Engineering validation does not by itself get an imaging-AI device cleared. The FDA does that, through a premarket pathway set by the device’s intended use and risk class. Under FDA medical device regulation, software that analyses a medical image for a diagnostic purpose stays a device even where other clinical decision support is excluded. Most AI devices reach the US market through a 510(k). A retrained model is not automatically covered by the existing clearance: it needs a documented risk assessment, and it can trigger a new 510(k) unless an authorised Predetermined Change Control Plan (PCCP) already covers the change. The evidence an imaging AI validation engagement produces goes into that submission. The submission itself, the pathway and the decision belong to the manufacturer and the agency.
How does FDA medical device regulation work for an imaging AI?
The FDA does not regulate AI as such. It regulates medical devices, and AI-enabled devices are among them. Section 201(h) of the FD&C Act defines a device as an instrument, apparatus or similar article intended for use in diagnosing, curing, mitigating, treating or preventing disease, or intended to affect the structure or function of the body. The definition also says the term does not include software functions excluded under section 520(o). The FDA reviews devices using a risk-based approach that looks at intended use and technological characteristics.
Intended use drives classification, pathway and evidence scope. Under 21 CFR 801.4, intended use can be shown by labelling claims, advertising, or oral or written statements by the manufacturer. The IMDRF definition likewise treats sales and marketing materials as information reflecting the manufacturer’s objective intent. In practice, a slide deck promising “detects lung cancer” can turn a research tool into a device claim.
The FDA also regulates by function, not by platform. Software performing a device function is a device whether it runs on a scanner workstation, a desktop or a phone. Software as a Medical Device (SaMD), in the IMDRF’s definition, is software intended for one or more medical purposes that performs them without being part of a hardware medical device. A model served from a cloud GPU that reads CT studies is SaMD. Firmware inside the scanner is not.
Why image analysis stays a device
Section 520(o)(1)(E) excludes some clinical decision support from the device definition, but only when all four of its criteria are met. The full criteria, and how IMDRF risk categories map onto FDA classes, are covered in our piece on intended use and risk classification. For imaging teams the first criterion settles the question: software intended to acquire, process or analyse a medical image for a 201(h) purpose remains a device. The FDA’s CDS guidance names enhancement, measurement, identifying normal and abnormal structures, sizing a suspected nodule, and CADe and CADx functions as image analysis that fails that criterion. It also lists software that detects and highlights abnormalities and grades their severity as a device function. Medical images include CT, X-ray, ultrasound and MRI, as well as images acquired for a medical purpose such as pathology or dermatology. The mechanics of those detection and diagnosis systems are covered in how computer-aided diagnosis software works.
That does not mean everything a life-sciences company builds near imaging is regulated. Many parts of a company’s pipeline sit outside the regulated portion even when the company assumes otherwise, so the first check is whether the specific application falls under the requirements at all. Where a team is genuinely unsure, the 513(g) process gives written FDA feedback on classification. The FDA generally responds within 60 days with a confidential letter.
510(k), De Novo or PMA: how Class I, II and III decide the pathway
Device class sets the pathway. Regulatory controls rise from Class I (low risk) to Class III (high risk). Class I devices carry general controls, including registration and listing, the quality system regulation, labelling and medical device reporting. Class II adds special controls and, for most devices, premarket notification. Class III requires premarket approval under 21 CFR Part 814.
A 510(k) shows that a device is substantially equivalent to a legally marketed predicate. It must have the same intended use, and either the same technological characteristics or different ones that raise no different questions of safety and effectiveness. Manufacturers must name a primary predicate that is most similar in indications and technology, and the FDA treats split predicates as inconsistent with the 510(k) standard. A predicate can be a device cleared through 510(k), a reclassified device, a preamendments device, or a device granted through De Novo.
De Novo classifies a novel device into Class I or II when general controls, or general and special controls, are enough and no legally marketed predicate exists. Congress created the pathway in 1997, and a 2012 amendment allowed sponsors to request it directly without first receiving a “not substantially equivalent” (NSE) decision on a 510(k). If the FDA finds a likely predicate during review, or finds that the device falls under a Class III regulation, the request is not eligible. A granted De Novo creates a classification regulation and lets the device serve as a predicate for later 510(k)s.
