The FDA has authorised 723 AI radiology tools. 33 were tested prospectively.

Artificial intelligence has entered radiology faster than any other clinical speciality, and the count of cleared products is the statistic usually cited as evidence of that. A research letter published in JAMA Network Open this month counted the same devices and asked a different question: how many were tested in a way that tells you whether they help [s1].

What the review covered

The authors reviewed AI and machine-learning-enabled devices with FDA premarket authorisation from November 1995 to June 2024, following PRISMA-ScR guidance for scoping reviews, and extracted testing details from the summary submission documents [s1]. Data were analysed between September 2024 and January 2025 [s1].

They identified 950 authorised AI/ML devices in total, of which 723 — 76% — were radiology devices [s1]. Growth has been recent and steep: only 33 devices were authorised between 1995 and 2015, 3% of the total, against 221 in 2023 alone, 23% of the total [s1].

The testing numbers

Of the 717 radiology devices with available submission documentation, 33 (5%) underwent prospective testing, 56 (8%) included a human in the loop, and 208 (29%) incorporated clinical testing of any kind [s1]. Only 15 devices used both prospective and clinical testing; six included all three [s1].

The regulatory explanation is in the pathway. Of the 950 devices, 924 were cleared through 510(k), 22 through de novo and four through premarket approval [s1]. The 510(k) route establishes substantial equivalence to an existing predicate device rather than requiring independent clinical data demonstrating performance or safety [s1].

The mismatch the authors highlight is between how the devices are regulated and how they are used. Most AI tools in radiology operate alongside a radiologist, yet only 56 were tested with any human operator [s1]. They cite a study of radiologists reading chest radiographs with AI assistance in which strong readers maintained their performance while weaker readers did not necessarily improve — a pattern that makes the combined human-plus-model performance difficult to predict from the model's standalone metrics [s1].

What November's own clearances look like

The pathway described in the review is visible in the same month's decisions. On 6 November the FDA cleared Aidoc's BriefCase-Triage under product code QAS, "radiological computer-assisted triage and notification software", via a Special 510(k) — the abbreviated route for modifications to a manufacturer's own cleared device [s2]. On 24 November it cleared a2z Radiology AI's a2z-Unified-Triage under the same product code and regulation, 21 CFR 892.2080, as a Traditional 510(k) [s3]. Both were reviewed by the radiology advisory committee and both reached the market by demonstrating substantial equivalence [s2][s3].

Neither clearance implies anything is wrong with either device. They are ordinary examples of the mechanism the review describes: computer-assisted triage tools, which reorder a worklist so that suspected findings are read sooner, reaching patients through a pathway built to compare a new product with an existing one rather than to measure whether patients do better.

What the review does not show

This is a research letter, and its limitations are real. The analysis depended on the completeness of FDA summary documents, which vary; testing may have been performed and not reported [s1]. It excludes tools not subject to FDA clearance at all [s1]. And the absence of prospective testing is not evidence that a device fails — it is evidence that the question was not asked in the submission.

It also does not follow devices after clearance. The authors note that the FDA has released a draft Total Product Life Cycle framework offering more structured oversight with manufacturer accountability, but that such guidance is not consistently applicable, while the European Union has adopted postmarket monitoring frameworks under which manufacturers may use monitoring systems developed for non-AI technologies [s1]. They observe that radiology AI device approvals have declined in the EU under evolving regulation, alongside a rise in peer-reviewed studies of clinical outcomes [s1].

What it means for a reader

The number of cleared AI tools is not a measure of how much evidence exists that they work. The two are close to independent: 97% of authorisations rest on equivalence to a predicate, and each new predicate can then anchor the next clearance [s1].

The review's practical suggestion is that the gap has to be closed downstream of the FDA. It argues institutional governance is needed to determine whether a device has undergone adequate testing and confirm it functions as intended in that setting, and points to health technology assessment as a way to evaluate clinical and economic value after approval [s1]. In other words, the question of whether a given tool helps a given hospital's patients is currently one the hospital has to answer for itself.

What to watch

Whether the Total Product Life Cycle framework moves from draft guidance to consistent practice is the regulatory variable [s1]. The evidentiary one is simpler: whether the share of submissions including prospective, human-in-the-loop testing rises from 5% and 8% [s1] as the field matures, or stays where it is because the pathway does not require it.

Sources

• FDA Approval of Artificial Intelligence and Machine Learning Devices in Radiology: A Systematic Review — JAMA Network Open, 2025-11-07
• 510(k) Premarket Notification K253265 — BriefCase-Triage, Aidoc Medical Ltd — US Food and Drug Administration, 2025-11-06
• 510(k) Premarket Notification K252366 — a2z-Unified-Triage, A2z Radiology AI Inc — US Food and Drug Administration, 2025-11-24