American Association of Physicists in Medicine

The American Association of Physicists in Medicine (AAPM), the professional organization of U.S. medical physicists, appreciates the opportunity to comment on FDA’s discussion paper, Considerations for the Regulation of Generative AI-Enabled Medical Devices (Docket FDA-2026-N-7874). Medical physicists are clinical experts responsible for acceptance testing, commissioning, independent verification, and quality management of devices in radiation therapy, medical imaging, and nuclear medicine.

AAPM’s principal recommendation is that FDA establish an explicit evidence chain linking manufacturer validation, clinical confirmation, site-specific acceptance and commissioning, post market surveillance, and controlled modification, with medical physicists providing site-specific technical leadership within multidisciplinary clinical teams. Benchmark success and nominal human review should not be treated as sufficient evidence of safe deployment. AAPM’s complete statement, with references, tables, and all 26 discussion questions, is attached; the points below are a summary.

1. Operationalize the risk framework. AAPM supports the activity-consequence framework as a heuristic, not a self-sufficient risk score. Sponsors should maintain a structured risk record covering credible errors, severity, detectability, propagation, reversibility, and cumulative exposure. Informational labels should not automatically imply low risk, since images and quantitative outputs can direct clinical action. Human supervision should be a demonstrated safeguard: reviewers need competence, independent information, adequate time, and authority to intervene.

2. Require physics-based, task-specific competency evidence. Benchmarks should follow the device’s hazards and claims. For reconstruction and synthesis, evaluate clinically relevant structure additions, deletion, displacement, distortion, quantitative bias, and performance across acquisition conditions. Visual realism alone is insufficient. For synthetic CT, connect image errors to target coverage, organ-at-risk dose, and proton range. Verification must use independent information; two engines sharing erroneous inputs can agree. Stochastic behavior should be tested under the deployed inference policy with repeated runs and prespecified perturbations.

3. Establish benchmark validity and sound statistics. Prespecify claims, target populations, reference standards, and acceptance criteria. Require external and temporal validation, subgroup analyses, data-leakage controls, and independent review of sponsor-developed benchmarks. Synthetic data should not substitute for validation using representative pathology or populations. Report confidence intervals and high-severity failures, not averages alone.

4. Preserve clinical confirmation and test the intended human-AI workflow. Clinical confirmation methods should be risk-proportionate and resolve the uncertainty that bench testing cannot. The integrity of human-AI workflow should be evaluated prospectively to assure unexpected model performance can be controlled by human oversight. User manuals should define error messages per IEC 63183:2019.

5. Make site-specific acceptance testing and commissioning an explicit deployment bridge. A designated medical physicist should lead risk-proportionate acceptance testing and commissioning, supported by a manufacturer deployment package. Sites should receive clear model identification and guidance on the datasets required for safe clinical deployment, with manufacturer on-site and remote technical support during acceptance and commissioning.

6. Connect site-specific and manufacturer surveillance. Site-specific technical monitoring, carried out by an expert team led by a medical physicist, is an operational foundation of post market assurance, complementary to manufacturer-led fleet surveillance and corrective action. Monitoring frequency should be guided by failure mode and effects analysis, with prespecified graded action thresholds and rollback.

7. Control changes across the complete configuration. Treat the device as model weights plus prompts, retrieval, preprocessing, decoding, tools, and interfaces, with versioned inventories and impact analysis. Predetermined Change Control Plans (PCCPs) should be bounded, with prohibited expansions and institutional release criteria. Foundation-model Device Master Files should be voluntary, version-specific, and non-authorizing, with contractual change notice and protection against silent substitution.

8. Evaluate agentic systems as bounded action sequences, with a validated boundary between proposing and committing actions, externally enforced constraints, fault and recovery testing, and full auditability. Existing interlocks and independent verification must not be bypassed by conversational interfaces, and mechanisms must confirm that action sequences do not cross boundaries from one patient to another mid-stream.

AAPM welcomes further discussion.

Physics-informed lifecycle quality assurance for generative AIenabled medical devices in radiation therapy, medical imaging and
nuclear medicine: American Association of Physicists in Medicine
(AAPM) response to FDA-2026-N-7874

Abstract

Generative artificial intelligence (GenAI) is transforming medical devices across radiation
therapy, medical imaging, and nuclear medicine, but its safe clinical implementation requires a
robust connection between premarket evidence and real-world clinical performance. Medical
physicists, including those certified by the American Board of Radiology (ABR), are clinical
experts who offer experimental and analytic approaches to site-specific acceptance, independent
verification, system commissioning, and quality management of medical devices in the areas of
radiation therapy, medical imaging, and nuclear medicine. The American Association of
Physicists in Medicine (AAPM) prepared this response to develop physics-informed
recommendations for the U.S. Food and Drug Administration (FDA) discussion paper on GenAIenabled medical devices, emphasizing medical physicists’ crucial role in premarket
benchmarking, post market monitoring, and clinical risk management frameworks. A targeted
analysis examined the FDA paper and its 26 questions, related agency documents, AAPM
guidelines, and selected reviews and primary studies published in leading scientific journals.
Recommendations were organized by failure mechanism, required evidence, stakeholder
responsibility, and lifecycle stage. This is a policy analysis, not a systematic review or completed
consensus process. The proposed response extends the FDA’s activity-consequence paradigm by
explicitly incorporating principles of measurement integrity, detectability, propagation,
intervention time, and effective supervision. While benchmark success supplied by the
manufacturer is important evidence, it is not sufficient for assuring safe and effective clinical
deployment. Medical physicists should conduct site-specific clinical confirmation, including
risk-proportionate acceptance testing, commissioning and ongoing quality assurance (QA).
Proposed clinical confirmation should include information preservation, quantitative accuracy,
dosimetric consequences, model uncertainty, and action sequences. Shared monitoring combines
physicist-led site-specific technical surveillance with manufacturer-led fleet surveillance without
transferring manufacturer accountability to clinical user. Whole-system change control,
foundation-model evidence, and independent constraints on agentic actions should preserve that
assurance as the medical devices evolve. In conclusion, AAPM recommends the FDA establish a
regulatory framework that explicitly connects product-level evidence with adequately resourced
site-specific quality management and lifecycle performance monitoring. For medical devices
within the domain of medical physics practice (i.e., radiation therapy, medical imaging, and
nuclear medicine), site-specific acceptance testing, commissioning, and ongoing quality
management are essential requirements for safe clinical deployment, which should be carried out
by medical physicists in collaboration with the multidisciplinary clinical team.