PMA applies to Class III devices. According to a MedDeviceGuide SaMD guide, PMA is relatively uncommon for software and applies to SaMD that drives life-critical decisions autonomously, such as setting radiation therapy dosing without physician override.
Pathway decision table
| Your situation | Likely route | What it hinges on |
|---|---|---|
| Class II imaging function, a predicate with the same intended use exists | 510(k) | Substantial equivalence to one primary predicate |
| Low-to-moderate risk, no predicate exists | De Novo (direct) | No likely predicate and no Class III regulation covering the device |
| 510(k) returned NSE for lack of predicate, new intended use or different technology | De Novo | NSE solely for inadequate performance data is generally ineligible |
| High-risk, Class III function | PMA | Premarket approval under Part 814 |
| Not sure the software is a device at all | 513(g) request | Written FDA response, generally within 60 days |
Timing differs too. A 2024 analysis in npj Digital Medicine reports mean review times of 150 days for 510(k)s and 399 for PMAs, with De Novo decisions taking on average 2.3 times as long as 510(k)s. The De Novo statute requires a classification order within 120 days of receipt.
How does FDA’s AI-enabled device guidance apply to an imaging model, and what does the AI device list tell you about predicates?
The FDA’s list of AI-enabled medical devices names authorised devices with their product codes and pathways, and the agency presents it as a way to understand the device landscape and regulatory expectations. Exponent reported that the list held 1,016 authorised AI/ML-enabled devices as of December 2024. The list is not comprehensive, though. It identifies devices mainly by AI-related terms in their authorisation summaries or classification. Treat it as a starting index, then confirm in the 510(k) and De Novo databases.
The MedDeviceGuide guide advises starting pathway analysis from that list’s product codes. Image-processing software for diagnostic review can fall under product code LLZ (21 CFR 892.2050), and radiology software such as Fujifilm’s Synapse 3D appears on the AI list under that code. If a device in your clinical area was authorised through De Novo and created a product code, a 510(k) referencing that De Novo as predicate may be open to you. The MedDeviceGuide guide gives precedents. A 2018 De Novo (DEN170073) for large-vessel-occlusion stroke triage created product code QAS, which later served as a predicate for competing triage clearances. Aidoc’s foundation-model CARE1 was cleared through 510(k) in February 2025 by referencing its own earlier triage clearances.
The same mechanism is visible outside imaging. Samsung’s Sleep Apnea Feature was granted De Novo as DEN230041, creating Class II product code QZW. Apple’s Sleep Apnea Notification Feature was then cleared through 510(k) with DEN230041 as predicate, although it uses accelerometer rather than blood-oxygen data. The FDA found that those technological differences raised no new questions of safety or effectiveness.
One caution on predicates. The npj Digital Medicine authors argue that 510(k) review does not take the nature of the algorithm into account, so a novel AI device can be found equivalent to a legacy predicate built on older technology. Those authors present this as a weakness of the system. A device without the predicate’s intended use is not substantially equivalent.
Is engineering validation enough to get an imaging-AI device cleared?
No, not on its own. Engineering validation is necessary, and it does not decide clearance. The IMDRF frames SaMD clinical evaluation as three kinds of evidence. Analytical validity means the output is accurate for a given input. Scientific validity means the output is associated with the clinical condition. Clinical performance means the output is clinically meaningful and associated with the target use. Analytical validity is generated during verification and validation inside the quality system and is always expected. For diagnostic SaMD whose risk profile requires it, the IMDRF adds clinical performance evaluation.
Engineering verification and validation testing generates analytical validity evidence. A submission typically also needs a clinical performance study. A statistical review of AI diagnostic validation describes the expected external study. It is run on the locked, market-ready version of the device, uses subjects entirely separate from training, follows a pre-specified protocol and analysis plan, and uses a prospective design where feasible. For aids to clinicians, such as radiology tools, a reader study comparing clinicians with and without the software often replaces a standalone performance study.
Why a model can pass every engineering test and still fail review
Passing every engineering test does not settle review, because the FDA first compares intended use with the predicate. Does the device match its predicate’s intended use? Does the evidence cover the intended population, including every applicable subgroup? Is the reference standard justified? The special controls for DEN230041 show the bar concretely. They require a study population representative of intended use, performance across subgroups such as age, sex, BMI and skin tone, comparison with a justified clinical reference such as polysomnography, and validation on data separate from training. In a 510(k), software validation data is one input among several that the FDA weighs when deciding whether the device is as safe and effective as the predicate.
