On May 5, 2026, at its annual meeting in Washington, DC, the American College of Radiology Council approved the ACR-SIIM Practice Parameter for Imaging Artificial Intelligence and announced that the ACR Data Science Institute had published, in the Journal of the American College of Radiology, the technical framework for Assess-AI — a national quality registry that the ACR describes as the world’s first AI quality registry for medical imaging. The registry had first launched operationally on November 18, 2024; the May 2026 announcement and the peer-reviewed publication formalized and documented an infrastructure that had by then been running for eighteen months.

The significance of Assess-AI for veterinary medicine is not that it monitors veterinary AI. It does not. Assess-AI is a human-medicine system, built by the human-medicine radiology profession, monitoring human-medicine imaging AI. The word “veterinary” does not appear in its technical framework. Its use cases are human conditions — intracranial hemorrhage, pulmonary embolism, pneumothorax, large vessel occlusion. Its regulatory context is the U.S. Food and Drug Administration’s authorization framework for AI-enabled medical devices, a framework that does not apply to veterinary AI at all.

The significance is the contrast. Assess-AI is documentary proof of what a radiology profession does when it takes seriously the proposition that clinical AI must be monitored after deployment. It built a national registry. It built a data-ingestion platform. It built analytics dashboards and benchmarking infrastructure and local case-review tools. It published the technical framework in a peer-reviewed journal. It approved a formal practice parameter governing how AI should be selected, evaluated before deployment, and monitored after. It created a recognized-facility designation for institutions that comply. The human-medicine radiology profession, having concluded that deployed AI requires structured oversight, built the structures.

The veterinary radiology profession reached a materially similar conclusion. In 2025, the American College of Veterinary Radiology and the European College of Veterinary Diagnostic Imaging jointly issued a position statement on artificial intelligence in veterinary diagnostic imaging and radiation oncology, published in the Journal of the American Veterinary Medical Association. That statement concluded that commercially available veterinary AI products do not currently meet the transparency and validation standards the profession requires, and it called for rigorous peer-reviewed research, unbiased third-party evaluation, and board-certified veterinary radiologists remaining in the diagnostic loop. The veterinary profession, through its two principal radiology specialty bodies, formally diagnosed the same problem the human side diagnosed: deployed clinical AI cannot be presumed to perform as marketed, and requires rigorous, independent oversight.

Then the veterinary profession built nothing. There is no veterinary Assess-AI. There is no veterinary national AI registry. There is no veterinary equivalent of the ACR-SIIM Practice Parameter, no veterinary recognized-facility designation, no veterinary post-deployment monitoring infrastructure of any kind. The diagnosis was issued and the prescription was written. The prescription was never filled.

This article documents that contrast in detail: what the ACR built and why, what the parallel infrastructure in the United Kingdom looks like, what the veterinary profession’s own position statement said, and what the structural consequences are of the veterinary profession having reached the human side’s conclusion without taking the human side’s action. It is a reference investigation, sourced to the ACR’s peer-reviewed technical framework, the ACR’s own announcement of the practice parameter, the Royal College of Radiologists’ published monitoring guidance, and the ACVR/ECVDI position statement. It names commercial veterinary AI products as examples where this publication’s prior coverage has already documented them, and it cross-references that prior coverage throughout.

The Problem Assess-AI Was Built to Solve

The ACR’s published technical framework is direct about why the registry exists. The framework states that, although the number of FDA-authorized AI products has grown rapidly and imaging AI is now integrated into routine clinical workflows at a substantial proportion of healthcare facilities, “many institutions report that real-world AI performance often falls short of the metrics demonstrated during the authorization process,” and that this shortfall limits clinician trust and increases downstream clinical workload.