1. Background

Medical Physics is a branch of physics associated with the practice of medicine, including
diagnostic medical physics, therapeutic medical physics, nuclear medical physics, and medical
health physics. The essential clinical responsibility of the Qualified Medical Physicist (QMP,
who is certified by an appropriate board) is to ensure the safe and effective delivery of radiation
to achieve a diagnostic or therapeutic result as prescribed in patient care. The medical physicist
performs or supervises the technical aspects of procedures necessary to achieve this objective.
The responsibilities of the medical physicist include: protection of the patient and others from
potentially harmful or excessive radiation; establishment of adequate protocols to ensure
accurate patient dosimetry; the measurement and characterization of radiation; the determination
of delivered dose; advancement of procedures necessary to ensure image quality; development
and direction of quality assurance programs; and assistance to other health care professionals in
optimizing the balance between the beneficial and deleterious effects of radiation; and
compliance with applicable federal and state regulations. The American Association of
Physicists in Medicine (AAPM), established in 1958, is the scientific, educational, and
professional organization of medical physicists in the U.S.

The AAPM welcomes the opportunity to comment on the FDA discussion paper, Considerations
for the Regulation of Generative AI-Enabled Medical Devices, under docket FDA-2026-N-7874.
The paper seeks early input rather than establishing guidance, regulatory expectations, or policy
changes. This response addresses the scientific and operational merits of the contemplated
approaches without assuming that any implementation is already required or within existing
statutory authority.1,2

AAPM’s principal recommendation is to establish an explicit evidence chain connecting
manufacturer validation, clinical confirmation, site-specific acceptance testing and
commissioning, post market surveillance, and controlled modification for GenAI-enabled
medical devices used for patient care in radiation medicine specialties (i.e., radiation oncology,
diagnostic radiology and nuclear medicine). Medical physicists, working within the
multidisciplinary clinical team should provide technical leadership for this chain at relevant
clinical practices. The role extends established medical physics responsibilities in technology
evaluation, quality assurance and safe clinical implementation. 3,4

Generative artificial intelligence (GenAI) is rapidly emerging as a transformative technology in
medical physics, with applications spanning radiation therapy, medical imaging, and nuclear
medicine. Built on foundation models (FMs), such as large language models (LLMs), large
vision models (LVMs), and large multimodal models (LMMs), GenAI can synthesize text,
images, and structured data to support clinical decision-making, treatment planning, contour
generation, workflow automation, quality management, education, and research. A scoping
review of GenAI, FMs, and LLMs in radiation therapy physics was recently published in one of
AAPM’s official journals, Medical Physics. This review surveyed emerging applications in
image analysis, workflow automation, clinical decision support, and quality assurance, while
identifying reproducibility, hallucinations, data privacy, and clinical integration as continuing
challenges.9 The review provides radiation-therapy-specific context for the present
recommendations, although it should not be interpreted as a formal AAPM practice guideline.

The risk of hallucinations is not limited to LLMs. Medical imaging studies have demonstrated
cross-institution generalization failures and clinically important weaknesses within subgroups
that aggregate performance can obscure. Reconstruction research further shows that image
priors, whether learned or conventional, can introduce realistic-appearing structures that are not
supported by the measured data. These studies establish plausible failure mechanisms, not a
universal failure rate for all GenAI-enabled medical devices.5,6,7 Importantly, FDA’s joint
transparency principles already identify information for site-specific acceptance testing or
validation as useful lifecycle information. AAPM recommends making that concept operational
through device-specific test methods, accountable personnel, and manufacturer support.8

2. Materials and Methods

2.1. Analysis of the FDA discussion paper

The analysis examined the full FDA discussion paper, including both figures, the benchmark
appendix, and all 26 discussion questions. Targeted English-language searches were conducted
from a wide variety of grounded sources, including official U.S. agencies, professional societies,
and medical literature. Search concepts combined major generative AI risks (e.g., hallucination,
distribution shift), medical physics practice areas, and clinical physics workflow (e.g.,
acceptance testing, commissioning, QA, post market monitoring, and version control). Sources
were selected for direct relevance to a regulatory question or physical failure mechanism, giving
priority to official documents, professional practice reports, reporting guidelines, primary
empirical or methodological studies, and directly relevant reviews.1,9

The analysis distinguished the January 2025 draft AI lifecycle guidance from final FDA
guidance on predetermined change control plans (PCCPs; August 2025), real-world evidence
(December 2025), and cybersecurity (February 2026). AAPM Task Group (TG) 273 and the
joint ESTRO-AAPM AI guidelines informed evaluation and implementation considerations.
These documents have different purposes and regulatory status; citation here does not imply that
a reporting checklist or professional recommendation is an FDA-recognized conformity
standard.10,11,12,13,14,15

2.2. Analytical framework and reporting

Each question in the FDA discussion paper was considered in terms of intended function,
credible failure mechanism, severity and reversibility of consequences, available independent
verification, evidence needed before patient exposure, and responsibility after deployment.
Illustrative imaging and therapy pathways were used to translate general concepts into testable
recommendations. The distinction between empirical evidence and policy judgment is preserved:
citations support the underlying mechanisms or existing guidance, whereas statements that FDA
or stakeholders should take an action are only recommendations. Appendix A maps the response
to every FDA discussion question.

3. Proposed AAPM Recommendations

3.1. Operationalize the risk framework rather than replace it

AAPM supports using activity and consequence as an organizing framework, but not as a selfsufficient risk score or a substitute for function-specific hazard analysis. Its activity axis
distinguishes non-directive information, action-directing information, clinician-supervised
actions, and fully autonomous actions; consequences range from limited to moderate to severe.
The FDA paper already discusses measurement functions, effective supervision, reversibility,
and other contextual considerations. The practical refinement is to specify how sponsors
document those considerations, demonstrate claimed safeguards, and connect residual risks to
evidence and monitoring.1

For each safety-relevant function, AAPM recommends a structured risk record identifying the
output, its downstream clinical use, credible errors, severity, exposure frequency, detection
mechanism, time available for intervention, and responsible reviewer. Following the processoriented approach of TG 100, this record should include both component failures and failures
created by interactions among acquisition, reconstruction, registration, interpretation, planning,
and delivery. Ordinal scores may help prioritize work, but a favorable aggregate score should not
mask an inadequately controlled catastrophic hazard.4

Five operational modifiers deserve particular attention. First, physical significance should be
evaluated in relation to the clinical task: an error near a high dose gradient or in a structure used
for range estimation may matter more than the same numerical error elsewhere. Second,
detectability should identify an actual independent measurement or review process, rather than
assume that a plausible image or fluent explanation reveals an error. Third, downstream
propagation should include multiple tools relying on the same erroneous upstream output.
Fourth, reversibility and intervention time should distinguish a revisable draft from an action
already committed to treatment. Fifth, cumulative and common-cause exposure should include
repeated fractions and deployment of the same defective configuration at multiple sites. These
are proposed extensions of established registration, dosimetric verification, and risk-management
practices to GenAI workflows.4,16,17

