How an FDA-Cleared Chest X-Ray AI Actually Gets Cleared

Before discussing what veterinary AI vendors do or do not publish, it is useful to be specific about what their human-medicine counterparts are required to demonstrate. The FDA’s 510(k) and De Novo pathways for AI/ML-enabled medical imaging devices impose a documented, public, statistically rigorous validation requirement that is the precondition for the product being legally marketable in the United States. Failure to meet the requirement results in clearance denial. The requirement is documented, the methodology is reviewed, and the validation results are published as part of the public clearance record.

The mechanics, in plain language: an AI/ML-enabled radiology device manufacturer prepares a pre-market submission that includes the device’s intended use, the algorithm description, the training methodology, the validation testing protocol, the test set composition, the reference-standard derivation, the performance results, and a Predetermined Change Control Plan addressing how the algorithm may be modified post-clearance. The submission is reviewed by FDA’s Center for Devices and Radiological Health. The reviewers scrutinize the test set for adequacy and representativeness, the reference standard for appropriateness (typically pathology-confirmed or expert-adjudicated consensus), and the resulting performance statistics for whether they meet the standard required for the indicated use. Where the FDA determines the validation is inadequate, clearance is denied or delayed. Where clearance is granted, the device’s full performance summary becomes part of the public 510(k) database, retrievable at the FDA’s accessdata website.

According to the 2025 year-in-review analysis of FDA AI/ML clearances published by Innolitics, the FDA cleared 295 AI/ML-enabled medical devices in 2025, with radiological computer-assisted detection and diagnosis software (product code QIH) representing approximately 25% of all clearances — roughly 75 cleared products in a single year, in a single device category. Each of those 75 products produced documented validation testing, against a defined reference standard, with public performance data. The median time from submission to clearance in 2025 was 142 days, with substantial variability based on the complexity of the indicated use and the rigor of the supporting evidence.

Two specific examples illustrate what the validation evidence looks like in practice. Qure.ai’s qXR chest x-ray triage product, which received FDA 510(k) clearance in September 2023 for pneumothorax and pleural effusion triage, was supported by a multicenter publication co-authored by Dr. Subba R. Digumarthy, thoracic radiologist at Massachusetts General Hospital. The published performance data reported up to 96% sensitivity and 100% specificity for pneumothorax detection. Bering Limited’s BraveCX chest x-ray triage solution, which received 510(k) clearance announced in September 2025, reported 95% to 97% specificity and AUC-ROC values of 0.96 for pleural effusion and 0.98 for pneumothorax. The BraveCX validation was supported by training data from over 1,000,000 chest x-rays with 50,000 board-certified radiologist-labeled images — a scale that, importantly, was disclosed and methodologically documented rather than asserted as a marketing claim.

The documentation standard is unambiguous. An FDA-cleared chest x-ray AI product produces, for each indicated use, the per-indication sensitivity, specificity, false-positive rate, and false-negative rate, along with confidence intervals, on a defined test set against a defined reference standard, in a public document. A purchaser of the product can verify the cleared algorithm version, look up the validation statistics for that version, and have confidence that material changes to the algorithm were either approved by the FDA or fall within a documented Predetermined Change Control Plan. The performance characteristics of the product the clinician is using are, in other words, a matter of public record.

What the Veterinary Side Looks Like by Comparison

None of the structure described above applies to commercial veterinary AI radiology products. There is no FDA pre-market submission requirement. There is no required validation testing protocol. There is no required test-set composition or reference-standard derivation. There is no required public disclosure of sensitivity, specificity, false-positive, or false-negative rates by indicated use. There is no Predetermined Change Control Plan requirement, no version traceability requirement, and no public performance database equivalent to the FDA’s 510(k) database. Veterinary AI radiology products are software products, sold to veterinary clinics on the basis of vendor-published marketing materials, with whatever validation evidence the vendor chooses to publish voluntarily, in whatever form the vendor chooses, with whatever level of methodological rigor the vendor finds commercially convenient.

The 2025 ACVR/ECVDI joint position statement on artificial intelligence in veterinary diagnostic imaging was unequivocal on the consequence: “There is currently no commercially available product for diagnostic imaging that meets these standards.” The “these standards” the position statement refers to are the standards of transparency, validation methodology, and safety documentation that the position statement identifies as the necessary conditions for veterinary AI adoption — standards that the position statement explicitly draws from the human-medicine AI publication norms and the FDA regulatory framework. The veterinary specialty colleges have, in other words, told the veterinary profession in writing that no current commercial veterinary AI radiology product meets the standards of validation that human radiology AI is held to as a precondition of being marketable. The position statement is a matter of public record. The vendors continue to market their products. The clinics continue to be presented with marketing materials that the specialty colleges have already documented as inadequate to the validation standards of human medicine.

The evidence base on the veterinary side, where it exists, is sparse but informative. Two recent peer-reviewed publications have addressed commercial veterinary AI radiology product performance directly. Each has produced findings that, when compared to the FDA-cleared human chest x-ray AI performance numbers cited above, document a substantial validation gap.

