Topic 04 · AI & software · Deep dive
A thousand clearances, one paid code: the strange economy of imaging AI
Radiology owns medical AI on paper: 1,104 of the 1,451 AI-enabled devices the FDA has ever authorised — 76% — belong to imaging. But regulatory clearance has run far ahead of evidence, reimbursement, and daily use. As of 2025 exactly one imaging-AI application held a paid CPT Category I code, and where the tools are genuinely deployed at population scale, the results are only now arriving.
The FDA cleared its first radiology AI in 1998 (ImageChecker mammography CAD). It took two decades to reach 500 total AI devices — and three more years to nearly triple that. Radiology's share has stayed locked around three-quarters throughout.
1998: where it started
The first cleared radiology AI on the FDA's list is ImageChecker, R2 Technology's mammography CAD (later Hologic) — a lineage that means today's clearance boom is CAD's second act, not its first.
6 → 295 per year
Annual FDA AI/ML clearances grew ~50-fold between 2015 and 2025. In 2025, 221 distinct manufacturers received at least one clearance — 183 of them exactly one. Only 9 companies managed four or more.
Foundation models arrive
In January 2026 Aidoc secured a foundation-model-powered clearance — a single body-CT triage tool covering 14 conditions (aortic dissection, appendicitis, bowel obstruction and others). The FDA is now working out how to tag foundation-model and LLM-based devices on its list.
The Radboud-founded Health AI Register (formerly aiforradiology.com) has tracked the European market since 2020. Its two landmark reviews measure both the market's growth and its persistent weakness: most published validation never rises above technical accuracy.
US Medicare remains the world's reference market for paying for imaging AI, and its verdict so far is stark: clearance is table stakes, payment requires outcome evidence, and outcome evidence is rare.
Why most will never be paid
CPT codes describe distinct procedures. Detecting fractures, nodules or incidentals is already inside the paid interpretation — the ACR's CPT advisor argues separate codes for that work would mean paying twice. ROI must come from throughput, not fee schedules.
The bar that two tools cleared
FFR-CT and CCTA plaque analysis earned Category I status the slow way: years as T-codes plus outcome trials showing added diagnostic value beyond the underlying scan. Both create new information no radiologist reports unaided.
Europe has no pathway at all
The EU leads on regulation (MDR, AI Act) but has no unified reimbursement route — adoption rides on hospital budgets and national innovation funds, and the AI Act's high-risk obligations land in 2026–27.
Regulation was never the bottleneck
The favourite statistic of every imaging-AI pitch deck — radiology's three-quarter share of all cleared medical AI — is true, and it is also the least informative number in this topic. The 510(k) pathway, through which nearly 95% of these devices pass, requires substantial equivalence to a predicate, not proof of patient benefit; median review took 142 days in 2025, and a quarter of submissions cleared in under 90. The result is a register that grew from roughly 500 devices at the start of 2023 to 1,451 by the end of 2025, in which equipment giants clearing embedded reconstruction and workflow modules sit beside 183 single-clearance startups. Counting clearances measures regulatory throughput. It does not measure clinical AI.
A register is not an evidence base
The Radboud group's longitudinal reviews of the European market are the closest thing the field has to an audit, and they show a catalogue outrunning its footnotes: products up 73% between 2020 and 2023, peer-reviewed validation up from 36% to 66% of products — but the share of evidence demonstrating actual clinical, outcome, or economic impact stuck at about a quarter. The US picture is harsher still: across 692 FDA-cleared devices from 1995–2023, 3.6% reported the race or ethnicity of their validation cohorts. For a purchaser, the practical implication is that the burden of local validation still sits with the buyer, which is exactly the cost the clearance count invites you to forget.
One paid code
The reimbursement funnel is this topic's honest headline. Of more than a thousand cleared radiology devices, exactly one held a paid CPT Category I code in 2025 — FFR-CT — with coronary plaque analysis joining it in January 2026. Both are cardiac CT applications that generate genuinely new information, validated through years of outcome trials while parked as temporary Category III codes. The ACR's own CPT advisor is blunt about the rest: fracture, nodule and incidental-finding detectors describe work already inside the paid interpretation, and coding them separately would mean paying twice. That leaves throughput, triage value, and add-on programmes like NTAP — capped, in the FY2025 hydrocephalus example, at $241.39 per inpatient case and expiring within three years — as the business case for nearly the entire catalogue.
Where it is used, it works — which sharpens the question
The strongest counterweight to the scepticism above is mammography screening, now the best-evidenced deployment in imaging AI. The MASAI randomised trial cut screen-reading workload 44% while detecting 29% more cancers — predominantly small, node-negative invasive disease — and its 2026 interval-cancer results confirmed non-inferiority with higher sensitivity. Germany's PRAIM study replicated the direction of effect across 463,094 real-world screens. Yet self-reported clinical use sits at 48% among surveyed European radiologists and around 30% in older US data, with response rates low enough that both are upper bounds. The gap between 1,104 clearances and this thin band of proven, adopted, and (rarely) paid applications is the defining statistic of radiology AI in 2026 — and the reason clearance counts should never be quoted as adoption.
On the data. The FDA list counts devices, not products: imaging hardware with embedded AI is included, vendor suites are sometimes split and sometimes bundled, and totals shift retroactively as the list is curated. Health AI Register counts are vendor-supplied with different inclusion criteria, so US and EU figures are not directly comparable. The "~12 with a Medicare payment route" figure is a modelled order-of-magnitude estimate, not an official census. Survey adoption numbers carry severe self-selection risk (the ESR survey's response rate was 2%), and the ACR 30% figure predates the current survey cycle. PRAIM was observational with radiologist-elected AI use. All figures span 2020–2026 vintages as labelled.