An information label should not automatically imply low risk. A synthetic CT volume, generated
contour, or quantitative result can direct clinical action even without an imperative sentence.
Conversely, not every automated action has severe consequences. Sponsors should justify the
activity level using the output’s practical influence and available safeguards. For interactive
systems, testing should include multi-turn conversations in which generic discussion becomes
patient-specific advice or a tool-executed action; risk assessment should encompass the most
consequential supported use, not only the initial screen.1

AAPM recommends evaluating human supervision as a measurable safeguard. The intended
reviewer must have relevant competence, access to independent information, adequate time, and
authority to reject or stop an action. Generalist performance should not substitute for specialist
competence when the function makes specialist-level judgments. Patient-facing functions should
demonstrate comprehension, appropriate escalation, and avoidance of misleading reassurance in
their intended populations. Under-escalation, unnecessary escalation, delay, and workload should
be reported separately rather than collapsed into an unvalidated universal utility score. These
recommendations build on FDA’s emphasis on the human-AI team and on clinical evaluation
that includes human factors.8,18

3.2. Require physics-based, task-specific competency evidence

AAPM supports complementary benchmarking and clinical confirmation. Benchmark selection
should follow the device’s hazards and claims, not a generic collection of medical examinations.
Within the FDA’s competency categories, measurement fidelity, geometric integrity, and
downstream dosimetric effects should be made explicit for imaging and therapy functions. TG
233 provides a task-based framework for CT performance evaluation; TG 132 and TG 219
provide relevant foundations for registration assessment and independent calculation-based
verification. Applying these foundations to GenAI will require device-specific test development,
not merely declaring compliance with existing reports.1,16,17,19

For image reconstruction and synthesis, AAPM recommends evaluating the clinically relevant
addition, deletion, displacement, and distortion of structures; preservation of low-contrast
findings; quantitative bias; and performance across supported acquisition conditions. Global
image similarity, visual realism, or average noise reduction should not be accepted as the sole
evidence for a diagnostic or quantitative claim. Measurement consistency is valuable but cannot
by itself verify structures that the acquired data do not adequately constrain. Bhadra and
colleagues’ analysis of reconstruction hallucinations provides a specific mathematical and
experimental basis for this limitation.7

For synthetic CT used in treatment planning, tests should connect image errors to target coverage
and organ-at-risk dose; for proton therapy, where CT-number errors translate directly into beam
range errors, tests should also address range accuracy and plan robustness. Published syntheticCT work has evaluated both image and dosimetric performance, including tissue-dependent
behavior, demonstrating why image-only evaluation is incomplete.20 As a logical consequence of
shared inputs, two dose engines can agree when both use the same erroneous synthetic anatomy
or density map. AAPM therefore recommends verification with sufficiently independent
information, selected for the anticipated failure mode, rather than treating computational
redundancy as automatic independence.7,17 Table 1 gives illustrative test specifications; its
proposed controls are not universal numerical tolerances.
Table 1. Proposed physics-informed evaluation specifications for GenAI-enabled medical
devices in medical physics practice domains.

Application Failure to challenge Proposed endpoint Independent evidence

Image A plausible structure is Task-specific detection and Reference measurements,
reconstruction added, removed, or measurement performance validated phantoms, and
or enhancement displaced; quantitative across supported dose and clinically adjudicated cases.7,19
signal is biased. acquisition settings.

Synthetic CT Anatomy or density error Target and/or organ-at-risk Reference imaging or
for treatment leads to dosimetric error, dosimetric error, increased measurements sufficiently
planning or range/robustness issues particle range uncertainty or independent of the synthesis
(in particle therapy) reduced plan robustness. model.17,20

Generated A small, high- Landmark and site-specific Independently reviewed
contours or consequence boundary is surface errors plus reference structures and siteimage mistaken despite downstream dose specific registration tests.16
registration favorable global overlap. consequences.

Treatment- Wrong patient, laterality, Clinically weighted error and Direct comparison with the
related text or units, prescription, or omission rates. authorized source record and
multimodal timing is carried into a independently implemented
extraction clinical action. identity/unit checks.4

Agentic A wrong or repeated tool Complete-trajectory safety, Controlled fault tests and
workflow action proceeds after stale permission compliance, independent clinical/physics
execution input or partial failure. recovery, and critical-limit authorization.4,13,17
preservation.

For generative text and multimodal outputs, AAPM recommends measuring unsupported
assertions, clinically significant omissions, contradictions, incorrect patient association, units,
laterality, temporal context, and unsafe task completion. Evaluation should specify how an openended output can be acceptable without matching one reference sentence. Structured expert
adjudication should report disagreement and use an explicit harm-oriented rubric. CLAIM 2024,
TRIPOD-LLM, and DECIDE-AI provide complementary reporting structures; completing them
improves transparency but does not itself establish safety.18,21,22

Stochastic behavior should be evaluated under the deployed inference policy. Sponsors should
prespecify repeated-run testing, relevant prompt paraphrases, input order, missing or
contradictory context, and tool-response perturbations. AAPM recommends reporting the
distribution of clinically consequential failures, not selecting the best of multiple outputs when
clinical deployment uses only one. Safety-critical boundary tests should include both appropriate
refusal and successful completion of permitted tasks. A device that avoids all difficult cases by
refusing may be safe on accepted cases yet ineffective for its claimed population; both coverage
and conditional error should therefore be reported.22

The same competency architecture can apply to vision-language and predictive world models,
but the task tests must change. AAPM proposes testing cross-modal patient and time alignment,
sensitivity to missing modalities, physical consistency, and temporal prediction error over the
interval in which an action is taken. Linguistic plausibility should not stand in for geometric or
quantitative accuracy. Where simulated futures drive an action, validation should test the action’s
consequences against independent observations, including failure under unexpected anatomical
or acquisition changes. This is a proposed extension, not an assertion that a single validated
benchmark presently covers all such architectures.