The Joslyn Commentary: Sensitivity Collapse on Difficult Cases

The Joslyn, Faulkner, Ma, and Appleby commentary published in Frontiers in Veterinary Science in June 2025 was published in response to a head-to-head comparison study by Ndiaye and colleagues earlier the same year, which had reported that a commercial veterinary AI product (SignalRAY/SignalPET) “performs almost as well as the best veterinary radiologist in all settings of descriptive radiographic findings.” The Joslyn commentary did not dispute the headline accuracy figure on the test set used. It documented, instead, several methodological issues that materially affected what the headline figure actually meant.

The first issue documented in the commentary was sensitivity collapse on difficult cases. The Ndiaye study had reported, in its own data, that the AI’s sensitivity dropped to 0.444 on cases where the radiological finding was high in ambiguity — that is, on the cases where specialist consultation provides the most clinical value, and where AI assistance would be most consequential to clinical decision-making. A sensitivity of 0.444 means that, on difficult cases, the AI missed more than half of the abnormal findings present. The commentary observed that “the strength” the headline claim referred to “lies in confirming normal cases” rather than in detecting abnormal ones — a fundamentally different clinical role from what AI radiology products are typically marketed as performing.

The second issue documented was the class-imbalance problem in the test set. Where a test set is heavily weighted toward normal cases, a trivial classifier that simply predicts “normal” for every image scores high accuracy without doing any actual diagnostic work. The Joslyn commentary documented that on the Ndiaye test set, an “always normal” classifier could score approximately 84% accuracy by virtue of class imbalance alone — meaning the AI’s reported headline accuracy figure was meaningfully reduced once the contribution of the trivial baseline was accounted for. The full data are in the commentary; the implication for the marketed performance figure is that the figure overstates clinically meaningful diagnostic capability.

The third issue documented was version traceability. The commentary explicitly noted that the commercial AI software was “continuously updated and does not have version numbers” — that the software the Ndiaye study evaluated was not necessarily the same software that a clinic adopting the product after publication would actually be using, and that the validation evidence therefore did not pin down a specific algorithm version against which clinic-side performance could be compared. This is precisely the regulatory failure mode that the FDA’s Predetermined Change Control Plan framework was designed to prevent in human medical AI. In veterinary AI, the failure mode is documented in the peer-reviewed literature and remains uncorrected in commercial deployment.

The fourth issue documented was the absence of CLAIM checklist adherence. The CLAIM checklist (Checklist for Artificial Intelligence in Medical Imaging), published in Radiology: Artificial Intelligence in 2020, is the 42-item documentation standard that peer-reviewed human medical imaging journals use to evaluate AI manuscripts. The Joslyn commentary documented that the original Ndiaye study did not consistently adhere to CLAIM standards, and that the proprietary nature of the commercial AI software was cited as justification for not disclosing the methodological details that CLAIM requires. The implication: the validation evidence published in support of commercial veterinary AI products does not currently meet the documentation standards that human medical imaging AI publication has required for six years.

The Ma 2026 External Validation Study: Sensitivity 71% to 90%

The Ma, Faulkner, Stander, Raisis, and Joslyn pilot study published in JAVMA in 2026 — the most recent peer-reviewed external validation work on commercial veterinary AI radiology products as of this article’s preparation — extended the methodological pattern documented in the 2025 commentary into a multi-platform comparison against pathology-confirmed reference standards. The study tested commercial veterinary AI services (including Caliber Vet AI from SK Telecom) against 53 general practice-sourced canine abdominal radiographs with confirmed diagnoses by definitive diagnostic tests (ultrasound, surgery, biopsy, CT, response to therapy, or follow-up imaging). The reported sensitivity range across the tested AI platforms was 71% to 90%, with documented “deficiencies in interpretation of general practice-sourced canine abdominal radiographs.”

To put the 71% to 90% sensitivity range in context: the FDA-cleared Qure.ai chest x-ray triage product reported up to 96% sensitivity for pneumothorax. The FDA-cleared Bering BraveCX product reported AUC-ROC of 0.96 to 0.98 for its indicated uses. The commercial veterinary AI products tested in the Ma 2026 pilot study, on pathology-confirmed canine abdominal cases, reported sensitivity ranges that would, in many cases, fall well below what FDA reviewers consider adequate performance for clearance in the human-medicine equivalent product category. The study explicitly documents that the AI services showed deficiencies in interpretation that are clinically meaningful — that is, that the products’ performance on the test population (general-practice-sourced canine abdominal radiographs, the actual clinical use case the products are deployed for) was not at the level the marketing implies.

The Ma 2026 study is, importantly, a pilot study with a sample size of 53 cases. Larger external validation studies, conducted with documented test-set composition and pathology-confirmed reference standards, are needed to establish definitive performance characteristics across the breadth of the products’ marketed indications. The pilot study’s importance is not the precise sensitivity figure on a 53-case sample; it is the existence of the work itself, conducted by independent academic investigators against pathology-confirmed reference standards on real general-practice clinical material, in a way that no vendor-published marketing material has matched. The pilot study demonstrates that external validation of commercial veterinary AI is possible to perform when investigators are willing to do the work, and that when the work is performed, the results are not at the level the vendor marketing materials suggest.

