The Food and Drug Administration drew a firm line this month around how much autonomy AI radiology software can earn before regulators sign off on it. A final order that took effect Thursday, September 17, requires several categories of AI-enabled imaging software — tools that flag suspicious cancer lesions, analyze scans and alert clinicians to urgent findings — to keep going through the agency’s traditional premarket clearance process before they can be sold to hospitals.
The petition that started it
The order formalizes a decision the FDA actually made months earlier. On April 1, the agency denied a citizen petition filed by Harrison.ai, a Sydney, Australia-based healthcare AI company, which had asked for a partial exemption from the 510(k) premarket-notification pathway for four categories of radiology software: computer-aided detection and diagnosis tools for suspicious cancer lesions, image-analysis systems, and triage-and-notification software that flags potentially urgent findings to clinicians. Harrison.ai argued that once a manufacturer already held clearance for a related device and maintained robust postmarket monitoring and clinician training programs, later updates to that software shouldn’t need to restart the full review process.
Why the FDA said no
The agency’s rationale was blunt. In its response, the FDA stated that “the petition and public comments did not demonstrate that premarket notification was unnecessary to provide reasonable assurance of safety and effectiveness.” In plain terms: internal company quality-control systems and after-the-fact monitoring are not substitutes for an independent government review before a diagnostic tool reaches a radiologist’s workstation. The September order effectively closes the door on that argument by codifying it as an enforceable rule rather than a one-off denial.
A market that has exploded in size
The stakes are unusually large because radiology is, by a wide margin, the epicenter of medical AI regulation. As of earlier this year the FDA had cleared more than 1,500 AI-enabled medical devices total, and radiology algorithms accounted for roughly three-quarters of them — over 1,160 cleared products. The agency has been authorizing new radiology AI tools at a pace of around 30 a month, with 68 new algorithms cleared in just the first quarter of 2026 alone. That volume is exactly why a company like Harrison.ai wanted a faster lane: for firms shipping frequent software updates, each version can trigger a fresh, costly and time-consuming review cycle.
Industry pushback and patient-safety defenders
Harrison.ai and other AI developers have long argued that iterative software, especially machine-learning models that improve with more data, doesn’t fit neatly into a regulatory framework built for static hardware. Forcing every meaningful update through 510(k) review, they contend, slows down improvements that could catch more cancers or reduce false alarms sooner. Patient-safety advocates and the FDA itself take the opposite view: radiology AI increasingly influences decisions about biopsies, urgent transfers and treatment triage, and a confident but wrong output can cause real harm at scale precisely because these tools run across thousands of patients a day. Without mandatory premarket review, critics argue, there would be no independent checkpoint before an updated algorithm starts making those calls in real hospitals.
What changes on the ground
For hospitals and health systems, the immediate effect is continuity rather than disruption — the rules they’ve operated under stay in place, so radiology departments using cleared AI tools for chest CTs, mammography or stroke imaging won’t see new gaps in oversight. For AI vendors, though, the order removes any hope of a shortcut: software updates covered under the four exempted categories still require 510(k) submission and clearance before hospitals can deploy them, which affects product roadmaps and revenue timing across the sector.
What’s next
The decision is likely to sharpen a broader debate in Washington over how the FDA should regulate AI that updates itself faster than any premarket pathway was designed to handle. Expect continued lobbying from AI device makers for a dedicated “predetermined change control” framework that allows pre-authorized types of algorithm updates without full resubmission — a middle path the FDA has floated in guidance documents but not yet finalized into binding rules. Until that framework exists, the message from this order is unambiguous: in the highest-volume corner of medical AI, human regulators still get the final word before software reaches the reading room.
Photo: cottonbro studio / PEXELS via Pexels