3.3. Establish benchmark validity and meaningful statistical evidence

Sponsors should specify the clinical claim, target population, reference standard, sampling
design, and acceptance criteria before evaluating the final system. AAPM recommends
independent external and temporal testing, patient-level separation where patients contribute
multiple images or encounters and prespecified analyses of clinically relevant subgroups and
acquisition environments. Cross-institution generalization and hidden-stratification studies
support this emphasis on performance beyond a pooled development distribution.5,6,14,21

A benchmark should demonstrate that its cases and scoring rubric represent hazards and
decisions within the intended use. Where a score is proposed as a gate to clinical evaluation or as
a proxy for safe deployment, sponsors should provide evidence linking that score to independent
clinically adjudicated performance. AAPM recommends fresh or sequestered test material,
provenance documentation, controls against training-test leakage, and independent review of
sponsor-developed benchmarks. Sponsor-specific tests can be necessary for novel functions but
should not be the only evidence when optimization to the test cannot be excluded.14,21,22

Synthetic data are particularly useful for controlled perturbations of anatomy, geometry, motion,
noise, acquisition parameters, and rare but physically specified faults. Their credibility should be
assessed against independent measurements and real cases in the relevant range. Synthetic data
are less persuasive as the sole evidence for rare diseases, underrepresented populations, or
realistic patient-clinician interactions when those characteristics are not independently verified.
AAPM recommends combining physical phantoms, digital simulations, and real clinical cases
for complementary purposes, rather than assuming that a larger synthetic dataset repairs absent
real-world representation.16,19,20 When a generative model creates test cases or reference labels,
evaluation should assess whether it shares training data, architecture-related weaknesses, or
systematic omissions with the device being tested. Independent clinical review and physically
grounded challenge cases should address those possible shared blind spots; agreement between
two models alone should not establish a reference standard.

Synthetic and real-world results should normally be reported separately. Combining them into
one clinical performance estimate should require a common estimand, a justified relationship to
the intended population, prespecified weighting, assessment of residual distribution differences,
and sensitivity analyses. Benchmark and clinical-confirmation observations should likewise not
be pooled merely because both yield an accuracy percentage. Sponsors should not count repeated
generations, slices, or correlated encounters as independent patients. These recommendations
apply the transparency and sampling principles of the cited evaluation guidance to the FDA’s
proposed evidence combination.14,21,22

Sample size should follow the harm-relevant endpoint and the precision or noninferiority claim.
Mean performance and statistical significance alone should not establish an acceptable safety
margin. As an illustration, observing zero events in 300 independent, identically distributed
opportunities leaves an approximately 1% upper one-sided 95% confidence bound on the event
probability, using the familiar rule of three. This is not a proposed acceptance threshold; the
calculation assumes independence and does not establish a bound after unaddressed distribution
shift.23 AAPM recommends explicit confidence intervals, analysis of missing outcomes, and
reporting of high-severity failures even when average performance satisfies a target.

3.4. Preserve clinical confirmation and test the intended human-AI workflow

Clinical confirmation should address uncertainty that the non-clinical assessment cannot resolve.
AAPM supports a risk-proportionate choice of methods, not a requirement for a randomized trial
for every device. Independent retrospective cases may be sufficient for some bounded outputs
with credible reference standards. Shadow-mode evaluation can identify integration and
population mismatches without influencing care but cannot by itself establish how clinicians
behave when they rely on the device. Prospective human-factors or workflow evaluation is
particularly important when human oversight is a claimed risk control; stronger comparative
designs may be needed for causal claims about patient benefit.1,12,18 Consistent with this need,
Qi et al. identify limited large-scale prospective clinical validation and call for coordinated multiinstitutional studies with standardized evaluation frameworks.9

The comparator should match the task and claim. Calibrated measurements, validated reference
images, expert-reviewed structures, or clinically justified dose criteria are preferable to a
popularity judgment where physical quantities can be assessed. For open-ended clinical
recommendations, AAPM favors an explicitly defined acceptable-care standard adjudicated by
appropriately qualified clinicians over an unexplained median-practitioner score. Both AIassisted and unassisted workflows should be assessed when the device is intended to assist a
clinician. A randomized clinical-vignette study found that access to an LLM did not significantly
improve physicians’ diagnostic reasoning even though the model alone performed better,
illustrating why model-alone performance does not establish team benefit.16,19,24

A realistic counterfactual — the care patients would receive without the device — can include
delayed specialist review or limited access to care, but sponsors should justify that comparison
with evidence from the intended setting. AAPM recommends reporting benefits relative to that
actual alternative while retaining minimum safety criteria for the task. The absence of an ideal
comparator should not justify hiding avoidable high-severity errors. Conversely, imposing an
unavailable idealized specialist workflow as the only comparator could obscure a genuine access
benefit; the relevant benefit-risk question must remain explicit.1 Moreover, there should be user
manuals that list error messages along with definitions. Error messages should follow the rules
outlined in IEC 63183:2019 so that users understand the cause of errors and what steps to take to
support safety in a clinical environment.

Qualified independent testing organizations can maintain sequestered datasets, perform
reproducible technical testing, and coordinate blinded adjudication. AAPM recommends
documented clinical and physics competence, conflict-of-interest management, auditability,
consistent methods, and protection against exclusive access to essential benchmarks. Third-party
participation should be optional where equally credible evidence can be generated otherwise and
should not transfer FDA decision-making authority or manufacturer accountability to a
certificate issuer. Publicly documented test specifications and multiple qualified providers would
be preferable to a single proprietary gatekeeper.1

3.5. Make site-specific acceptance and commissioning an explicit deployment
bridge

Product-level clinical confirmation and site-specific deployment verification answer different
questions. The former supports the device’s claimed performance across its intended conditions;
the latter establishes whether a specific implementation operates within those conditions and
integrates safely with the site’s equipment and workflow. AAPM recommends this distinction
expressly. Site-specific commissioning should not mean repeating a pivotal trial or obtaining a
new FDA authorization at every site. Neither should manufacturer benchmark success be treated
as sufficient proof of site-specific readiness.3,8,15

For imaging and therapy systems within the medical physics practice, a designated medical
physicist with required credentials should lead technical acceptance and commissioning with
responsible physicians and other clinical professionals. Acceptance should verify the delivered
configuration against documented specifications. Commissioning should establish supported
site-specific uses, baseline performance, integration integrity, independent checks, failure
response, and training before clinical release. The scope should reflect the application and the
medical physicist’s specialty and competence; professional qualification alone does not establish
competence in every new AI technology.3,4,16

AAPM recommends a manufacturer-supplied deployment package containing supported
acquisition and patient conditions, contraindications and limitations, version identification,
representative validation results, test assets, instructions for site-specific testing, proposed
performance measures, known failure modes, monitoring interfaces, update procedures, and a
safe fallback workflow. Sites should be provided with information clearly marked on the model
and guidance on the minimum or range of datasets required for a medical device to be safely
deployed in a clinical setting. Sites should be able to retrieve the information needed for
independent verification without reverse-engineering the product. This operationalizes the
acceptance-testing and lifecycle information already contemplated in the FDA’s transparency
principles.8