The framework identifies the specific mechanisms. AI models are, in the framework’s words, “notoriously fragile and prone to issues like performance degradation when local data is different from training data, including variations in scanner manufacturers or patient demographics.” Off-label use — applying a model outside its authorized indications — can produce worse-than-expected performance. And FDA-authorized models are generally locked: they cannot be retrained or fine-tuned without additional regulatory authorization, except within the scope of an authorized Predetermined Change Control Plan, which means a model that begins to drift cannot simply be quietly corrected. The framework concludes that detecting and addressing performance drift “is critical to maintaining clinical reliability,” and that “without mechanisms for robust monitoring, the risk of sub-optimal performance by these models remains a pressing concern.”

This is a remarkable set of admissions for a profession’s flagship body to publish in its own journal. It is the human-medicine radiology establishment stating, in peer-reviewed print, that FDA authorization is not a guarantee of real-world performance; that AI models degrade when they encounter data different from their training distribution; that the same model performs differently at different facilities and on different patient populations; and that none of this can be managed without infrastructure built specifically to detect it. Every one of those statements has a veterinary analog, and the veterinary analogs are in several respects more severe — a point this article returns to in detail below.

How the Registry Works

Assess-AI operates within the ACR’s National Radiology Data Registry, the established multi-registry infrastructure through which thousands of U.S. imaging facilities already submit quality data. Any U.S. imaging facility with an imaging AI model deployed in a real-world clinical setting is eligible to participate, and the ACR has noted that the more than 5,000 facilities already participating in at least one National Radiology Data Registry program are eligible to participate in Assess-AI under their existing agreements.

Participating facilities submit three categories of de-identified data through ACR Connect, the ACR’s locally-installed software platform that connects to a facility’s radiology information system, picture archiving and communication system, and electronic health record. The three categories are: the imaging AI outputs produced in the clinical workflow; the de-identified text of the radiology report for the exam; and the imaging study metadata drawn from DICOM headers, including patient age and sex. The framework notes that the registry does not currently collect source image pixel data or patient-specific health information — the de-identification is handled locally, before transmission, and only the study date of service is retained as protected health information.

What the registry does with that data is the analytically important part. It computes concordance — agreement — between the AI model’s output and a surrogate label extracted from the radiology report. Because re-reviewing every study against a fresh expert-annotated reference standard would not scale, the registry instead uses large language model prompting pipelines to extract, from the text of the report the interpreting radiologist already wrote, a structured label indicating whether the target finding was present or absent. The AI output and the report-derived surrogate label are then compared. When they agree, the case is concordant; when they disagree, it is discordant. Monthly concordance rates are computed and plotted over time, with confidence intervals, against registry benchmarks.

The ACR is careful — and this care is itself instructive — about what concordance does and does not mean. The framework states explicitly that the report-derived labeling method produces concordance that “represents alignment with the radiologist’s report (not equivalent to clinical accuracy).” It is not a pathology-confirmed ground truth. It is a measure of whether the AI agrees with the radiologist, not a measure of whether either the AI or the radiologist is correct. The framework acknowledges the limitations of this approach directly and at length: radiology reports contain hedging and diagnostic uncertainty, so surrogate labels extracted from them may themselves be uncertain; the large language models used for extraction are subject to drift and require ongoing validation; and incorporation bias is a concern, because if the radiologist viewed the AI output before writing the report, the report is no longer independent of the AI result.

This combination — a pragmatic, scalable surrogate paired with explicit, published acknowledgment of exactly what the surrogate cannot establish — is a model worth naming clearly. The ACR did not wait for a perfect monitoring methodology before building monitoring infrastructure. It built a good-enough system, documented its limitations honestly, and committed to improving it. The alternative — building nothing until a perfect system is available — is the de facto position of the commercial veterinary AI sector, and the ACR’s example demonstrates that the alternative is a choice, not a necessity.

What the Registry Found — and Why It Matters

The ACR’s published framework includes early results that, although presented as an infrastructure overview rather than a definitive performance evaluation, demonstrate exactly why post-deployment monitoring is necessary. At the time of the analysis, the registry held data from 16 participating facilities, representing 48,371 patients and 62,835 imaging studies.