The Validation-Statistics Comparison Table

The asymmetry in validation evidence between FDA-cleared human chest x-ray AI and commercial veterinary AI radiology products is documented in the table below. The human-side figures are drawn from the FDA’s public 510(k) database and the published clearance announcements; the veterinary-side figures are drawn from the only peer-reviewed external validation literature available as of this article’s preparation. Where vendor-published claims exist on the veterinary side, they are characterized as such.

Product / Study	Reported sensitivity	Reported specificity / AUC	Reference standard	Validation source
Qure.ai qXR-PTX (FDA-cleared, human)	Up to 96% (pneumothorax)	100% specificity	Multicenter, expert-adjudicated	FDA 510(k); peer-reviewed, MGH co-author
Bering BraveCX (FDA-cleared, human)	Reported with AUC 0.96–0.98	95%–97% specificity	Multicenter, board-certified labeled	FDA 510(k); 1M+ training corpus disclosed
SignalPET / SignalRAY (Ndiaye 2025 evaluation)	0.444 in difficult/ambiguous cases (Joslyn 2025 commentary)	“Higher specificity than radiologists” but on class-imbalanced test set	Radiologist consensus, not pathology	Frontiers Vet Sci 2025; commentary documents issues
Multiple commercial vet AI services (Ma 2026 pilot)	71% to 90%	Per-platform variability; deficiencies documented	Pathology-confirmed (53 dogs)	JAVMA 2026 external validation pilot
SignalPET / Vetology / Antech (vendor marketing)	Not consistently published per indication	Not consistently published per indication	Not disclosed at CLAIM-checklist standard	Vendor marketing materials only

The pattern in the table is the central finding of Part A. The FDA-cleared human chest x-ray AI products operate to a published sensitivity-specificity standard, with public documentation, with version traceability. The commercial veterinary AI radiology products, where they have been independently evaluated at all, have produced sensitivity figures that range from 0.444 on difficult cases to 71% to 90% on pathology-confirmed cases — figures that would not pass FDA review for the human-medicine equivalent product category, and that the vendors do not publish in any form comparable to the FDA’s 510(k) database. The vendor-published marketing claims, where they exist, are not produced to the documentation standards that human medical AI publication has required for six years, per the explicit conclusions of the 2025 ACVR/ECVDI position statement and the 2025 Joslyn et al. commentary.

The FDA-Side Structural Separation: Why Vendor and Provider Are Different Companies, By Design

One feature of the FDA-cleared human radiology AI category that does not get discussed enough, but that turns out to matter enormously for understanding the veterinary side, is that the AI software vendor and the radiology service provider are, almost without exception, different companies. This is not an accident. It is a structural feature of how the category is regulated, financed, and economically organized in U.S. human medicine, and it produces the alignment of incentives that makes FDA-cleared AI radiology products generally trustworthy when they are used.

Qure.ai is a software company headquartered in New York and Mumbai. It develops AI algorithms for radiology and licenses them to hospitals, imaging centers, and radiology groups. Qure.ai does not employ U.S. radiologists who bill insurance for radiology reads. It does not own hospital systems. It does not operate teleradiology services in competition with the radiology groups that buy its software. Its revenue comes from software licensing — not from displacing the radiologists who would otherwise read the studies. The company that buys Qure.ai’s qXR software, deploys it in a hospital ICU, and uses it to triage chest x-rays for pneumothorax is a different company entirely — typically a hospital, an integrated delivery network, or a radiology group that bills insurance and Medicare for the radiology read itself.

Bering Limited, the manufacturer of FDA-cleared BraveCX, is a software company headquartered in London. It does not own U.S. radiology practices. Aidoc is an Israeli software company that licenses its FDA-cleared AI portfolio to U.S. health systems. Each of the approximately 75 FDA-cleared radiological CAD products that received clearance in 2025 represents the same structural pattern: a software company developing the AI, and separate radiology service providers buying and deploying it. The economic boundary between the two roles is the structural feature that aligns incentives.

The implication of the separation matters. The AI vendor’s revenue depends on producing software that radiology service providers want to buy. If the software is inaccurate, radiology service providers stop buying it, because the radiologists at the radiology service providers are the ones who bear malpractice exposure when the AI is wrong. The radiology service provider’s revenue depends on producing accurate reads to maintain referrals from clinicians and avoid malpractice exposure. The radiology service provider does not capture additional revenue when the AI vendor’s software displaces the work of the radiology service provider’s own radiologists, because the radiology service provider is the radiology service provider. The two companies are economically separate, and their incentives are aligned by being separate. Neither company captures revenue from the other’s primary activity. Neither company has a direct financial incentive to deploy AI that performs below the standard the radiologists at the radiology service provider would tolerate.