A site-specific test plan should include representative and challenging cases, supported scanner
and protocol combinations, geometric and quantitative integrity, patient and data association,
transfer to downstream systems, unavailable services, and recovery after interruption.
Radiotherapy commissioning should test the complete path to the point of clinical action, not
only file generation. Release should be documented by the accountable technical and clinical
owners, with defined restrictions for untested configurations. AAPM proposes reacceptance or
partial recommissioning after safety-relevant model, scanner, protocol, interface, or workflow
changes.3,16,17

The term on-site should refer to meaningful site accountability and the practical ability to inspect
and intervene where needed, not an automatic requirement for continuous physical presence
besides every AI application. AAPM recommends an on-site medical physicist when physical
testing, equipment integration, or the consequence and timing of intervention require it. Shared
or remote medical physics support may be appropriate when access, response times,
responsibilities, and escalation are demonstrably adequate. This proposed flexibility should
support access without diluting competence or shifting un-resourced obligations onto small or
rural facilities.3 Table 2 assigns the responsibilities for different stakeholders.
Table 2. Proposed allocation of lifecycle responsibilities for GenAI-enabled medical devices in
radiation therapy, medical imaging, and nuclear medicine.

Accountable Primary responsibility Operational evidence or handoff
participant

Device manufacturer Substantiate intended use; supply Versioned deployment package, test assets,
implementation tools; evaluate monitoring access, safety notices, update
changes; conduct product-wide evidence, and coordinated response to
surveillance and corrective action. cross-site signals.1,8,11

Foundation-model Supply relevant upstream information Version-specific evidence, update and
provider and agreed version/change support to retirement notice, audit support, and
device sponsors. authorized Device Master File (MAF)
reference where available.1,25

Medical physicist Lead site-specific technical acceptance, Documented baseline, site-specific test
commissioning, relevant ongoing results, action limits, restrictions, technical
quality assurance, and technical release recommendation, and escalation
assessment of changes. record.3,4

Clinical team and Own clinical review, workflow design, Named clinical owner, training,
healthcare institution staffing, access controls, and demonstrated oversight, patient-level
institutional release and incident review, and a resourced fallback
response. process.3,18

FDA, independent Maintain distinct regulatory, Transparent evidence expectations,
evaluators, and independent testing, and professional- auditable evaluation methods, nonexclusive
professional method development roles. test infrastructure, and competence
organizations development; FDA retains regulatory
decisions.1

3.6. Connect site-specific surveillance to manufacturer-led post market
monitoring

AAPM recommends that the FDA recognize physicist-led site-specific technical monitoring as
an operational foundation for post market assurance in imaging and therapy, integrated with
clinical review and manufacturer-led surveillance across sites. A minimum monitoring record
should identify the relevant model and application versions, prompt and retrieval configuration,
preprocessing, acquisition context, external tools, output, review status, and consequential action,
subject to data minimization and access controls. Manufacturer monitoring should remain
responsible for detecting cross-site signals and initiating coordinated corrective action; sitespecific quality assurance is complementary, not a substitute.1,8,13

The monitoring plan should track all eligible uses and attempts, not only successfully completed
outputs. AAPM recommends separating technical failures, clinically consequential errors,
appropriate refusals, unnecessary refusals, overrides, near misses, and delays. Clinical review
samples should include routine cases as well as enriched high-risk and underrepresented groups,
with appropriate weighting for population estimates. The plan should distinguish changes in
input distributions, changes in device behavior, and changes in clinical performance; an inputdrift alert is a reason to investigate, not proof of clinical failure or safety.1,14,21

Monitoring frequency should be justified by use volume, detectability, severity, and time to harm
rather than a universal annual or monthly interval. AAPM proposes immediate transaction-level
checks for patient identity, units, permissions, and critical limits; periodic review of actual
clinical outputs; and triggered reevaluation following changes, incidents, or emerging subgroup
signals. When monitoring is credited as a preventive safeguard, the total time to detect, assess,
and intervene must be shorter than the relevant time to irreversible harm. Retrospective
surveillance remains useful for learning but cannot prevent an action already completed.4,17

Prespecified action thresholds should trigger graded responses: investigation, intensified review,
restriction of use, suspension of an affected function, or rollback to a validated configuration.
AAPM recommends assigning an owner and response deadline for each threshold, rehearsing
recovery, and recording the reasoning behind continued use after a signal. Statistical control
limits should not replace clinical severity thresholds. Lack of observed events in a small
subgroup should be reported as limited evidence, not a demonstration of safety.4,23

Machine-based supervisory agents may assist triage, anomaly detection, and workload reduction,
but should be evaluated as safety-relevant components. AAPM recommends testing their missedhazard rates, false alarms, latency, subgroup behavior, and failure when the primary and
supervisory systems share data or model dependencies. Different names or prompts do not
establish independent failure modes. For high-consequence quantitative actions, independent
physics checks or externally enforced constraints should supplement generative reviewers. A
supervisory agent should not be the sole unvalidated arbiter of another agent’s safety.1,4,17

AAPM would support accepting selected residual premarket uncertainty only when the
unresolved issue is explicit, the expected benefit justifies exposure, monitoring can detect the
relevant problem in time, interventions are effective and affected patients and sites can be
identified. This should not become a general exchange of weaker premarket evidence for
promises of future observation. When errors are difficult to detect before irreversible highconsequence treatment, stronger premarket evidence and preventive controls should remain
necessary. The clinical consequences of suspension or delayed treatment should also inform the
fallback plan.1,4

3.7. Control changes across the complete safety-relevant configuration

AAPM recommends treating the deployed device configuration as more than a model name.
Safety-relevant changes may involve model weights, system prompts, retrieval sources,
embedding or ranking components, preprocessing, decoding settings, tools, permissions,
orchestration, or user interfaces. Sponsors should maintain a versioned configuration inventory
and a traceable connection between each clinical output and its applicable configuration. Sitespecific prompts or retrieval corpora should be configurable only within documented, validated
boundaries unless a further change assessment is performed.1,22 Qi et al. also identify prompt
sensitivity and model-version changes as challenges to reproducibility and clinical validation,
reinforcing the need to document and evaluate the deployed configuration.9

The extent of reevaluation should follow a documented impact analysis. A demonstrably
nonfunctional change may justify verification within the quality system, whereas a change
affecting quantitative output, refusal behavior, clinical scope, or tool execution warrants targeted
testing and regression testing of critical unchanged functions. AAPM recommends additional
clinical confirmation when changed behavior invalidates the assumptions supporting prior
confirmation, and site-specific reacceptance when the change affects a site’s technical baseline
or workflow. These are evidence recommendations; the applicable regulatory pathway must still
be determined under FDA requirements and the device’s authorization.10,11
Predetermined Change Control Plans (PCCPs) for medical devices can support bounded, planned
modifications with defined development, validation, implementation, and impact-assessment
methods. AAPM recommends that GenAI PCCPs specify performance boundaries, prohibited
functional expansions, affected subgroups, monitoring, notification, site-specific release criteria,
and rollback. A promise to rerun benchmarks should not function as advance permission for
unknown changes in intended use or autonomy. Where a future modification cannot be
adequately characterized within an authorized plan, sponsors should obtain the appropriate
additional regulatory assessment rather than assume coverage.11