For the intracranial hemorrhage use case, the ACR fit an exploratory mixed-effects logistic regression to examine what factors were associated with concordance. The result: patient sex, patient age, and scanner manufacturer were all significantly associated with AI concordance, each at a p-value below 0.0001. Increasing patient age was associated with lower concordance. Female patients showed modestly higher concordance than male patients. And the scanner manufacturer — the brand of the equipment producing the image — was significantly associated with whether the AI agreed with the radiologist. Beyond all of those measured factors, the framework reports that “the variability of the facility random intercept indicated meaningful between-facility differences in baseline performance not explained by patient age, sex, or scanner manufacturer.”

Read that finding plainly: the same category of AI model, monitored across real facilities, performed measurably differently depending on the patient’s age and sex, depending on which manufacturer’s scanner produced the image, and depending on the facility itself in ways that could not be fully explained even after accounting for those variables. This is precisely the performance variation the framework warned about — fragility to local data variation, divergence across deployment settings — and it was detectable only because the monitoring infrastructure existed to detect it. Without Assess-AI, every one of those facilities would have continued using its AI model with no way of knowing that performance varied this way.

The Governance Loop

Detecting a performance signal is only useful if something can be done about it. The Assess-AI framework describes what the ACR calls a complete governance loop. When a facility’s concordance for a given AI model drops — falling below the registry benchmark, or below a confidence interval, or below the facility’s own historical performance — the facility’s dashboard surfaces the signal. A local administrator can then use the Forensics App, a component of ACR Connect that runs locally at the facility, to securely re-identify the discordant cases and review them: the imaging study, the radiology report, the AI output, and the surrogate label, all together, so a reviewer can determine why the case was discordant. Structured feedback from that local review is returned to the registry.

The framework notes that the corrective actions a facility might take after such a review include updating scanner protocols, modifying data pipelines, re-training personnel, or contacting the AI model’s manufacturer. The loop is closed: monitoring detects the signal, local review investigates the cause, and corrective action addresses it — with the registry capturing the feedback so that the monitoring improves over time.

The ACR-SIIM Practice Parameter for Imaging AI, approved at the same May 2026 meeting, formalizes the surrounding governance expectations. Per the ACR’s announcement, the practice parameter applies to physicians, technologists, medical physicists, informatics teams, data scientists, and administrators who deploy or use AI in imaging workflows, and it covers AI tool selection, pre-deployment evaluation, ongoing performance monitoring, and patient privacy protection. Facilities that implement AI in accordance with the parameter can earn the ACR Recognized Center for Healthcare-AI designation. The registry is the measurement instrument; the practice parameter is the governance standard; the recognized-facility designation is the incentive. Together they constitute an integrated system.

The Royal College of Radiologists’ Parallel Infrastructure

The ACR is not alone, and that fact matters for understanding whether post-deployment AI monitoring is a fringe idea or an emerging standard. In the United Kingdom, the Royal College of Radiologists has published guidance titled “Post-deployment monitoring and safety reporting of AI medical imaging devices in clinical practice,” and has developed an RCR AI registry.

The RCR guidance draws a careful three-part distinction that is worth reproducing because it clarifies the conceptual landscape. Evaluation, in the RCR’s framework, is the assessment of an AI medical device before or at the point of deployment, to confirm expected performance. Auditing is the structured, periodic review of outputs against reference standards, to verify ongoing accuracy and compliance. Monitoring is the continuous or routine observation of operational performance and safety indicators, to detect change over time. The RCR treats all three as complementary but distinct activities that together form a complete oversight practice. The guidance recommends tracking AI-human disagreement rates, liaising with the vendor when AI performance changes so the model can be recalibrated or retrained, and sharing audit findings across National Health Service networks so that performance trends are spotted early across institutions rather than discovered in isolation.

The convergence is the point. Two of the most established radiology professional bodies in the world — one in the United States, one in the United Kingdom — independently reached the same conclusion: that deployed clinical AI requires structured, ongoing, post-deployment oversight, and that the profession itself must build the infrastructure to deliver it. They did not wait for regulators to mandate it. They did not wait for vendors to volunteer it. The professional bodies built it themselves, because the professional bodies concluded it was part of the responsible practice of radiology.