This separation is the precondition for the FDA-cleared category to function the way it does. It is also, as Part B will document, the precise feature that does not exist on the veterinary side, where a single corporate parent owns both the AI software and the radiology service that competes with the AI software’s external market. The conflict of interest that the human-side structure prevents by design is, on the veterinary side, the central design feature.

The Mars Petcare Consolidation: Documented Corporate History

To understand why commercial veterinary AI radiology software exists in the form it does, it is necessary to understand who owns the companies that produce it and how the corporate consolidation of veterinary services in the United States has structured the financial incentives that shape product design choices. The history is well-documented in SEC filings, corporate press releases, and trade-press reporting, all of which constitute primary-source evidence not subject to factual dispute.

Mars Incorporated, the privately held global conglomerate best known for confectionery and pet food brands (M&Ms, Snickers, Pedigree, Royal Canin, Whiskas), began acquiring U.S. veterinary services companies in earnest with the 2007 consolidation of Banfield Pet Hospital under its corporate umbrella. The Banfield network, a chain of veterinary practices originally established in 1955 and operated through partnership agreements with PetSmart retail locations, gave Mars an initial entry into the U.S. veterinary services market at the primary-care level. In 2015, Mars acquired BluePearl Veterinary Partners, the specialty referral hospital network that, prior to the acquisition, had been the largest privately held specialty veterinary hospital chain in the United States. The BluePearl acquisition gave Mars ownership of approximately 50 specialty hospitals at the time of the transaction, employing significant numbers of board-certified specialists across veterinary internal medicine, surgery, oncology, neurology, and diagnostic imaging.

The transformative transaction came in January 2017, when Mars announced the acquisition of VCA Inc. (NASDAQ: WOOF) for approximately $9.1 billion in an all-cash transaction including $1.4 billion in outstanding debt. The acquisition was completed in September 2017. The transaction is documented in VCA’s SEC filings (Form DEFA14A, January 9, 2017) and Mars’s own corporate announcements. The VCA acquisition gave Mars ownership of nearly 800 small-animal veterinary hospitals across the U.S. and Canada, Antech Diagnostics (a nationwide clinical laboratory and imaging services company operating in all 50 states and Canada), Sound Technologies (an animal diagnostic imaging equipment and PACS company), and Camp Bow Wow (a dog day-camp franchise that has since been spun off from the Mars Petcare structure). The combined Mars Petcare veterinary services portfolio, post-VCA acquisition, included Banfield, BluePearl, VCA, and Pet Partners — making Mars, by trade-press characterization, “the biggest vet provider in the country.”

The vertical integration that resulted is consequential for understanding the AI radiology product category. The same corporate parent now owns: (1) the primary-care clinical practices that generate the radiograph studies (Banfield, VCA), (2) the specialty referral hospitals that employ the board-certified veterinary radiologists who would otherwise be the consulting specialists external to the corporate structure (BluePearl), (3) the diagnostic laboratory and imaging services company that processes the studies (Antech Diagnostics), (4) the imaging equipment and PACS infrastructure that captures and stores the studies (Sound Technologies), and (5) the AI radiology product that purports to interpret the studies (Antech RapidRead). Each of these is a revenue-generating service touchpoint in the cumulative cost of producing a veterinary radiograph interpretation for a pet owner. The corporate parent captures revenue at each touchpoint. The structural significance of this fact for AI radiology product design is the subject of the next section.

The number of corporate-owned veterinary practices Mars controls post-consolidation has grown substantially since the 2017 acquisition closed. Per dvm360 reporting referenced in JAVMA News coverage and Yahoo Finance reporting from January 2025, “the number of VCA clinics has jumped 25% under Mars, while BluePearl has nearly doubled to about 100 hospitals” since the 2017 acquisition. The U.S. corporate-veterinary-medicine consolidation is broader than Mars (other major consolidators include private equity-backed groups like Thrive Pet Healthcare and a number of independent rolls-ups), but Mars is the largest single consolidator in the category.

The Conflict of Interest That Doesn’t Exist on the Human Side, and the Anticompetitive Tying That Replaces It

This is the structural argument the rest of Part B builds on, and it deserves to be stated as plainly as possible: on the human side, the AI software vendor and the radiology service provider are different companies. On the veterinary side, in the dominant corporate consolidator’s structure, they are the same company. The conflict of interest that the human-side category prevents by economic separation is, on the veterinary side, the central design feature of the largest corporate participant. Once this is recognized, the validation gap documented in Part A and the labor-math impossibility documented in Parts One and Two become not separate issues but a single integrated phenomenon, with a single structural cause.

Recall the human-side structure. Qure.ai develops AI and licenses it to hospitals and radiology groups. Bering Limited develops AI and licenses it to hospitals and radiology groups. Aidoc develops AI and licenses it to hospitals and radiology groups. The AI vendor’s customer is the radiology service provider. The two companies have aligned but separate incentives. The AI vendor wants its software to be accurate enough that radiology service providers want to buy it. The radiology service provider wants accurate reads to maintain referrals and avoid malpractice exposure. Neither company captures revenue when the other does its primary job poorly. The radiology service provider does not benefit financially when the AI vendor produces inaccurate software. The AI vendor does not benefit financially when the radiology service provider’s radiologists produce inaccurate reads. The boundary between the two companies is the structural feature that produces the alignment.