For third-party foundation models, AAPM recommends contractual advance notice of safetyrelevant changes, stable version selection, documented support periods, access to prerelease
testing, and a transition plan when a version is retired. Technical controls should detect
configuration changes and prevent silent substitution; behavioral sentinel testing should
supplement, not replace, trustworthy version identification. An unannounced update should
trigger risk-based investigation and restriction or fallback when the prior safety evidence no
longer applies. Sponsors remain accountable for their device even when they did not initiate the
upstream change.1,11,13

3.8. Support reusable foundation-model evidence without blanket
authorization

AAPM supports exploring voluntary, version-specific Foundation Model Device Master Files
(MAFs) as a mechanism to reduce repeated submission of confidential upstream evidence. The
existing MAF mechanism permits authorized reference to information held by FDA; a master
file is not a marketing application or independent approval. The discussion paper similarly
distinguishes foundation-model documentation from authorization of a particular medical
device.1,25,26

AAPM recommends a structured content specification covering model identity and provenance,
relevant training-data limitations, supported modalities and context, known healthcare failure
modes, subgroup evaluation, refusal and safety-policy behavior, security testing, change history,
update commitments, audit capabilities, and retirement arrangements. The file should distinguish
evidence about the base model from evidence about specific adaptations. A date-stamped change
log and clear responsibility for updates would be essential to practical usefulness. Device
sponsors should still receive the safety-critical information needed to operate and monitor their
products, even when trade-secret detail is supplied confidentially to FDA.1,8,13

Participation incentives should emphasize reusable evidence and reduced duplication, not
automatic regulatory exemptions. AAPM recommends equivalent routes for adequate upstream
information, including authorized supplier documentation and independent assessment, when a
provider does not maintain an MAF. Nonparticipation alone should not be disqualifying, but
missing safety-relevant evidence should not be presumed irrelevant. Neither a favorable model
card nor an MAF should replace end-to-end validation of the intended device.25,26

Guardrails should be explicitly defined regarding the conditions for which agentic systems will
operate. AAPM recommends implementing mechanisms for confirming that sequences do not
cross boundaries from one patient to another mid-stream, which is crucial for patient safety.

3.9. Evaluate agentic systems as bounded action sequences

Agentic radiotherapy planning is no longer solely hypothetical: a published feasibility study
evaluated LLM-driven planning agents in lung and cervical cancer cases. That small planning
study supports technical feasibility, not prospective evidence that autonomous delivery is safe.27
AAPM therefore recommends evaluation of complete action sequences, including unsuccessful
paths and recovery, rather than extrapolation from language accuracy or isolated tool-call
success.

For high-consequence workflows, AAPM recommends a validated boundary between proposing
an action and committing it. Permitted tools, patient context, units, target objects, allowable
parameters, and execution privileges should be constrained outside the generative model where
feasible. Irreversible treatment actions should require appropriate clinical authorization and
independent, validated safeguards; claims of autonomous operation should require corresponding
evidence rather than inheriting assurance from a supervised configuration. Existing physical
interlocks and independent verification should not be bypassed by a conversational
interface.4,13,17

Testing should include incomplete or conflicting inputs, incorrect tool responses, loss of
connectivity, stale patient information, repeated requests, partial execution, and malicious
instructions embedded in retrieved material. AAPM recommends checking that retries cannot
duplicate consequential actions and that a partial failure returns the system to a defined safe state.
Rollback of software or a plan does not reverse radiation already delivered; therefore,
recoverability should be assessed at the clinical action level, not only at the database level.4,13

Auditability should preserve relevant inputs, retrieved sources, outputs, tool commands and
responses, configuration identifiers, approvals, and timestamps. A generated explanation or
apparent self-critique should not be accepted as proof that the action was correct. Independent
safety constraints and observed performance should determine whether an agent can proceed.
Autonomous expansion of tool privileges or clinical scope should be prohibited outside validated
change-control procedures. These are proposed agent-specific safeguards grounded in process
risk management and device cybersecurity, not assertions that current guidance already mandates
each detail.4,13

4. Discussion

4.1. The distinct contribution of medical physics

The central contribution of this response is a connection between regulatory evidence and sitespecific clinical operation. A benchmark asks whether a defined system meets test criteria.
Clinical confirmation asks whether those criteria correspond to acceptable performance for the
intended use. Site-specific acceptance and commissioning ask whether the delivered system,
configured with a particular site’s equipment, data, personnel, and workflow, remains within the
demonstrated conditions. Ongoing monitoring asks whether that assurance persists. Treating
these as complementary rather than interchangeable follows the practical logic of medical
physics quality management and FDA transparency principles.3,4,8
This approach is especially important when an output appears credible, but its critical properties
are not directly visible. A synthetic image can be visually persuasive without independently
verified measurement accuracy; a generated plan can satisfy a computed objective while
depending on an incorrect input. The cited reconstruction and synthetic-CT studies support these
distinctions, but do not establish that every generative method is inferior to conventional
methods.7,20 The policy implication is task-specific verification.

4.2. Proportionality, feasibility, and accountability

AAPM recommends that the FDA distinguish information and capabilities that manufacturers
should provide from professional activities that healthcare institutions should organize. Device
labeling, deployment, monitoring access, and supplier change management should not rely on
unspecified institutional resources. Institutions, in turn, should identify personnel with authority,
protected time, and technical access to carry out the site-specific tasks on which safe deployment
depends. Physicians retain responsibility for clinical judgments within their roles, whereas
medical physicists provide technical assurance within theirs. Shared responsibility should make
ownership more explicit, not more diffuse.1,3

Implementation should be proportional. Shared test libraries, interoperable logs, reusable site
protocols, and qualified regional physics support could reduce duplicated effort. AAPM
recommends the collaborative development of representative challenge datasets and test methods
with patient, clinical, physics, industry, and regulatory participation, followed by assessment of
reproducibility and clinical relevance. Evaluation costs and workforce burden should be
measured during pilots; this analysis supplies no cost-effectiveness estimate. Lower-resource
sites should receive practical implementation support.