This transatlantic convergence forecloses one possible defense of the veterinary profession’s inaction. It is not possible to argue that post-deployment AI monitoring is an unproven, experimental, or premature idea that the veterinary profession is reasonably waiting to mature. It has matured. It is being implemented at national scale on two continents by the human-medicine radiology establishment. For veterinary medicine, post-deployment AI monitoring is not a frontier that has not yet been reached. It is a gap that has not been filled.

What the Veterinary Radiology Bodies Said in 2025

The veterinary radiology profession cannot claim it was unaware of the problem. In 2025, the American College of Veterinary Radiology and the European College of Veterinary Diagnostic Imaging — the two principal veterinary radiology specialty bodies, the direct counterparts of the ACR and the RCR — jointly issued a position statement on artificial intelligence in veterinary diagnostic imaging and radiation oncology, published in the Journal of the American Veterinary Medical Association.

The position statement’s conclusions were not equivocal. It concluded that commercially available veterinary AI products do not currently meet the transparency, validation, and evidence standards that the profession requires in order to rely on them. It called for rigorous peer-reviewed research validating veterinary AI products. It called for unbiased third-party evaluation — evaluation independent of the vendors selling the products, rather than vendor-funded or vendor-controlled validation. And it called for board-certified veterinary radiologists remaining in the diagnostic loop, rather than being displaced by AI systems operating without specialist oversight.

That is, in substance, the veterinary radiology profession’s two flagship specialty bodies stating formally that the commercial veterinary AI products currently being marketed to and adopted by veterinary practices have not been validated to the standard the profession’s own experts consider necessary — and calling for exactly the kind of rigorous, independent, ongoing evaluation that, on the human side, produced Assess-AI and the RCR registry.

This publication has documented the commercial veterinary AI landscape that the ACVR/ECVDI statement was responding to. The Phantom Radiologists series examined the foundational training-data claims of commercial veterinary AI vendors including SignalPET, Vetology, and Antech RapidRead, and calculated that the “trained on millions of specialist-reviewed cases” marketing claims cannot be reconciled with the documented size of the veterinary radiology specialist workforce. The analysis of AI primary reads and the veterinary regulatory gap documented that veterinary AI operates outside the FDA authorization framework, the state practice act frameworks, and the reimbursement-gatekeeping mechanisms that constrain AI on the human side. The examination of institutional inaction by the ACVR, AVMA, and AAHA documented the pattern of professional bodies issuing statements without building enforcement or oversight mechanisms. The ACVR/ECVDI position statement is the veterinary radiology profession’s formal acknowledgment of the problem. Assess-AI is what the acknowledgment looks like when it is followed by construction.

The Diagnosis Without the Cure

Place the two professions side by side. Both have flagship radiology specialty bodies. Both of those bodies concluded, in formal published statements, that deployed clinical AI cannot be presumed to perform as marketed and requires rigorous, independent, ongoing evaluation. On the human side, that conclusion was followed by the construction of Assess-AI, the ACR Connect data platform, the DART analytics platform, the Forensics App, the ACR-SIIM Practice Parameter, and the ARCH-AI recognized-facility designation — an integrated, national, peer-reviewed-and-documented oversight infrastructure. On the veterinary side, the conclusion was followed by nothing of the kind.

The veterinary profession issued the diagnosis and did not build the cure. And the veterinary diagnosis, properly understood, describes a more dangerous condition than the human one — because the veterinary AI products the ACVR/ECVDI statement was evaluating were never subject to the FDA pre-market authorization process in the first place. On the human side, the monitoring infrastructure is a second layer of assurance on top of a regulatory floor. On the veterinary side, there is no regulatory floor, and now there is no monitoring layer either. The veterinary profession identified an unguarded gap and then left it unguarded.