An intended use worded too broadly invites a higher classification. Excluding low-quality images can improve sensitivity and specificity yet bias results, since poor images sometimes overlap with genuinely hard cases.
Does an existing FDA clearance cover a retrained model?
Only if the change stays below the regulatory threshold or falls inside an authorised PCCP. Under 21 CFR 807.81(a)(3), a new 510(k) is required when a change could significantly affect safety or effectiveness, or when it is a major change in intended use. The significant effect can be positive or negative, so “the new model is better” is not an exemption. The FDA’s AI/ML discussion paper expects that many AI modifications, including retraining on new data and architecture changes, would be subject to premarket review under the software-modifications guidance.
The decision belongs to the 510(k) holder, and the FDA recommends recording the justification for submitting or not submitting in the change control records. The FDA’s software-change guidance asks for a risk-based assessment first. Verification and validation then confirm a “no submission” decision. They cannot overturn a “could significantly affect” finding, even when every test passes. A clean regression run does not replace the risk assessment.
Which changes to a cleared imaging model need re-evaluation
Use this as a diagnostic checklist against the FDA’s software-change criteria:
- New or modified hazard. The change creates or modifies a hazard whose harm is serious or worse, and it is not already mitigated in the cleared device. When all three hold, a new 510(k) is likely.
- Risk control change. New or modified risk controls are needed to prevent significant harm.
- Clinical performance shift. The change could significantly affect clinical functionality or performance specifications tied to the intended use.
- Algorithm rewrite. A complete rewrite of the core algorithm can require a new 510(k) even with the same claims and risk profile.
- Intent to improve outcomes. A change made to significantly improve clinical outcomes or mitigate a known risk likely requires one.
- New inputs or new use. Supporting new scanner vendors or new input types, widening the patient population, or moving from “aid in diagnosis” to “definitive diagnosis” are input or intended-use modifications.
- Usually documented to file. Cybersecurity-only fixes, changes that restore the most recently cleared specification, and cosmetic changes typically do not need a submission.
Ignoring this has consequences. RegulatoryIQ reports a September 2026 FDA warning letter to O.N. Diagnostics. The letter found that a new machine-learning algorithm for vertebral landmarking, and a move from desktop to web architecture, each required a new premarket notification. No PCCP was involved.
How do adaptive models work under a PCCP, and what does that require of the eval harness?
A PCCP lets a manufacturer get specific planned changes authorised inside the original submission, then make them later without a new marketing submission for each one. Section 515C of the FD&C Act, enacted on 29 December 2022, provides the authority. A PCCP can be authorised through 510(k), De Novo or PMA. The FDA’s general PCCP draft guidance sets out three components: a Description of Modifications, a Modification Protocol, and an Impact Assessment. Exponent reports that the final AI-specific guidance keeps the same three-part structure for AI-enabled device software functions. The earlier discussion paper used different terms for the same idea: SaMD Pre-Specifications for what may change, and an Algorithm Change Protocol for how.
TechnoLynx builds the evaluation and lineage evidence such a plan relies on, as an engineering output an auditor can read and sign against.
What cannot go in a PCCP matters as much. The joint FDA, Health Canada and MHRA principles limit PCCPs to changes within the device’s original intended use. The FDA’s draft guidance lists new patient populations and changes needing new clinical data among modifications generally not appropriate. It also excludes changes that could introduce a new risk not acceptable before mitigation. Its sleep-apnoea worked example names an alternate input data type and a significant architecture change as out of scope. A PCCP is not a licence for a model to drift into a new clinical claim.
What cleared plans actually allow
So far the plans the FDA has cleared mostly describe locked models updated on a schedule, not live learning in the field. Apple’s sleep-apnoea PCCP states that it includes no adaptive algorithms that continuously learn in the field, and that every modification is trained, tuned and locked before release. RegulatoryIQ’s review of the FDA’s PCCP list found the same pattern elsewhere: retraining a locked model, validating against fixed acceptance criteria and releasing uniformly to every installed device. It also reports one exception, VIRTUE QA (K260597), which retrains six machine-specific models daily on each machine’s own data, with gated activation that keeps the prior validated model if a candidate errors. By RegulatoryIQ’s count, the share of AI-enabled authorisations carrying a PCCP rose from 1.3% in 2023 to 11.6% in the first half of 2026.
