guide cited in the dossier found PCCP references on roughly 10% of listed devices by April 2026.
The clearance count also says little about clinical maturity. A mammography CADe tool, stroke LVO triage notifier, chest X-ray abnormality detector, and reporting assistant can all sit under radiology while carrying different operational risks.
| Category | Common use | Buyer risk to test locally |
|---|---|---|
| Stroke LVO / ICH triage | Time-sensitive alerting | Time-to-action, neurologist review, false positives |
| PE triage | Acute CTA prioritization | Scanner mix, alert burden, subgroup performance |
| Mammography / DBT | Detection support | Reader workflow, recall-rate effect, threshold tradeoff |
| Fracture detection | Radiograph triage | Missed fracture classes, pediatric or geriatric fit |
| Report automation | Drafting or workflow support | Hallucination risk, sign-off controls, audit logs |
A product's panel category is only a label. The purchasing decision lives inside the labelled claim.
How To Read 510(k) Summary AI Medical Device
Read the 510(k) summary as a bounded engineering artifact: intended use, predicate, performance testing, labelling, and workflow role. 21 CFR 807.92 sets summary content requirements, but buyers often need vendor-supplied evidence for subgroup, scanner, and prospective validation.
Start with the exact intended use. Look for patient population, modality, body region, clinical task, workflow setting, and the required human role.
Then inspect performance. A useful summary may disclose sensitivity, specificity, AUC, reference standard, and comparator logic, but buyers should avoid assuming every summary includes all the fields they need. The research dossier flags 21 CFR 807.92 as the content requirement for 510(k) summaries, while also noting that demographic subgroup reporting is not uniformly enforced.
The missing fields usually matter most. Many summaries do not provide prospective validation, multi-site external validation, demographic subgroup performance, scanner-protocol stratification, or real-world drift results.
A buyer-grade review should ask for these artifacts before contracting:
- Exact K-number and FDA summary.
- Dataset site count, geography, scanner vendors, and protocol mix.
- Standalone sensitivity, specificity, AUC, and operating threshold.
- Subgroup results by age, sex, race or ethnicity, and key comorbidities.
- Human-factors evidence for the workflow claim.
- Local silent-mode validation before clinical use.
- Post-deployment monitoring thresholds and escalation rules.
- MDR, recall, and MAUDE history for the product family.
This is where generic "FDA cleared" language fails. A cleared product can still be a poor fit for a hospital whose imaging protocols, disease prevalence, or staffing model differs from the validation set.
PCCP And QMSR Changed Diligence
PCCP is the 2026 buyer's pressure test for adaptive AI. FDA's December 4, 2024 final guidance on Predetermined Change Control Plans lets manufacturers pre-specify certain modifications, validation methods, and update controls.
That matters because model updates are routine in production software. Scanner firmware changes, acquisition-protocol changes, retraining data, and threshold adjustments can all change device behavior.
A PCCP gives the vendor an approved envelope. It does not permit arbitrary retraining, off-label expansion, or silent changes outside the declared modification scope.
The quality-system baseline also moved. FDA's QMSR final rule was published on February 2, 2024, and FDA's QMSR page says the regulation became effective on February 2, 2026, incorporating ISO 13485:2016 by reference.
Cybersecurity belongs in the same diligence file. FDA's medical-device cybersecurity FAQ and the 2025 Federal Register cybersecurity guidance notice point buyers toward SBOMs, vulnerability management, patch cadence, and incident response. This article uses those sources as procurement diligence prompts, not as a standalone legal conclusion about a particular submission.
Clinical Validation Is The Real Purchase
Clearance proves the vendor got through a regulatory gate. Clinical AI validation proves the model is useful at your site.
The failure mode is familiar to anyone who has deployed ML outside the training distribution. A model that performs well on one hospital's scanners and case mix can degrade when moved to a new scanner fleet, new protocol, different prevalence, or different patient population.
The literature gives buyers a reason to insist on local evidence. Work indexed by the NCI on AI recognition of patient race in medical imaging showed that imaging models can learn demographic signals that clinicians cannot reliably see.
The research dossier also cites subgroup and site-transfer studies showing that model performance can degrade materially across sites, scanner vendors, protocols, and patient mixes; the important procurement point is the direction of risk, not a universal numeric penalty.
Silent-mode evaluation is the practical minimum. Run the model without affecting care, compare outputs to local ground truth, stratify performance, and measure alert volume before go-live.