The Conditions Transfer, and They Intensify

The ACR’s framework justified Assess-AI by reference to a specific set of structural conditions. Each of those conditions has a veterinary analog, and in several cases the veterinary version is materially worse. The point of this section is not abstract argument; it is to walk through the ACR’s own stated rationale, condition by condition, and show that the rationale applies to veterinary medicine with at least equal and often greater force.

Condition One: Real-World Performance Diverges From Pre-Market Claims

The ACR built Assess-AI partly because real-world AI performance often falls short of the metrics demonstrated during FDA authorization. On the human side, there is at least a pre-market metric to fall short of — the FDA authorization process produces a documented performance baseline against which real-world divergence can be measured. Veterinary AI has no such process. As this publication documented in its coverage of the veterinary AI regulatory gap, commercial veterinary AI products are not subject to FDA pre-market authorization. There is no required, standardized, independently reviewed pre-market performance baseline for a veterinary AI radiology product. The marketing claim is, very often, the only performance figure that exists. On the human side, monitoring detects divergence from a regulated baseline. On the veterinary side, there is no baseline — which means monitoring is not a second layer of assurance, it is the only possible layer.

Condition Two: Models Are Fragile to Local Data Variation

The ACR cited AI fragility to variation in scanner manufacturers and patient demographics. The veterinary version of this condition is more severe for a reason this publication has documented at length in the Phantom Radiologists infrastructure analysis: breed-specific anatomic variation. Human chest radiograph AI does not have to contend with the difference between a Dachshund’s thorax and a Great Dane’s thorax. Veterinary AI does. The canine species alone spans a range of body conformations more anatomically diverse than the entire human population. A veterinary AI model trained predominantly on one set of breeds and deployed on a practice population skewed toward different breeds is encountering exactly the local-data-variation fragility the ACR warned about — amplified by a degree of anatomic heterogeneity that has no human-medicine equivalent. The ACR’s own data showed that scanner manufacturer alone significantly affected concordance; veterinary AI faces scanner variation plus breed variation plus the species variation between canine and feline patients.

Condition Three: Off-Label Use Produces Worse-Than-Expected Performance

The ACR cited off-label use — applying a model outside its authorized indications — as a performance risk. On the human side, there are authorized indications: the FDA authorization specifies what the model is cleared to do, which makes off-label use at least definable. Veterinary AI products are not authorized for specific indications by any regulator. The concept of “off-label” presupposes a label. Where the boundaries of a veterinary AI product’s validated performance lie — which species, which breeds, which body regions, which presentations, which equipment — is, in the absence of regulatory definition and in the absence of the rigorous third-party evaluation the ACVR/ECVDI statement called for, frequently undefined. A veterinary practice cannot keep a model within its validated boundaries when those boundaries have not been established.

Condition Four: Locked Models Cannot Adapt to Drift

The ACR noted that FDA-authorized models are generally locked and cannot be retrained without further authorization, which makes drift detection critical because drift cannot simply be silently corrected. The veterinary situation is different in a way that cuts both directions and is, on balance, worse. Veterinary AI models are not subject to the FDA’s lock; a vendor can in principle update a veterinary model at any time. But that is not reassurance — it is the opposite. On the human side, a locked model at least cannot change without a documented, authorized process, so the facility knows the model is stable between authorized updates. A veterinary AI model can change underneath the practices using it, with no regulatory documentation, no version control visible to the customer, and — critically — no monitoring registry that would detect the resulting performance change. The human side has locked models plus drift monitoring. The veterinary side has unlocked models and no drift monitoring. The combination of mutability and invisibility is the worst of the available configurations.

Condition Five: The Same Model Performs Differently Across Facilities

The ACR’s exploratory analysis found meaningful between-facility performance differences not explained by patient age, sex, or scanner manufacturer. There is no structural reason to expect veterinary AI would be exempt from comparable between-facility variation — veterinary practices differ in equipment, in technique, in patient population, in case mix, in image quality. But on the veterinary side, that variation is entirely invisible. A veterinary practice using a commercial AI radiology product has no registry benchmark, no peer comparison, no facility-level concordance trend, and no mechanism by which it would ever learn that the product performs differently in its practice than the vendor’s marketing implied. The between-facility variation the ACR can now see and investigate is, in veterinary medicine, simply happening in the dark.