4.3. Limitations and priorities for further work

This response is a targeted policy analysis, not an exhaustive evidence synthesis or a formally
endorsed practice guideline. Some evidence predates current GenAI architectures and supports
general failure mechanisms rather than device-specific incidence. Several proposed controls,
including monitoring cadence, agentic acceptance criteria, and the content of reusable
foundation-model evidence, require prospective operational evaluation. No universal numerical
tolerance or safe autonomy threshold is established here.

Future work should test whether the proposed site-specific and manufacturer monitoring
measures detect clinically meaningful deterioration with acceptable workload; evaluate
interobserver consistency in harm adjudication; develop modality-specific reference datasets;
and examine performance after actual scanner, protocol, and model changes. FDA should engage
AAPM and other relevant organizations in this work without presuming that any organization
has committed resources or that a professional report automatically acquires regulatory status.

5. Conclusions

AAPM recommends a lifecycle approach that retains the FDA’s risk heuristic but connects it to
observable hazards, task-specific evidence, and enforceable clinical safeguards. Non-clinical
benchmarking should complement, not replace, clinical confirmation; neither should substitute
for risk-proportionate site acceptance and commissioning. For radiation therapy, medical
imaging, and nuclear medicine applications within medical physics practice, physics-led sitespecific technical quality assurance should be a recognized operational foundation of post market
monitoring, supported by manufacturer-provided information, tools, and coordinated
surveillance. Whole-configuration change control, reusable but non-authorizing foundationmodel evidence, and independent constraints on agentic actions should preserve that assurance as
devices evolve.
Appendix A. Responses Mapped to the FDA Discussion Questions
The entries below are the response’s recommendations, not quotations or restatements of FDA
policy. Question numbering follows the discussion paper. The cited main-text sections provide
the supporting rationale and references.1

A.1. Risk assessment: Questions 1-6
Question 1. Retain the two axes as a heuristic and add a structured, auditable record of physical
significance, detectability, propagation, reversibility, and effective intervention. See Section 3.1.

Question 2. Judge directiveness by practical influence on clinical action, including images and
quantitative outputs, not wording or disclaimers alone. See Section 3.1.

Question 3. Test patient comprehension, false reassurance, and escalation in the intended
population; demonstrate rather than presume professional supervision. See Section 3.1.

Question 4. Require task-specific competence and specialist comparators for specialist claims;
general medical knowledge should not substitute for physics or clinical specialty expertise. See
Sections 3.1-3.4.

Question 5. Evaluate complete multi-turn interactions and the most consequential supported use,
including transitions from information to tool execution. See Sections 3.1 and 3.9.

Question 6. Report missed escalation, unnecessary escalation, delay, and burden separately,
using context-specific acceptance criteria rather than a universal combined score. See Sections
3.1 and 3.6.

A.2. Premarket evaluation: Questions 7-17
Question 7. Support benchmarking plus clinical confirmation, followed by a distinct, riskproportionate bridge to site-specific acceptance and commissioning. See Sections 3.2-3.5.

Question 8. Scale evidence to consequence, autonomy, detectability, and available safeguards; a
supervision label should reduce evidence only when its effectiveness is demonstrated. See
Sections 3.1 and 3.4.

Question 9. Add explicit physical, geometric, dosimetric, and end-to-end integration tests. Use
relevant AAPM methods and reporting frameworks without treating them as universal regulatory
conformity standards. See Section 3.2 and Table 1.

Question 10. Require benchmark-to-clinical relevance, protected external testing, provenance,
and independent scrutiny of sponsor-developed tests. See Section 3.3.
Question 11. Select confirmation methods to resolve the actual remaining uncertainty; shadow
mode cannot alone validate reliance-dependent human behavior or clinical benefit. See Section
3.4.

Question 12. Prespecify endpoints, independent sampling units, uncertainty, and subgroup
analyses. Pool evidence only under a justified common estimand and transparent weighting. See
Section 3.3.

Question 13. Use credible synthetic data for controlled physical perturbations, not as a substitute
for unverified pathology or population representation; avoid shared generator-evaluator blind
spots. See Section 3.3.

Question 14. Prefer physical reference standards where available and explicitly adjudicated
acceptable-care criteria otherwise; evaluate the intended human-AI team. See Section 3.4.

Question 15. Justify the real alternative with setting-specific evidence while preserving
minimum safety requirements. See Section 3.4.

Question 16. Support competent, independent, auditable third parties with nonexclusive methods
and alternative evidence routes; retain FDA and manufacturer accountability. See Section 3.4.

Question 17. Retain architecture-neutral competency domains but add modality alignment,
physical consistency, temporal prediction, and action-consequence testing as applicable. See
Section 3.2.

A.3. Post market monitoring and changes: Questions 18-24
Question 18. Accept selected residual uncertainty only when detection and intervention can
control the identified risk; do not trade away prevention of poorly detectable irreversible harm.
See Section 3.6.

Question 19. Combine site-specific technical checks, sampled clinical review, and fleet
surveillance; justify cadence and change-triggered reevaluation by the hazard and time to harm.
See Section 3.6.

Question 20. Validate supervisory agents, including shared failure modes, and retain
independent checks for high-consequence actions. See Section 3.6.

Question 21. Assign a medical physicist with required credentials technical leadership sitespecific clinical ownership to relevant clinicians, institutional resources, and manufacturer
responsibility for product surveillance and corrective action. See Sections 3.5-3.6 and Table 2.

Question 22. Use impact analysis to scale targeted, regression, clinical, and site-specific testing;
determine quality-system, PCCP, or additional submission pathways under applicable
requirements. See Section 3.7.
Question 23. Bound PCCPs by function, performance, validation methods, and prohibited
expansions; do not grant open-ended permission for uncharacterized future changes. See Section
3.7.

Question 24. Combine upstream notification agreements, stable version selection, configuration
identification, prerelease testing, and validated fallback; retain sponsor accountability. See
Section 3.7.

A.4. Foundation models and agents: Questions 25-26
Question 25. Support voluntary, current, version-specific MAFs and equivalent evidence routes,
without treating documentation as model authorization or replacing device-specific evidence.
See Section 3.8.

