Understanding FDA 510(k) concepts for pathology AI software

A 510(k) is a U.S. premarket notification pathway for certain medical devices. It is not a generic approval of artificial intelligence, and a 510(k) discussion cannot establish the regulatory status of a particular pathology software product without reviewing that product's intended use, classification, evidence, labeling, and FDA decision.

The central concept is substantial equivalence

FDA describes a 510(k) as requiring a demonstration of substantial equivalence to a legally marketed predicate device. The comparison considers intended use and, where technology differs, whether the differences raise different questions of safety and effectiveness. The predicate comparison is specific to the submission and does not mean that two products are identical.

Intended use controls the regulatory question

The intended use describes what the software does, for whom, with which inputs, and in what setting. Marketing language can contribute to how intended purpose is understood. Descriptions such as detection, diagnosis, triage, measurement, risk prediction, and decision support should therefore be reviewed with regulatory and clinical specialists rather than treated as interchangeable terms.

Evidence must match the claimed use

A performance plan should reflect the intended population, specimen and image conditions, reference standard, clinically relevant thresholds, and expected users. Depending on the device and claim, evidence may include analytical, technical, and clinical performance testing. A result from a curated research dataset is not by itself evidence that software is safe or effective for clinical diagnosis.

Labeling and workflow limitations matter

A submission includes information such as indications for use, specifications, performance, and proposed labeling. Labeling should identify intended users, inputs, outputs, warnings, limitations, and conditions in which the software should not be used. A concurrent decision-support output should be framed as information for a qualified professional, not as an autonomous diagnosis or a replacement for professional judgment.

For pathology AI, the safest public framing is precise: explain the regulatory concepts, state whether the discussion is educational, identify the intended-use caveat, and avoid presenting any software output as a diagnosis.