When a radiology AI vendor tells you their product has "FDA clearance," your next question should be: clearance for what, exactly? This isn't a gotcha question. It's the question that determines whether the cleared indication actually covers the way you plan to use the tool.
The 510(k) premarket notification pathway is the most common FDA clearance route for AI radiology software. Understanding what 510(k) clearance means — and what it doesn't — is increasingly important for hospital imaging departments, because the market now includes cleared tools being marketed for uses adjacent to but not identical to their cleared indication, and non-cleared tools being deployed in ways that create institutional liability exposure.
How 510(k) Clearance Works for Software Devices
A 510(k) submission does not require the manufacturer to prove their device is safe and effective in an absolute sense. It requires them to demonstrate substantial equivalence to a legally marketed predicate device. For a new AI radiology tool, this typically means identifying a prior cleared CAD or decision support product with a similar intended use, and showing that the new device performs comparably on a defined set of performance metrics.
The FDA evaluates the submission — typically over 3 to 12 months for a standard 510(k) — and either clears the device, issues an additional information request, or determines that substantial equivalence has not been demonstrated (at which point the manufacturer would need to pursue a De Novo classification or the more rigorous PMA pathway).
What gets cleared is the specific indicated use described in the submission. A device cleared for "detection of pulmonary nodules on low-dose CT in adults aged 50-80 with a smoking history of 20+ pack-years" has been cleared for exactly that. Using the same device for incidental nodule detection on CT angiography studies performed for pulmonary embolism workup, or for nodules in patients outside the specified age range, is an off-label use. The FDA generally allows clinicians to use cleared devices off-label at their discretion, but the institutional liability environment is different, and the performance data validating that specific use doesn't exist.
Intended Use vs. Indications for Use
510(k) submissions define two related but distinct concepts: Intended Use and Indications for Use. Intended Use describes the general purpose of the device — "computer-aided detection software intended to identify and flag potential findings on medical images for radiologist review." Indications for Use is narrower and specifies the clinical situation — patient population, imaging modality, anatomical region, clinical context.
Buyers should ask to see the cleared Indications for Use statement, not just the cleared Intended Use. Vendors sometimes emphasize the broader Intended Use language in marketing materials, which can create the impression that clearance is broader than it actually is. The FDA's 510(k) database (accessible at accessdata.fda.gov) allows you to look up the cleared summary for any cleared device, including the Indications for Use language, and this is a reasonable step for any department conducting formal due diligence on a tool they're considering deploying clinically.
The Locked Algorithm vs. Adaptive Algorithm Problem
Traditional 510(k) clearance is predicated on a locked algorithm — the device that gets cleared is the device that gets deployed, and any subsequent change to the model that could affect performance requires either a new 510(k) submission or, under more recent FDA guidance, a predetermined change control plan (PCCP) that was reviewed as part of the original clearance.
This creates a real tension with how modern machine learning-based radiology tools are developed. A model that continues to learn or update based on new data after deployment is not a locked algorithm, and traditional 510(k) clearance doesn't cover ongoing model updates without additional regulatory work. The FDA's 2021 action plan for AI/ML-based Software as a Medical Device (SaMD) and subsequent guidance on PCCPs are an attempt to create a regulatory framework for adaptive algorithms, but this is still evolving and most currently cleared AI radiology tools either use frozen model versions or have limited update pathways.
From a buyer's perspective, this matters because a vendor's statement that they "regularly update and improve the model" may be technically in tension with their cleared device specification. Asking directly — "is the version of the model we deploy the cleared version, and how do model updates relate to your clearance?" — is a reasonable due diligence question. A serious vendor will have a clear answer.
Clinical Decision Support vs. Medical Device Software
Not all AI radiology software requires 510(k) clearance. The FDA's 21st Century Cures Act and subsequent guidance created an exemption for certain clinical decision support (CDS) software that meets specific criteria: it must not acquire, process, or analyze a medical image, and the clinician must be able to independently review the basis for the recommendation and is not expected to rely primarily on the software output.
Some radiology AI tools — particularly those focused on workflow prioritization, report drafting assistance, or structured reporting templates — may fall under the CDS exemption and would not require 510(k) clearance. The key functional distinction is whether the software is directly analyzing images (which generally requires clearance as a SaMD) or operating on derived data such as structured report fields or DICOM metadata (which may fall under CDS).
This is a meaningful distinction for buyers, because a tool that hasn't pursued FDA clearance isn't necessarily non-compliant — it may legitimately fall within the CDS exemption. The question to ask is: has the vendor done a formal regulatory assessment of their device classification, and what is their documented rationale for the pathway they've chosen? A vendor who hasn't done this analysis is a different risk profile from a vendor who has a clear regulatory strategy that places them in the CDS exemption with documented reasoning.
What Clearance Doesn't Tell You About Performance
510(k) clearance is a regulatory determination about substantial equivalence to a predicate device. It is not a determination that the device is clinically superior to alternatives, that it will perform well on your patient population, or that it will integrate smoothly into your workflow. Cleared status is a floor for regulatory compliance, not a ceiling for what you should evaluate before deployment.
A device can be cleared and perform poorly on your specific case mix. The validation dataset used in the 510(k) submission reflects the manufacturer's test cohort — which may differ from your patient demographics, your CT scanner models, your reconstruction protocols, and your clinical indication mix in ways that affect real-world performance.
We're not saying cleared status is irrelevant — it's a meaningful threshold and it signals that the manufacturer has engaged with the regulatory process seriously enough to document their device's performance. But it's one input into a procurement decision, not a substitute for the department's own evaluation of whether the tool performs as claimed in their specific environment.
Questions to Ask Every Vendor
When evaluating any AI radiology tool, these questions should be part of a standard due diligence checklist:
- Is this device 510(k) cleared? If yes, what is the cleared Indications for Use? Can you provide the cleared summary from the FDA 510(k) database?
- If not cleared, what is your regulatory classification rationale? Has this been formally documented and reviewed by regulatory counsel?
- Is the deployed algorithm a locked version? What is the relationship between model updates and your cleared device specification?
- If you have a PCCP, what kinds of model changes does it permit without a new submission?
- What was the composition of the validation dataset used in your 510(k) submission — scanner models, slice thickness, patient demographics, disease prevalence?
- What post-market surveillance data do you collect, and is any of it publicly available?
The answers to these questions won't be uniformly favorable for any single vendor, and a vendor who can't answer them clearly is telling you something important about their regulatory maturity. The FDA's evolving guidance on AI/ML-based SaMD is genuinely complex and the landscape will continue to shift over the next several years. Departments that understand the current framework are better positioned to make procurement decisions that hold up as that framework evolves.