Question 26. Evaluate complete action sequences and recovery; constrain permissions and
critical parameters externally and require justified authorization and safeguards before
irreversible actions. See Section 3.9.
References
1. US Food and Drug Administration. Considerations for the Regulation of Generative AI-Enabled Medical
Devices: Discussion Paper and Request for Feedback. 2026. Accessed September 25, 2026. https://www.fda.gov/
media/194242/download

2. US Food and Drug Administration. Considerations for the Regulation of Generative AI-Enabled Medical
Devices: Discussion Paper and Request for Feedback. Docket FDA-2026-N-7874. Accessed September 25, 2026.
https://www.fda.gov/medical-devices/digital-health-center-excellence/considerations-regulation-generative-aienabled-medical-devices-discussion-paper-and-request

3. Lim SB, Busse N, Clements J, et al. AAPM Medical Physics Practice Guideline 10.b: Scope of practice for
clinical medical physics. J Appl Clin Med Phys. 2025;26(11):e70230. doi:10.1002/acm2.70230

4. Huq MS, Fraass BA, Dunscombe PB, et al. The report of Task Group 100 of the AAPM: Application of risk
analysis methods to radiation therapy quality management. Med Phys. 2016;43(7):4209-4262.
doi:10.1118/1.4947547

5. Zech JR, Badgeley MA, Liu M, Costa AB, Titano JJ, Oermann EK. Variable generalization performance of a
deep learning model to detect pneumonia in chest radiographs: A cross-sectional study. PLoS Med.
2018;15(11):e1002683. doi:10.1371/journal.pmed.1002683

6. Oakden-Rayner L, Dunnmon J, Carneiro G, Ré C. Hidden stratification causes clinically meaningful failures in
machine learning for medical imaging. In: Proceedings of the ACM Conference on Health, Inference, and Learning.
2020:151-159. doi:10.1145/3368555.3384468

7. Bhadra S, Kelkar VA, Brooks FJ, Anastasio MA. On hallucinations in tomographic image reconstruction. IEEE
Trans Med Imaging. 2021;40(11):3249-3260. doi:10.1109/TMI.2021.3077857

8. US Food and Drug Administration, Health Canada, Medicines and Healthcare products Regulatory Agency.
Transparency for Machine Learning-Enabled Medical Devices: Guiding Principles. June 2024. Accessed September
25, 2026. https://www.fda.gov/medical-devices/artificial-intelligence-enabled-medical-devices/transparencymachine-learning-enabled-medical-devices-guiding-principles

9. Qi XS, Wang Y, Yang X, et al. Generative AI, foundation models and large language models in radiation therapy
physics: Clinical applications, challenges, and future directions. Med Phys. 2026;53(8):e70616.
doi:10.1002/mp.70616

10. US Food and Drug Administration. Artificial Intelligence-Enabled Device Software Functions: Lifecycle
Management and Marketing Submission Recommendations. Draft guidance. January 2025. Accessed September 25,
2026. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/artificial-intelligence-enableddevice-software-functions-lifecycle-management-and-marketing

11. US Food and Drug Administration. Marketing Submission Recommendations for a Predetermined Change
Control Plan for Artificial Intelligence-Enabled Device Software Functions. Final guidance. August 2025. Accessed
September 25, 2026. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/marketingsubmission-recommendations-predetermined-change-control-plan-artificial-intelligence

12. US Food and Drug Administration. Use of Real-World Evidence to Support Regulatory Decision-Making for
Medical Devices. Final guidance. December 2025. Accessed September 25, 2026. https://www.fda.gov/regulatory25
information/search-fda-guidance-documents/use-real-world-evidence-support-regulatory-decision-making-medicaldevices

13. US Food and Drug Administration. Cybersecurity in Medical Devices: Quality Management System
Considerations and Content of Premarket Submissions. Final guidance. February 2026. Accessed September 25,
2026. https://www.fda.gov/regulatory-information/search-fda-guidance-documents/cybersecurity-medical-devicesquality-management-system-considerations-and-content-premarket

14. Hadjiiski L, Cha K, Chan HP, et al. AAPM task group report 273: Recommendations on best practices for AI
and machine learning for computer-aided diagnosis in medical imaging. Med Phys. 2023;50(2):e1-e24.
doi:10.1002/mp.16188

15. Hurkmans C, Bibault JE, Brock KK, et al. A joint ESTRO and AAPM guideline for development, clinical
validation and reporting of artificial intelligence models in radiation therapy. Radiother Oncol. 2024;197:110345.
doi:10.1016/j.radonc.2024.110345

16. Brock KK, Mutic S, McNutt TR, Li H, Kessler ML. Use of image registration and fusion algorithms and
techniques in radiotherapy: Report of the AAPM Radiation Therapy Committee Task Group No. 132. Med Phys.
2017;44(7):e43-e76. doi:10.1002/mp.12256

17. Zhu TC, Stathakis S, Clark JR, et al. Report of AAPM Task Group 219 on independent calculation-based
dose/MU verification for IMRT. Med Phys. 2021;48(10):e808-e829. doi:10.1002/mp.15069

18. Vasey B, Nagendran M, Campbell B, et al. Reporting guideline for the early-stage clinical evaluation of decision
support systems driven by artificial intelligence: DECIDE-AI. Nat Med. 2022;28(5):924-933. doi:10.1038/s41591022-01772-9

19. Samei E, Bakalyar D, Boedeker KL, et al. Performance evaluation of computed tomography systems: Summary
of AAPM Task Group 233. Med Phys. 2019;46(11):e735-e756. doi:10.1002/mp.13763

20. Thummerer A, Seller Oria C, Zaffino P, et al. Deep learning-based 4D-synthetic CTs from sparse-view CBCTs
for dose calculations in adaptive proton therapy. Med Phys. 2022;49(11):6824-6839. doi:10.1002/mp.15930

21. Tejani AS, Klontzas ME, Gatti AA, et al. Checklist for Artificial Intelligence in Medical Imaging (CLAIM):
2024 update. Radiol Artif Intell. 2024;6(4):e240300. doi:10.1148/ryai.240300

22. Gallifant J, Afshar M, Ameen S, et al. The TRIPOD-LLM reporting guideline for studies using large language
models. Nat Med. 2025;31(1):60-69. doi:10.1038/s41591-024-03425-5

23. Hanley JA, Lippman-Hand A. If nothing goes wrong, is everything all right? Interpreting zero numerators.
JAMA. 1983;249(13):1743-1745. https://pubmed.ncbi.nlm.nih.gov/6827763/

24. Goh E, Gallo R, Hom J, et al. Large language model influence on diagnostic reasoning: A randomized clinical
trial. JAMA Netw Open. 2024;7(10):e2440969. doi:10.1001/jamanetworkopen.2024.40969

25. US Food and Drug Administration. Device Master Files. Accessed September 25, 2026. https://www.fda.gov/
medical-devices/premarket-submissions-selecting-and-preparing-correct-submission/device-master-files

26. US Food and Drug Administration. PMA Special Considerations. Section: Device Master Files. Accessed
September 25, 2026. https://www.fda.gov/medical-devices/premarket-approval-pma/pma-special-considerations
27. Wang Q, Wang Z, Li M, et al. A feasibility study of automating radiotherapy planning with large language
model agents. Phys Med Biol. 2025;70(7):075007. doi:10.1088/1361-6560/adbff1