Outcome claims need a higher bar. Stroke triage tools may plausibly improve door-to-needle or door-to-thrombectomy time, but that evidence does not automatically transfer to PE, fracture detection, mammography, or report automation.
Procurement Checklist For Radiology AI
A buyer-grade dossier should force every vendor into the same evidence format. Otherwise, platform language will drown the device-level facts.
| Dimension | Ask this | Good evidence |
|---|---|---|
| Intended use | Does the cleared claim match our clinical task? | FDA summary and label match the deployment scenario |
| Population fit | Did validation include patients like ours? | Subgroup sensitivity and specificity tables |
| Scanner mix | Were our modalities and protocols represented? | Multi-site, multi-vendor test evidence |
| Workflow | Who sees the alert and who acts? | Human-factors evidence and escalation policy |
| PCCP | Which updates are pre-specified? | Documented modification scope and validation rules |
| Cybersecurity | How are SBOMs and vulnerabilities handled? | SBOM current within 6 months and patch SLA |
| Monitoring | How is drift detected after go-live? | Monthly dashboard with thresholds and review owner |
| Liability | What happens after a miss or recall? | Contractual obligations, indemnity, recall workflow |
Do not accept aggregate AUC as the main answer. AUC can hide threshold behavior, subgroup failures, and a false-positive rate that overwhelms a busy reading room.
Also separate diagnostic performance from workflow performance. A tool can detect a target well and still slow the radiologist down if the alert appears in the wrong system, arrives too late, or forces duplicate review.
What This Means For You
If you're a health-system buyer, treat radiology AI like clinical infrastructure. The product touches patient prioritization, diagnostic attention, audit trails, cybersecurity, and liability.
Start with the FDA record, then build a local evidence file. The record tells you what the vendor is allowed to market. Your validation tells you whether the model belongs in your workflow.
If you're a founder, a 510(k) is no longer enough for sophisticated buyers. You need a clean intended-use story, PCCP strategy, subgroup evidence, deployment monitoring, and integration proof in PACS, RIS, and reporting systems.
And if you're a radiologist, insist on version visibility. Every alert should be traceable to a model version, input study, timestamp, user action, and review outcome.
FAQ
Are large language models cleared by FDA for radiology reporting?
As of the July 6, 2026 research snapshot, the dossier found no standard-pathway FDA clearance for a generative or agentic radiology AI product. Report-automation products may have regulated components.
For detection, diagnosis, or autonomous-reporting claims, treat this as editorial diligence: verify the exact device, intended use, pathway, and authorization status on FDA's list and the applicable 510(k), De Novo, or PMA record before relying on it.
Is Viz.ai stroke AI covered by Medicare?
No blanket answer, and this article is not asserting current coverage for any specific deployment. Some stroke-triage technologies have been associated with CMS New Technology Add-on Payment pathways in defined inpatient contexts, but that does not mean product-wide Medicare coverage, every-hospital reimbursement, outpatient coverage, non-stroke coverage, or off-label coverage.
Buyers should verify the current CMS IPPS rule, eligible codes, hospital billing fit, and product-specific indication through the CMS FY 2026 IPPS page and their reimbursement counsel.
What does PCCP cover in AI radiology?
A PCCP can cover pre-specified model modifications, validation methods, and update controls that FDA reviewed as part of the marketing submission. It does not cover arbitrary retraining, off-label expansion, or silent changes outside the declared modification scope.
Is FDA clearance enough for clinical deployment?
No. Clearance is necessary for marketing many medical-device claims in the United States, but deployment needs local validation, workflow testing, cybersecurity review, quality-system checks, and post-market monitoring.
Sources
- FDA Artificial Intelligence-Enabled Medical Devices
- The Imaging Wire: Top 10 radiology AI vendors by FDA authorizations
- The Imaging Wire: FDA updates AI list with new clearances
- IQVIA: FDA publishes AI/ML-enabled medical device list
- FDA: Artificial Intelligence in Software as a Medical Device
- Federal Register: PCCP marketing submission recommendations
- Federal Register: QMSR final rule
- 21 CFR Part 814: Premarket Approval of Medical Devices
- FDA: Quality Management System Regulation
- FDA: Cybersecurity in medical devices FAQ
- Federal Register: Cybersecurity in medical devices guidance
- NCI CDAS: AI recognition of patient race in medical imaging
- CMS FY 2026 IPPS proposed rule page
- Plainstamp: FDA PCCP-AI builder's guide
- LinkedIn tracker: 1247 FDA Authorized AI-Enabled Medical Devices