Now consider the Mars-Antech-RapidRead structure. The same corporate parent (Mars Petcare) owns the AI software (Antech RapidRead), the radiology service that operates in competition with independent veterinary teleradiology providers (Antech Diagnostics teleradiology, which sells specialist reads to clinics that do not own their own in-house radiologists), the imaging hardware and PACS infrastructure where the studies are captured (Sound Technologies), the specialty hospital network that employs the specialists whose reports become the training data (BluePearl), and the clinical practices that generate the radiograph studies the AI was trained on (Banfield, VCA). The AI software vendor and the radiology service provider — which are different companies on the human side — are the same company on the veterinary side. There is no economic boundary that produces aligned-but-separate incentives. There is a single corporate parent that captures revenue at every step.

The conflict of interest, stated as a methodological proposition, is this: the same company that develops the AI software, sells the AI subscription, employs the specialists whose work is used to train the AI, runs the radiology service that competes with independent teleradiologists, and processes the laboratory follow-up generated by the AI’s output, is also the company whose marketing materials describe the AI as “trained on millions of specialist-reviewed cases” without disclosing that the specialists are corporate-employed, the radiology service is corporate-owned, and the training-data harvesting was conducted within the corporate IT infrastructure. The methodological consequence is that there is no independent party in the structure to verify the marketing claims. The specialists who labeled the training data work for the same corporate parent that sells the AI. The radiology service that would, in a properly separated structure, be the customer auditing the AI’s accuracy is, in this structure, the same company selling the AI. The clinics that adopt the AI based on the marketing are the only external party in the chain — and they have no access to the underlying training methodology, validation evidence, or specialist-labor disclosures that would let them evaluate what they are buying.

The market-structure proposition is sharper. The vertical integration of AI vendor, radiology service provider, specialist labor pool, imaging hardware, and clinical practices into a single corporate parent is the textbook architecture of anticompetitive tying. Independent veterinary teleradiology providers — small businesses operated by individual board-certified veterinary radiologists or by independent group practices — are not competing on the merits of clinical accuracy against an equivalently structured competitor. They are competing against a vertically integrated corporate ecosystem that can underprice them on the AI subscription side (because the displaced specialist labor cost is captured internally rather than passed to the customer), tie the AI subscription to the corporate parent’s other diagnostic services (laboratory work to Antech, imaging to Sound, specialty referral to BluePearl), and use the bundled cost structure to displace independent teleradiology consumption regardless of comparative diagnostic accuracy.

The displacement is not driven by the AI’s accuracy relative to the independent teleradiologist. As Part A documented, the available external validation evidence on commercial veterinary AI suggests sensitivity ranges of 71% to 90% on pathology-confirmed cases (Ma 2026) and 0.444 on difficult cases (Joslyn 2025), figures that, on the human side, would not pass FDA review for the equivalent product category. The displacement is driven by the cost arbitrage of the vertically integrated corporate structure, in which the lower per-study cost of the AI subscription captures market share from the independent teleradiologist whose business model requires charging a per-study fee that fully compensates the specialist labor. Independent teleradiology providers are, in this structural sense, in the same competitive position as a small independent business competing against a vertically integrated corporate parent that can underprice on one component of the bundle while monetizing other components — the textbook market-structure concern that competition law has, in other industries, recognized as anticompetitive tying.

This article does not assert that any specific corporate action constitutes a per se violation of any specific competition law statute. It asserts, descriptively, that the structural conditions described above match the architecture that competition policy in other healthcare service categories has, in some cases, been used to address. The cumulative impact of the structure on the independent veterinary teleradiology profession — a profession the ACVR’s own leadership has described as facing a workforce crisis in which existing demand exceeds available supply — is that the workforce crisis is being addressed not by training more board-certified veterinary radiologists, but by deploying AI substitution within a vertically integrated corporate ecosystem that captures the displaced specialist labor cost as internal margin. The professional-supply problem and the corporate-monetization opportunity are, in this structure, the same phenomenon viewed from two different angles.

The Captive Specialist Labor Pool

The relevance of the corporate consolidation to the AI radiology product category is direct. When a clinic adopts Antech RapidRead in place of external specialist consultation, the revenue that would otherwise have flowed to an independent teleradiology service or to a local board-certified veterinary radiologist instead flows internally within the Mars Petcare corporate structure. The displaced specialist consultation revenue does not disappear from the system; it is captured by the corporate parent that has substituted AI for specialist review. Where the specialist consultation would have cost the clinic $85 to $250 per study (the typical range for U.S. veterinary teleradiology consultation, with the lower end applying to routine standard-turnaround reads and the upper end to stat reads, complex cases, MRI and CT consultations, and after-hours coverage), the AI subscription costs substantially less per study — in some cases under $5 per study, in others structured as flat-rate clinic subscriptions that distribute the AI cost across high-volume clinics at marginal per-study costs approaching zero.

