Every one of the 212 product codes FDA reviews under its radiology panel, positioned by how long a 510(k) takes and how many have ever cleared. Sixty-one of them have never been cleared through 510(k) at all. One of those is the code FDA moved on August 10.
A device class is a description of what evidence you have to bring. Class III means a premarket approval application. Class II means a 510(k) plus special controls. On August 10 the FDA proposed moving digital breast tomosynthesis — product code OTE — from one side of that line to the other.
The chart below is the whole neighborhood OTE is about to move into.
Horizontal position is the median days from submission received to decision. Vertical position is total 510(k) clearances ever, on a square-root scale. Circle size is clearances since 2016. Codes with no 510(k) history sit in the shaded rail on the left — there is no review time to plot, because nothing has ever been reviewed.
Sixty-one radiology product codes have zero 510(k) clearances in the entire FDA database. Some are obsolete — pneumoencephalographic chairs, betatrons, film dryers. Some are Class III, where 510(k) is not the pathway. And four of them are brand-new software codes FDA created before anyone shipped a product.
| Code | Device name | 510(k)s |
|---|
Pool every radiology 510(k) ever decided and split it by whether the product code is one of the sixteen image-analysis software codes. The AI median is 143 days. Everything else is 79 days. That is an 81 percent gap, and it is the kind of number that ends up on a conference slide.
It is also mostly an artifact. Fifteen of those sixteen codes were created in 2018 or later, so the comparison is 2018–2026 submissions against a baseline that runs back to 1976 — and radiology review times rose sharply after 2022 for everyone. Hold the year constant and watch.
The gap does not vanish. It halves — from 64 days to 37 — and in 2024 it is 12. Most of what looked like “FDA is slower on AI” was “FDA got slower, and AI arrived late.”
Buried at 91 FR 51412, FDA writes that task-based studies using anthropomorphic phantoms with structured background “may serve as an alternative to clinical studies.” The codified special control names three acceptable routes to diagnostic accuracy:
“Objective task-based assessment of diagnostic accuracy of the device, conducted using human subjects, structured physical phantoms, or in silico methodologies, or a combination of these approaches.”
The order defines an in silico trial as a study estimating device performance “based on computational modeling in a virtual population.” That is a regulator putting build the evidence on the same list as collect it — for the platform installed in 94 percent of the 9,107 MQSA-certified facilities in the country.
And the order contains zero instances of “artificial intelligence,” “machine learning,” or “computer-aided detection.” The whole software special control is: “Software verification, validation, and hazard analysis must be performed.” Nine words. Every red dot in the fourth quadrant of that scatter reads pixels produced by this platform.
A product code is not a market, and an empty rail is not an empty category. OTE has zero 510(k)s because DBT is Class III. Four original PMAs and twenty-six PMA supplements were reviewed under that code and none of them appear anywhere in this chart. The 510(k) database only sees one of the two doors.
Median review time is not review burden. Days from receipt to decision includes every FDA hold clock and every sponsor response. A 250-day median on a code with seven clearances tells you almost nothing; raise the min-clearances slider and the far-right cloud thins out fast.
The AI bucket is hand-built. Sixteen codes selected by name. QIH — “Automated Radiological Image Processing Software” — is 182 of the 354 and covers everything from a quantification utility to a triage model. Redraw the boundary and the medians move.
2026 is partial. The file runs through decision date 24 Jul 2026. The 2026 column in the year-by-year view is roughly seven months.
A proposed order is not a rule. One product code, open for comment until October 9. Not a clearance, not a guidance, not a green light for simulated AI evidence.
If FDA will accept a virtual clinical trial for a tomosynthesis system, what is the principled argument against accepting one for the model reading its output — and who builds the virtual patient population both of them would have to agree on?
The docket closes October 9.