One predicate rule affects competitors. Only the version cleared before any PCCP changes may be used as a predicate.
What the eval harness has to do
Under a PCCP, the Modification Protocol the eval harness carries out is part of the marketing authorisation, because the manufacturer must implement modifications the way the authorised plan says. The harness therefore needs:
- A sequestered test set. Access controls and an audit trail documenting test-set independence as new data is added.
- Pre-defined acceptance criteria for each modification, fixed before the change is built.
- Leakage controls. Measures limiting information leakage when part of the test set is reused across evaluations.
- Per-change and aggregate verification. Validation of the whole device after each modification and of all modifications together.
- Fail-closed behaviour. An unresolvable acceptance failure is recorded and the change is not shipped.
- Revert and stop mechanisms. A way to roll back a change that misses its performance criteria.
- Version tracking and user communication for every release.
- Labelling discipline. Labelling reflects only modifications already implemented.
A modification made outside the protocol’s methods likely needs a new marketing submission before release. The FDA encourages a Q-Submission to get feedback on a proposed PCCP before filing.
Where 21 CFR Part 820 and Part 11 meet the validation evidence
Since 2 February 2026, Part 820 has been the Quality Management System Regulation (QMSR). It incorporates ISO 13485:2016 by reference and now explicitly requires risk management. Part 820 was amended, not retired. The QMSR applies to finished device manufacturers who intend to distribute devices commercially.
Three things connect it to an engineering team’s work. First, the FDA treats every software modification, from a hotfix to a retrain, as a design change under Part 820. Second, device changes must be documented within the quality system, and the FDA recommends recording the “submit or not” justification in change control records. Third, the FDA expects device software to be verified and validated together with its computing platform. A model validated on one inference stack and deployed on another has a gap in that chain. Note that those FDA documents cite the earlier quality system regulation’s section numbers, such as 21 CFR 820.30. Check citations against the current Part 820 text before reusing them.
Part 11 is narrower than broad interpretations suggest. The FDA’s Part 11 guidance says the agency interprets the scope narrowly: Part 11 applies when records required by predicate rules, or submitted to the FDA, are kept electronically in place of paper. The FDA exercises enforcement discretion on specific Part 11 validation and audit-trail requirements, while predicate-rule validation obligations still apply. So eval logs, dataset manifests and signed reports that serve as required design records need controlled, attributable storage. Not every training-job log does. Our guide to computer system validation in pharma covers the system-level side.
What can TechnoLynx do here, and where does engineering validation stop?
TechnoLynx builds the validation evidence an imaging-AI submission draws on, as an engineering output that a reviewer, auditor or compliance owner can read and sign against. That covers task-specific evaluations, lineage and traceability for production decisions, and evidence packs aligned to named published rubrics. Our life sciences AI engineering page and the parent page on the imaging AI validation engagement describe the engagement itself.
The boundary is explicit. TechnoLynx does not provide policy advice or certification. We produce the artefacts that an audit and a submission need. Under FDA rules, the decision whether a change needs a new 510(k) rests with the 510(k) holder. A team with an in-house RA lead and a mature ML platform may not need outside engineering help. Where the engineering evidence is the weak link, TechnoLynx builds it. The distinction is set out further in validation evidence versus regulatory clearance.
FAQ
When did the FDA start regulating medical devices?
The 1976 Medical Device Amendments to the FD&C Act first gave the FDA authority to regulate medical devices.
Is 21 CFR 820 obsolete?
No. Part 820 was amended into the Quality Management System Regulation, effective 2 February 2026, which incorporates ISO 13485:2016 by reference.
What does 21 CFR 801.4 mean?
21 CFR 801.4 says how intended use is established: through labelling claims, advertising materials, or oral or written statements by the manufacturer or its representatives.
What does Class I mean for a medical device?
Class I devices are subject to general controls, such as establishment registration, device listing, quality system requirements, labelling and medical device reporting.