The economic logic of the substitution is straightforward and is documented in the explicit business strategy of corporate veterinary services consolidation. Per the JAVMA News reporting on the 2017 Mars-VCA acquisition, “These corporate roll-ups, they exist to acquire to get bigger… companies see the potential for increasing practices’ profits through economies of scale.” The economy-of-scale logic, applied to the cost of specialist consultation specifically, is that AI substitution at the per-study cost differential between specialist consultation and AI subscription captures the difference for the corporate parent. At an estimated annual U.S. veterinary radiograph volume in the hundreds of millions of studies, even a modest per-study cost differential translates to substantial annual margin available to whichever corporate structure displaces the maximum specialist consultation volume with AI subscription volume.

The structural complication is the source of the training data the AI was built on. The AI radiology product, to be commercially marketable, requires training data labeled by board-certified specialists. Where does that labeled training data come from when the corporate parent owns the specialty hospital network that employs the specialists? The answer the corporate structure makes available is that specialists employed by BluePearl read studies as part of their normal clinical employment, with the resulting reports and labels generated within the corporate IT infrastructure that the corporate parent owns and operates. Whether those specialists explicitly consented to or were specifically compensated for the AI-training use of their clinical labor — beyond their regular salary as BluePearl-employed clinicians — is a matter the corporate parent has not publicly disclosed in any form this publication has reviewed.

The result, structurally, is that the “specialist-reviewed training corpus” claim that Antech RapidRead’s marketing makes is descriptive of clinical reports generated by Mars-employed specialists in the course of their normal Mars-paid clinical work, and harvested by the corporate parent for downstream AI training use. This is a different proposition from external specialist labeling commissioned and paid for as a discrete labeling project with documented methodology, and a different proposition from peer-reviewed AI training datasets where the labeling methodology is independently documented and reviewed. The marketing claim is technically accurate — board-certified veterinary radiologists were involved in producing the labels, in the sense that they wrote the reports the labels were derived from — but it is not equivalent to what the same claim would mean in the human-medicine FDA-cleared AI training context, where labeling is a discrete documented project with explicit methodology and the specialists involved are typically external to the device manufacturer.

The Cost-Reduction-At-Expense-Of-Quality Dynamic

Once the corporate revenue model is mapped, the validation gap documented in Part A becomes structurally explicable. An AI radiology product that displaces specialist consultation captures margin for the corporate parent whether or not the AI’s diagnostic accuracy matches that of the displaced specialist. The misdiagnosis cost — false negatives that miss disease, false positives that drive unnecessary further workup — is borne by the practicing veterinarian who relied on the AI output and by the patient whose care was directed by it. The corporate parent that sold the AI subscription does not bear the misdiagnosis cost. The financial incentive structure therefore favors AI deployment regardless of the AI’s accuracy relative to the displaced specialist.

This is not a theoretical observation. It is the structural logic of any service substitution where the substituting product is cheaper than the substituted service and the quality differential is borne by a third party. In the human medical AI category, the FDA pre-market clearance requirement is the regulatory mechanism that prevents the substitution from occurring without documented evidence that the substitute performs adequately. The pre-market clearance requirement, applied to chest x-ray AI products like Qure.ai’s qXR or Bering’s BraveCX, requires the device manufacturer to demonstrate, to FDA’s satisfaction, that the AI performs at the level its indicated use requires before the product can legally be marketed. Where the demonstration cannot be made, the product cannot be marketed. The clinical risk of inadequate AI substitution is, in effect, internalized by the device manufacturer, who bears the regulatory cost of demonstrating adequate performance.

In the veterinary AI category, no equivalent regulatory mechanism exists. The clinical risk of inadequate AI substitution is, structurally, externalized — borne by the practicing veterinarian, the clinic, and the patient, rather than by the AI vendor or the corporate parent. The financial incentive to deploy AI substitution is not constrained by a corresponding requirement to demonstrate that the AI substitution performs at an adequate level for the indicated use. The category therefore evolves to maximize the financial incentive (lower cost per study) without a corresponding evolution toward demonstrated diagnostic accuracy (validation evidence at the level human-side AI is required to produce).

The 2025 ACVR/ECVDI position statement on AI was explicit on the structural incentive misalignment: AI products sold into veterinary clinical workflows carry the potential for clinical harm to patients, and the standards required to mitigate that potential — transparency about training data, validation against pathology-confirmed reference standards, version traceability, and published per-indication performance data — are not currently met by any commercially available product. The position statement does not characterize the absence of these standards as the result of inadequate effort by the vendors. It characterizes it as the absence of a regulatory framework that requires the standards to be met as a precondition of commercial sale. The category is structured to permit, and therefore to produce, commercial products that do not meet the validation standards human medicine requires.

