The least glamorous part of AI in radiology may turn out to be the most important: telling people what changed.
That sounds simple. It is not. A diagnostic AI tool may be trained on one dataset, validated on another, deployed inside a PACS or reporting workflow, monitored after release, and then updated when the model, threshold, input, interface, or intended environment changes. Somewhere in that chain, a radiologist is expected to decide whether to trust a box, a score, a contour, a triage flag, or a sentence.
The FDA has been building a framework for this problem in pieces. Three recent documents are worth reading together: the final guidance on Predetermined Change Control Plans, the draft guidance on AI-enabled device software function lifecycle management and marketing submissions, and the joint FDA/Health Canada/MHRA transparency principles for machine learning-enabled medical devices.
Read separately, they can look like regulatory paperwork. Read together, they point to a more practical obligation for diagnostic radiology: if software is going to change over time, the clinical team needs to know what changed, why it changed, how it was tested, and what that means at the point of care.
The PCCP is about planned change
The FDA’s final PCCP guidance, issued in August 2025 after an earlier December 2024 version, is meant for AI-enabled device software functions. The basic idea is appealing: if a manufacturer can describe a set of planned modifications up front, along with the method for making and validating those modifications, then some future changes may be implemented under that authorized plan rather than requiring a new marketing submission each time.
The guidance describes three core PCCP components: a Description of Modifications, a Modification Protocol, and an Impact Assessment. In plain language: what might change, how the change will be made and tested, and why the benefits and risks remain acceptable.
For radiology AI, this matters because many useful models are not static in the way we wish they were. Scanner protocols change. Patient populations shift. Reconstruction methods evolve. A model that worked reasonably well on one local dataset may not behave the same way after a subtle workflow or input change. A PCCP does not make those problems disappear. It creates a structured lane for certain anticipated changes.
The lifecycle guidance is about the whole device
The draft lifecycle guidance, issued in January 2025 and explicitly labeled “not for implementation,” is broader. It gives FDA’s current thinking on what should be included in marketing submissions for devices that include AI-enabled device software functions.
What stands out is the total product lifecycle framing. The draft guidance does not treat an AI model as a math object floating above clinical reality. It asks for information about the device description, user interface, labeling, risk management, data management, model development, validation, performance monitoring, cybersecurity, and other lifecycle considerations.
In radiology, an AI output is rarely just an output. It appears somewhere. It interrupts or does not interrupt someone. It changes a worklist, a report, a measurement, or a follow-up recommendation. It may be visible to the radiologist, technologist, referring clinician, patient, or no one at all. The FDA argues the same model can have different safety implications depending on how it is embedded.
The lifecycle view forces the uncomfortable but necessary question: how does the technical change become a clinical change?
Transparency ties the pieces together
The transparency principles make the practical obligation more explicit. The FDA page defines transparency as the degree to which appropriate information about a machine learning-enabled medical device, including intended use, development, performance, and when available, logic, is clearly communicated to relevant audiences. It also emphasizes who needs the information, why they need it, what information is relevant, where it should appear, when it should be communicated, and how communication should be designed.
This is where the regulatory documents become clinically interesting. A PCCP may authorize a path for future modification. Lifecycle documentation may describe how the system is developed, validated, monitored, and maintained. But transparency is what connects those facts to the people using the tool.
For a radiologist, transparency does not necessarily mean opening the neural network and pretending that every parameter is interpretable. Sometimes that is not possible, and most of the time it is not useful in daily practice. The more practical version is knowing enough to use the device appropriately: intended use, intended users, patient population, inputs, outputs, known limitations, performance characteristics, subgroup concerns, failure modes, monitoring plan, and whether a displayed result is meant to inform or replace a human judgment.
The hardest part is timing. A 40-page Model Card is not transparency (it isbalso not a “card”). A model card is useful only if it answers the question being asked by the person asking it. A release note is helpful only if the local team understands whether the change affects their scanners, protocols, population, or workflow.
What this means for radiology practices
Radiology practices evaluating AI-enabled SaMD should probably stop asking only, “Is this FDA-cleared?” That question still matters, but it is incomplete.
A better set of questions might be:
- Does this device include a PCCP, and what types of changes are covered?
- How will we be notified when the model, threshold, data inputs, or user interface changes?
- What local validation or acceptance testing is expected before or after an update?
- What performance monitoring is available after deployment?
- What limitations, subgroup performance concerns, or known failure modes should radiologists actually see?
- Where will this information appear in the workflow: labeling, documentation, dashboard, PACS overlay, alert, release note, or governance review?
These questions are not just for vendors. They are also for health systems.
There is a temptation to treat regulatory clearance as the end of the story, for AI tools that even qualify as regulated devices.
AI-enabled medical devices will need more than good retrospective performance. They will need credible plans for change, lifecycle management, and transparency. In diagnostic radiology, the last of these may be the bridge between a technically authorized device and a clinically trustworthy one.