FDA 510(k) for Radiology AI: Reading the Cleared Device List and What It Means for Procurement

The FDA's database of cleared medical devices lists over 700 AI/ML-enabled medical devices as of late 2024, with radiology representing the single largest category — over 60% of cleared AI/ML devices are imaging-related. This number is frequently cited by radiology AI vendors as a market signal of regulatory maturity. It is a real number, from a real database, and it does reflect genuine regulatory activity. But what "cleared" means, how to read the cleared device list, and what it actually tells you about a product's fitness for your environment are questions that require more careful reading than the headline suggests.

The 510(k) Process: What It Does and Doesn't Establish

FDA 510(k) clearance under 21 CFR Part 807 establishes that a new device is substantially equivalent to a legally marketed predicate device in terms of intended use and technological characteristics. For radiology AI, the 510(k) review examines: the intended use description (exactly what the device is designed to do, in what clinical context, with what patient population), the technological characteristics versus the predicate (algorithm type, training data scope, performance validation methodology), and any special controls required for the device type.

510(k) clearance establishes substantial equivalence — it does not establish that the device is optimal, that its published performance is reproducible in your specific environment, or that it is clinically superior to alternatives. Two 510(k)-cleared ICH detection systems can have substantially different real-world performance profiles depending on their training data, threshold configurations, and the populations they were validated on. Clearance tells you that FDA reviewed the submission and found it substantially equivalent to a predicate. It does not tell you which system will perform better in your overnight CT head volume.

The clearance also applies specifically to the intended use as described in the 510(k) submission. If a cleared ICH detection system is used in a way that is outside its cleared intended use — for example, if a hospital uses it on patient populations substantially different from those described in the submission — that's a regulatory use outside clearance, even if the vendor supports the configuration technically.

Reading the FDA AI/ML Device Database

The FDA maintains a public list of cleared AI/ML-enabled devices at its website, updated periodically. The list includes: device name, vendor name, 510(k) number, date of clearance, device class, and a brief description of intended use. Key fields to look at when reviewing a clearance for radiology AI procurement:

• Intended Use: This is the most important field. It defines exactly what clinical task the device is cleared for, in what imaging modality, and for what patient population. A device cleared for "detection of intracranial hemorrhage on non-contrast head CT in adult patients" has a narrow scope — it is not cleared for pediatric patients, for contrast-enhanced head CT, or for hemorrhage quantification if quantification wasn't in the cleared intended use.
• K-number: The unique identifier for the 510(k) submission, used to look up the full submission summary in the FDA's CDRH 510(k) database. The submission summary contains the detailed description of testing and validation performed.
• Predicate Device: The device or devices cited as predicates for substantial equivalence. Understanding the predicate chain tells you how the regulatory review was structured and what performance standards were applied.
• Special Controls: Class II medical devices (which most radiology AI systems are) may have specific performance testing requirements documented in special controls guidance documents.

Aidoc's ICH detection system (K183327) and Viz.ai's LVO detection system (K193482) are examples of cleared radiology AI devices with publicly available submission summaries. RapidAI's ICH system (K203804) is another. Reading the full submission summary for a competitor's cleared device is instructive for understanding the standard performance validation methodology that FDA expects — which then helps you evaluate whether a vendor's published claims are consistent with what a 510(k) submission would actually require.

The SaMD Framework and Classification

Radiology AI detection systems are Software as a Medical Device (SaMD) under the International Medical Device Regulators Forum (IMDRF) framework that FDA has adopted. The risk classification for SaMD considers: (1) the significance of the information provided to the clinical decision — does it drive the diagnosis directly, or is it a triage aid where a human makes the final decision? and (2) the state of the healthcare situation — life-threatening, serious, or non-serious.

Most current radiology AI triage systems are designed to assist radiologist workflow (triage prioritization, worklist management) rather than to make independent diagnoses. This "non-autonomous" or "assistant" function tends to place them in lower risk SaMD categories than fully autonomous diagnostic systems. The clinical validation requirements under the AI/ML Action Plan (FDA, 2021) and the Predetermined Change Control Plan (PCCP) pathway reflect this stratification — systems that assist but don't replace human judgment face different validation requirements than systems that produce final diagnoses.

FDA's AI/ML Action Plan, published in January 2021, outlined a framework for ongoing regulation of AI/ML SaMD — specifically addressing the challenge that AI models can change over time through retraining, which traditional 510(k) clearance (a one-time point-in-time review) doesn't naturally accommodate. The PCCP pathway, formalized through FDA guidance in 2023, allows AI manufacturers to pre-specify planned modifications to their algorithms and have those modification plans reviewed as part of the initial clearance — reducing the need for supplement submissions for each incremental model update.

What "510(k) Submission in Progress" Actually Means

Pacslens's regulatory status is straightforward: we are preparing 510(k) submissions for selected indications under 21 CFR Part 892 (radiology devices). We reference cleared predicate devices including Aidoc (K183327), Viz.ai (K193482), and RapidAI (K203804) in our intended use mapping and predicate device analysis. We have not yet received clearance, and we do not represent Pacslens as a cleared device in any clinical context that would require cleared device status.

This is the honest position for an early-stage radiology AI vendor at the current stage of development. Pursuing clearance before deploying commercially — even in pilot contexts — is the correct approach, and we intend to complete clearance submissions before moving to broader commercial deployment.

For community hospitals evaluating early-stage radiology AI vendors for pilot programs, the relevant questions about regulatory status are: Has the vendor engaged FDA on their 510(k) pathway (Pre-Sub meeting, written feedback)? What is the intended use description for the clearance submission? What predicate devices are being cited? Is the vendor tracking their deployment status in a way consistent with FDA's pre-clearance use expectations?

GMLP and the Quality System Connection

Good Machine Learning Practice (GMLP) is the set of principles that FDA, Health Canada, and MHRA (UK) have jointly developed for AI/ML medical device development. GMLP covers: data management practices for training and testing data, model performance testing on representative populations, transparency of algorithm outputs to users, monitoring for real-world performance drift, and human-AI interaction design that maintains appropriate human oversight.

GMLP is not a certification standard with a formal audit program — it is a guidance framework. But vendors whose development practices align with GMLP are better positioned for a successful 510(k) submission because the GMLP principles map directly to the types of documentation FDA reviewers look for in AI/ML device submissions. When evaluating vendor technical credibility, asking about GMLP alignment is a proxy for asking about the quality of their development and validation practices.

IEC 62304 (Medical device software lifecycle processes) is the complementary standard that governs the software engineering practices for medical device software — version control, change management, testing, and defect tracking. An AI system that tracks software changes under an IEC 62304-aligned development process has a documented trail of what changed, when, and why, which is essential for post-market surveillance and for managing the regulatory implications of model updates under a PCCP or supplement submission.