Who Bears the Cost When the AI Is Wrong

The clinical and economic costs of false positives and false negatives in commercial veterinary AI radiology output fall predictably across stakeholders, and the predictability of where they fall is itself part of the structural argument. A false negative — the AI returning a “normal” result on a study that actually contains a clinically significant abnormality — produces clinical risk borne by the patient (delayed diagnosis, missed treatment opportunity, potentially preventable adverse outcome) and by the practicing veterinarian (responsibility for the missed diagnosis, exposure to malpractice claims, reputational harm). The AI vendor and the corporate parent that sold the subscription do not bear the false-negative cost. They have, in some cases, contractual disclaimers in their terms of service that explicitly limit their liability for clinical outcomes — a structure broadly similar to the IDEXX clinical-content disclaimer documented in our companion analysis of IDEXX’s contract terms.

A false positive — the AI returning a “abnormal” result on a study that does not contain the indicated abnormality — produces different costs but they fall on the same stakeholders. The patient receives unnecessary further workup, additional imaging studies, possibly invasive diagnostic procedures, and the financial cost of the unnecessary care. The practicing veterinarian carries the burden of explaining the result to the client, ordering the follow-up workup, and managing the clinical disposition of a finding the AI generated that may not exist. The AI vendor and the corporate parent that sold the subscription, again, do not bear the false-positive cost. They benefit, in fact, from a portion of the false-positive workflow, where the follow-up imaging or laboratory work is performed within the same corporate structure (Antech Diagnostics for laboratory work, Sound Technologies for imaging equipment, BluePearl referral for specialist consultation that becomes necessary precisely because the AI output requires interpretation).

The structural irony of the false-positive case is that, in a vertically integrated corporate ecosystem where the AI vendor, the diagnostic laboratory, the imaging hardware, and the specialty referral are all owned by the same parent, false-positive output by the AI generates additional internal corporate revenue at every downstream service touchpoint. This is not an allegation of intentional design; it is a description of the financial-flow consequence of vertical integration when the AI output is wrong in the false-positive direction. Where the AI is wrong in the false-negative direction, the clinical cost falls outside the corporate structure (on the patient and the clinic). Where the AI is wrong in the false-positive direction, additional revenue flows inside the corporate structure (through the laboratory, imaging, and specialty referral subsidiaries). The financial-incentive asymmetry between false-negative and false-positive errors is, in a vertically integrated ecosystem, not symmetric.

The IDEXX Vertical Integration Comparison

The Mars-VCA-Antech-BluePearl consolidation is the most prominent example of vertical integration affecting the veterinary AI category, but it is not unique. IDEXX Laboratories — the publicly traded clinical laboratory and reference diagnostic services company (NASDAQ: IDXX) — has pursued its own vertical integration strategy through the rollout of in-clinic laboratory equipment, reference laboratory services, software services (Cornerstone, ezyVet practice management), and AI-enabled diagnostic features. IDEXX’s commercial structure, contract terms, and clinical-content disclaimer language is documented in this publication’s prior coverage of the IDEXX contract terms, including IDEXX’s Section 9.2.3 disclaimer that explicitly states the clinical content provided “does not constitute an opinion, medical advice, diagnosis” — precisely the kind of liability-shifting contractual structure that places clinical risk on the practicing veterinarian rather than on the company that sold the diagnostic service.

The IDEXX comparison is useful because it documents that the corporate-consolidation revenue model in veterinary services is not specific to Mars. The same structural dynamics — vertical integration of clinical practice, diagnostic services, and AI-enabled product features, with contractual arrangements that shift clinical risk from the vendor to the practicing veterinarian — exist across multiple major veterinary services consolidators. The revenue model is structural to the corporate-veterinary-medicine consolidation pattern, not idiosyncratic to any one company’s business strategy.

What an Honest Disclosure Standard Would Look Like

The conclusion that the validation gap and the revenue-model dynamics of commercial veterinary AI radiology are structural rather than incidental does not lead to the recommendation that the products should be banned, that the corporate consolidation should be reversed, or that any specific company should be subject to regulatory enforcement action. It leads to a more limited recommendation: that the disclosure standards human medical AI is held to as a precondition of commercial sale should also be met by veterinary medical AI as a precondition of clinical adoption, and that, in the absence of an FDA-equivalent regulatory mechanism in the veterinary space, the burden of demanding the disclosure rests with the clinics that adopt the products and the specialty colleges that set the standards of practice the clinics are expected to follow.

The disclosure standard that would address the validation gap is documented in the CLAIM checklist: per-indication sensitivity, specificity, false-positive rate, and false-negative rate, on a defined test set, against a defined reference standard, with the test-set composition and reference-standard derivation documented at a level of detail sufficient for an independent reviewer to evaluate. Algorithm version traceability, specifying the precise algorithm version that any cited validation study addressed and the change-control policy governing how the deployed version differs from the validated version. Training-data methodology documentation, specifying the percentage of labels generated by board-certified specialists, by residents, by general-practice clinicians, by automated NLP extraction, and by AI-generated pseudo-labels.

The disclosure standard that would address the revenue-model dynamics is the kind of corporate-relationship transparency that regulated industries are routinely required to produce: explicit identification of when the company selling the AI product, the company processing the laboratory work, the company supplying the imaging hardware, and the company providing specialist referral are all owned by the same corporate parent. Disclosure of contractual arrangements with the practicing veterinarians whose specialist labor was used to label training data, including consent and compensation terms. Clarity, in the marketing of the AI product, about what the “specialist-reviewed” training corpus claim actually means in terms of how the specialists were involved and what they were compensated for.

