How Are AI-Enabled Medical Imaging Platforms Regulated by the FDA?

AI-enabled medical imaging devices are regulated as medical devices whose software uses artificial intelligence, authorized through FDA's existing premarket pathways and tracked on the public AI-Enabled Medical Device List. The list exists, per FDA, to identify AI-enabled devices authorized for US marketing and show when devices use AI.

What is the AI-Enabled Medical Device List?

The list is a curated resource maintained by FDA's Digital Health Center of Excellence. Per the agency's AI-enabled devices page, the devices in the list have met FDA's applicable premarket requirements, including a focused review of overall safety and effectiveness, which includes an evaluation of study appropriateness for the device's intended use and technological characteristics. Each entry links to the FDA database record, which contains releasable information such as summaries of safety and effectiveness.

Two caveats come straight from the agency. The summaries are not all-inclusive and do not include most of what a sponsor may have submitted, and the list is not a comprehensive resource of every AI-enabled device; it was built primarily by identifying AI-related terms in marketing authorization summaries. Imaging dominates the list, which reflects where algorithms reached clinical maturity first: triage of CT and MRI studies, lesion detection and measurement, and image reconstruction.

How do imaging algorithms usually reach the market?

Most AI-enabled imaging devices enter through the 510(k) pathway, demonstrating substantial equivalence to a predicate device. Under the premarket notification framework, a submitter must receive an order finding the device substantially equivalent before marketing, and the comparison runs on intended use and technological characteristics. For algorithm products, predicates are frequently earlier-cleared versions of the same software, or a device of the same type, with performance bench data standing in for clinical trials.

The evidence expectations scale with claimed function. A device that flags a suspected large-vessel occlusion for prioritized review is validated on retrospective reader studies against reference standards; a device that quantifies a measurement must show agreement with the accepted method. What clearance does not confer is autonomy: most cleared algorithms are labeled as aids to the interpreting physician, whose read remains the diagnosis.

What makes AI devices different from other software devices?

Three properties distinguish them in regulatory terms. First, the model is trained on data, so the submission must describe the training, tuning, and test datasets and their separation. Second, performance is statistical, which is why summaries report sensitivity and specificity by use case rather than a single accuracy figure. Third, some models are designed to change after authorization, which engages FDA's discussion of predetermined change control plans: a sponsor specifies up front what parts of the model may update and under what limits, so that changes can occur within the cleared boundaries.

None of this required a new statutory pathway. FDA has so far handled AI devices through existing classifications, special controls where a De Novo created a category, and guidance on changing algorithms. The practical consequence is that an imaging algorithm's regulatory status is readable from its authorization letter and database summary, exactly as for any other device.

What should a buyer or hospital evaluator check?

The regulatory trail answers a short list of questions.

1. The authorization pathway and date, from the FDA database entry.
2. The stated intended use: triage aid, detection aid, quantification, or reconstruction.
3. The population and imaging conditions in the validation studies.
4. Whether the model is locked or operates under a predetermined change control plan.
5. The labeling's statement of the clinician's role in the final interpretation.

The AI list makes the first item a five-minute check, and the database summaries carry most of the rest. Devices absent from the list are not necessarily unauthorized, given the list's own stated limits, but absence plus no database record is a warning sign worth resolving before procurement.

What kinds of imaging algorithms are on the market?

The category is broader than detection tools. Cleared AI-enabled imaging devices span several functional families: triage and notification software that reprioritizes worklists for suspected findings, computer-aided detection and diagnosis tools that mark lesions for the radiologist, quantification software that measures structures or change over time, and reconstruction algorithms that use learned models to produce diagnostic images from less acquired data.

Each family carries its own validation logic. Triage tools are validated on sensitivity to the critical finding and time-to-notification; detection tools on reader performance with and without the algorithm, often in multi-reader multi-case studies; quantification tools on agreement with reference measurements; reconstruction tools on image quality and low-dose performance versus conventional pipelines. The intended use statement in the database summary tells the reader which logic applies.

The common thread is that the algorithm operates inside a clinical workflow that retains a human interpreter. Most authorizations are written as aids, and the summary documents state the studied use conditions, which is what a procurement evaluation should test against local practice.

What is a predetermined change control plan?

Some machine-learning models are designed to improve after deployment, retrained on new data. A predetermined change control plan is the mechanism by which a sponsor and FDA agree in advance on what may change: which parts of the model, trained on what data, validated by what method, with what limits. Changes inside the plan's boundaries can proceed without a new submission; changes outside them cannot.

The plan matters for imaging because model drift is real. Scanner fleets, protocols, and patient populations differ across sites, and a model that quietly degrades at the edge of its training distribution is a safety question, not just a performance question. The plan converts that risk into a documented, auditable schedule.

For evaluators, the practical question is whether a vendor's model is locked, adaptive under a plan, or unclassified on the point. A locked model behaves predictably but freezes its performance; an adaptive model under a plan carries governance obligations; an unspecified answer is a reason to pause the procurement conversation.

How should a hospital evaluate a cleared algorithm?

A disciplined evaluation separates the regulatory record from the local validation question. The sequence below reflects what the authorization documents can and cannot answer.

1. Pull the FDA database entry from the AI-enabled device list and read the intended use and study description in the summary.
2. Map the studied population, scanner platforms, and acquisition parameters against local practice.
3. Confirm the claimed function, triage, detection, quantification, or reconstruction, matches the clinical need.
4. Check whether the model is locked or operates under a predetermined change control plan.
5. Run a local silent-mode evaluation on the institution's own case mix before clinical reliance.

The last step is the one the regulatory record cannot supply. Clearance means the device met FDA's requirements for its stated intended use as studied, not that it will generalize to every scanner fleet and population. Institutions that treat the authorization as the beginning of their evidence, rather than the end, get the value these tools promise and avoid the failures that make headlines.

The regulatory posture also shapes what vendors can build next. Because authorization attaches to an intended use and studied conditions, expansion to a new imaging modality, a new patient population, or a new scanner platform is a new regulatory question, however similar the underlying model. The device list shows this pattern plainly: authorizations cluster by family, with each cluster grown through successive, individually cleared steps rather than a single sweeping approval. That granularity is slow by design, and it is what keeps the market auditable.