The Comparison, Summarized

Oversight Element	Human Radiology	Veterinary Radiology
Pre-market regulatory authorization of AI products	Yes — FDA AI/ML device authorization	No — no FDA pathway applies
National post-deployment AI monitoring registry	Yes — ACR Assess-AI (and RCR registry in the UK)	No — none announced or published
Formal professional-society AI practice parameter	Yes — ACR-SIIM Practice Parameter for Imaging AI	No equivalent practice parameter
Recognized-facility designation for AI governance	Yes — ACR ARCH-AI designation	No equivalent designation
Profession formally identified the AI validation problem	Yes	Yes — 2025 ACVR/ECVDI position statement
Profession built infrastructure addressing the problem	Yes	No
Facility-level concordance benchmarking against peers	Yes — via Assess-AI dashboards	No mechanism exists
Structured local discordant-case review workflow	Yes — ACR Forensics App	No equivalent
Peer-reviewed publication of the monitoring framework	Yes — JACR, 2026	No framework to publish

The table’s penultimate row is the one that should give the veterinary profession pause. On the question of whether the profession identified the AI validation problem, both columns read “yes” — the veterinary profession, through the ACVR/ECVDI position statement, did identify it. It is the final row, whether the profession built infrastructure addressing the problem, where the columns diverge. The veterinary profession’s failure is not a failure of awareness. It is a failure of action after awareness.

Nothing About the Architecture Is Human-Medicine-Specific

It would be a mistake to conclude from this article that a veterinary monitoring registry is impossible because veterinary medicine lacks some essential ingredient the human side possesses. The core architecture of Assess-AI is not human-medicine-specific. Consider its components in turn. De-identified data ingestion from practice imaging systems: veterinary practices run picture archiving and communication systems and practice management systems from which de-identified AI outputs, report text, and DICOM metadata could be drawn, exactly as on the human side. Normalization of AI outputs across vendors: the veterinary AI market has a finite, enumerable set of commercial vendors whose outputs could be mapped to standard values. Surrogate label extraction from report text using large language models: veterinary radiology reports are text documents from which a large language model can extract present/absent labels using precisely the prompting methodology the ACR describes. Central concordance computation, benchmarking dashboards, local case-review tools: none of these is biologically specific to humans.

The veterinary profession also has the institutional bodies that could host such a registry. The ACVR is an established specialty organization. The veterinary informatics community has technical capacity. The DICOM standard already accommodates veterinary terminology through a dedicated working group. The data infrastructure question for a veterinary monitoring registry is a question of assembly and resourcing, not of invention.

There are genuine veterinary-specific complications, and an honest account names them. The breed and species heterogeneity that makes veterinary AI fragile also makes veterinary benchmarking harder — a concordance benchmark may need to be stratified by species and potentially by breed group to be meaningful. The veterinary radiology specialist workforce that would adjudicate discordant cases is, as this publication has documented extensively, small relative to clinical demand. Veterinary practices are more fragmented and less likely to participate in existing national quality registries than the 5,000-plus facilities already in the ACR’s National Radiology Data Registry. These are real design challenges. But they are design challenges of the kind the ACR itself navigated — the ACR’s framework is candid that its own surrogate-labeling approach is imperfect and that its benchmarking does not yet adjust for case mix. The ACR did not let imperfection prevent construction. It built the imperfect-but-useful system and committed to improving it.

The barrier, then, is not technical feasibility. It is institutional will and resource commitment. The ACR chose to commit the resources. The Royal College of Radiologists chose to commit the resources. The veterinary professional bodies have not made the equivalent choice — and the 2025 ACVR/ECVDI position statement proves the choice has not been made for lack of awareness, because the awareness is on the record.

---

---