FDA refuses to dispense screening radiological detection software: radiologist cannot see drift

With a petition to exempt four categories of image detection and sorting aids from 510(k), the Agency responded that there was no guarantee that a human reader would detect a discrepancy before it produced a diagnostic error.

The final order issued by the FDA on September 17, 2026, ends a request filed by Harrison.ai for a partial exemption from premarket notification — the so-called 510(k) procedure — for certain artificial intelligence software used in radiology. The petition covered four generic categories of devices and six distinct product codes, including computer-aided diagnostic software for suspicious cancer lesions, medical image analyzers, and triage and reporting tools.

The central argument of the petitioner was that any evolution of such a device likely to affect its safety or effectiveness would be detected without difficulty by its users, qualified radiologists, either by visual examination of the images that accompany the release of the software, or by routine tests. The agency rejects this reasoning on a specific point: for a given patient, the concordance between the reader's interpretation, the output of the software and the actual state of the disease is not necessarily known, or even available, before this release has led to a diagnostic error. In other words, the human control invoked does not arrive in time.

The FDA is not completely closing the door and is referring manufacturers to two existing systems: predetermined change control plans, which make it possible to supervise certain changes to a model in advance without going through a full submission, and its Q-Submission program to discuss with the agency in advance. The decision comes at a time when more than 1,600 medical devices incorporating AI have been authorized for marketing in the United States, the vast majority of which have been authorized through the 510(k) route.