clinicians.build · interactive · august 11, 2026

Zero

On August 10 the FDA proposed moving digital breast tomosynthesis from Class III to Class II. Product code OTE has never had a single 510(k) clearance. Here is the whole mammography lineage, one dot per clearance, 1983 to now — and what happened the last time FDA did this.

Primary source: Radiology Devices; Reclassification of Digital Breast Tomosynthesis System
FDA proposed order · 91 FR 51406 · Docket FDA-2026-N-7630 · published August 10, 2026
Grounding data: MIMI Labs · FDA 510(k) release database + device classification file, vintage through 24 Jul 2026

A device class is not a safety rating. It is a description of what evidence you have to bring. Class III means a premarket approval application — your own clinical data, reviewed from scratch. Class II means a 510(k): show you are substantially equivalent to something already cleared, and meet the special controls.

FDA is proposing to move digital breast tomosynthesis across that line. Comments close October 9, 2026. To see what the move means, look at the paperwork trail of every device code in the mammography lineage.

Four lanes, one dot per 510(k) clearance

Each dot is one FDA 510(k) clearance with decision Substantially Equivalent, plotted by decision year. Press play and watch the lanes fill.

FDA 510(k) database · radiology panel · decisions 1983–2026 · 497 clearances shown
1983
IZH mammo x-ray
0
MUE full-field digital
0
OTE tomosynthesis
0
AI / CAD software
0
1983. Screen-film and early digital mammography systems clear through 510(k) at a steady clip. Tomosynthesis does not exist yet.

The lane that never fills

OTE stays empty for the entire chart. Not because nothing happened — four original premarket approval applications and twenty-six PMA supplements were reviewed under that code, starting with Hologic’s Selenia Dimensions 3D on February 11, 2011. But a PMA is not a 510(k), and the 510(k) database has nothing to show.

That is what a Class III platform looks like from the outside: fifteen years of installed base, ninety-four percent market penetration, and a submission pathway only four original applications have ever walked.

The last time this happened

Full-field digital mammography — product code MUE — is the precedent FDA cites by name in the order: “FFDM systems…are currently regulated as class II devices with special controls and have been without any significant safety signals since 2010.”

FDA published that reclassification as a final rule on November 5, 2010, codified at 21 CFR 892.1715. Look at the MUE lane. It is empty until 2011, then it fills. Forty clearances in fifteen years, at a median of 169 days from receipt to decision. Reclassification did not open a floodgate. It opened a door, and a modest number of applicants walked through it slowly.

MUE first 510(k): 2011 · 40 clearances through Jul 2026 · median review 169 days · 24 of the 40 landed in the last decade

The sentence that matters

Buried at 91 FR 51412, FDA writes that task-based performance studies using anthropomorphic phantoms with structured background “may serve as an alternative to clinical studies.” The codified special control then lists three acceptable ways to show diagnostic accuracy:

Proposed special control 1(i)(B) · 91 FR 51415

“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.”

An in silico trial is defined in the order itself as a study estimating device performance “based on computational modeling in a virtual population.” A regulator just put build the evidence on the same list as collect it.

And the order contains zero instances of “artificial intelligence,” “machine learning,” or “computer-aided detection.” The entire software special control is one sentence: “Software verification, validation, and hazard analysis must be performed.”

Meanwhile, in the fourth lane

The red lane is the sixteen radiology product codes FDA created for image-analysis software — automated processing, computer-assisted triage, CADe/CADx, AI-guided acquisition. All but one were established in 2018 or later. Four clearances in 2018. 354 by July 2026.

Every one of those devices reads pixels produced by a platform whose reconstruction pipeline FDA just proposed governing with nine words.

Four of those codes are empty

FDA built the shelf before anyone shipped a product to put on it. These four radiology software product codes exist and have never had a clearance:

CodeDevice name510(k)s
⚠︎ The one to notice is QVD. FDA created a product code specifically for machine-learning imaging software with a predetermined change control plan — and it has never been used. In today’s DBT order, PCCPs appear in the preamble discussion and nowhere in the codified special controls. The mechanism for governing a model that changes after clearance exists on paper in two places, and is empty in both.

Where this reading is thin

Stress-test the chart before you quote it

A product code is not a market. OTE has zero 510(k)s because DBT is Class III, not because nothing was submitted. The four PMAs and twenty-six supplements are real regulatory work that this database does not show. An empty lane is an artifact of which door the submission went through.

2026 is a partial year. The 510(k) file used here runs through decision date 24 Jul 2026. Every 2026 bar in this chart is roughly seven months of a twelve-month year. Do not read the last column as a decline.

“AI/CAD software” is a hand-built bucket. Sixteen product codes, chosen by name. QIH alone — “Automated Radiological Image Processing Software” — is 182 of the 354 and covers everything from a quantification tool to a triage model. Product-code counts flatter the AI story because FDA minted new codes at exactly the moment the category grew.

A proposed order is not a rule. This is one product code, open for comment until October 9. It is not a clearance, not a guidance, and not a green light for simulated evidence for AI. Do not put it on a slide that way.

The question

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 comment docket is open until October 9. It is the cheapest lobbying a two-person company will ever do.