None of these disclosures is exotic or unprecedented. All of them are routine in regulated industries where conflicts of interest, vertical-integration dynamics, and validation-evidence asymmetries have been recognized as creating risk to patients and to professional standards of care. The structural failure documented in this series is not the absence of any individual disclosure. It is the absence of a regulatory mechanism that requires the disclosures to be made as a precondition of commercial sale, and the structural absence of independent validation infrastructure that would substitute for the missing regulatory requirement.

The Audit Mechanism That Doesn’t Exist: How Would Anyone Even Know?

The final structural observation in this article is the one that brings the validation gap, the corporate consolidation, the conflict of interest, and the anticompetitive tying together into a single integrated point: there is no external audit mechanism that would allow an outside party to verify what fraction of nominally specialist-reviewed veterinary radiograph reads in the vertically integrated corporate ecosystem actually involved meaningful specialist review of each individual case. The accountability blackout this absence creates is, in some ways, the most consequential structural failure in the entire category.

Consider how the equivalent verification works on the human side. A radiology service provider that bills Medicare for a chest x-ray read submits a claim that is auditable by CMS, by the hospital’s quality assurance committee, by the radiology group’s own internal compliance review, by the specialist’s malpractice carrier, and by external auditors retained by any of these parties. The radiologist who signs the report is identifiable by national provider identifier (NPI), and the specific report is linked to a specific patient encounter, a specific date, a specific imaging study, and a specific radiologist. Where AI assistance is used in producing the read, the role of the AI is documented in the radiologist’s report and is reviewable. The radiologist’s signature is a personal professional attestation that the radiologist actually reviewed the study and reached the conclusion the report describes. False attestation — signing reports the radiologist did not actually review — is professionally and legally consequential, with documented enforcement mechanisms across all of CMS, state medical boards, hospital credentialing committees, and malpractice litigation.

The veterinary side has none of these audit mechanisms. There is no Medicare to bill, no CMS to audit, no hospital quality assurance committee external to the corporate parent to review the work, no malpractice carrier in most cases, no equivalent of state medical board enforcement of false attestation in radiology specifically, and no public report database where external researchers could examine what fraction of nominally specialist-reviewed reads actually contain documented specialist analysis versus stamped sign-off on AI output. A specialist working inside the Mars Petcare corporate ecosystem who signs a report that the AI generated has no external party verifying that the specialist actually reviewed each case independently. The specialist’s employment relationship is with the corporate parent. The corporate parent is also the company selling the AI subscription that may have produced the report. The corporate parent is also the company operating the radiology service that the report nominally represents. There is no independent external party in the chain to ask the question: did the specialist actually look at this study, or did the specialist’s signature attach to the AI output without meaningful human review?

This article does not allege that any specific specialist has signed reports without reviewing them. It documents the structural absence of any mechanism that would allow an outside party to evaluate the question — and the structural significance of that absence in a corporate ecosystem where the financial incentive favors maximum throughput, where the AI subscription revenue depends on the AI’s output being treated as clinically useful, and where the specialist labor cost is the largest single variable expense the corporate parent could potentially reduce by encouraging specialists to sign-off on AI-generated reports more efficiently. The accountability blackout is not an allegation. It is a description of the structural conditions under which an allegation would be impossible to substantiate or refute, because no mechanism exists to gather the evidence that would settle the question.

The implications for clinics and pet owners are severe. A clinic adopting Antech RapidRead, or any commercially marketed veterinary AI radiology product that bundles AI output with nominal specialist review, has no way to know, on any individual case, whether the specialist whose name appears on the report actually examined the study independently of the AI’s output, or whether the specialist’s sign-off was, in effect, a rubber stamp on whatever the AI produced. The clinic has no way to compare the rate of meaningful specialist review across vendors, across products, across corporate structures. The pet owner who is paying for the diagnostic interpretation of their pet’s radiograph has no way to know what they are buying. The professional integrity of the veterinary radiology specialty as an external check on diagnostic accuracy depends on the assumption that specialist sign-off represents specialist review. The structural absence of any mechanism to verify that assumption is, in the corporate consolidation environment documented in the previous sections, a substantial gap in the integrity infrastructure of the category.

The five-disclosure standard described in the previous section — training-data methodology, version traceability, validation statistics, corporate-relationship transparency — is necessary but not sufficient to address this gap. What would be sufficient is the kind of audit infrastructure that human medicine has built over decades: a publicly accessible record of which specialist signed which report, on what date, for what study, with what role for AI in the interpretation, retrievable by independent researchers and regulators in a manner that does not depend on the corporate parent’s voluntary cooperation. The veterinary profession has not built this infrastructure. The corporate consolidators have no reason to build it on their own initiative, because the absence of the infrastructure is, in the structural argument this article documents, the precise feature that makes the corporate consolidation business model work.

---

---