Artificial intelligence can flag findings on a veterinary radiograph, but it cannot replace a diagnostic read. The American College of Veterinary Radiology and the European College of Veterinary Diagnostic Imaging stated in 2025 that no commercially available veterinary imaging AI product currently meets the profession's standards for transparency, validation, and safety, and that a qualified veterinarian, preferably a board-certified radiologist, should always review AI output.
Ask a room of practice owners whether AI can read an X-ray and you will get three answers: yes, no, and it depends what you mean by read. The honest answer sits closer to the third. Software can detect patterns on a radiograph and return a list of suspected findings in seconds. What it cannot do is take clinical responsibility for that interpretation, account for the patient in front of you, or tell you what it missed. That gap is not a temporary limitation waiting on a better model. It is the reason the profession's own radiology colleges have drawn a line, in writing, about how these tools should be used. This post covers what the line says and what it means for how you set up imaging in your practice.
Key Facts
- The ACVR and ECVDI published a joint position statement on artificial intelligence in the Journal of the American Veterinary Medical Association on March 19, 2025 (Appleby RB, DiFazio M, Cassel N, Hennessey R, Basran PS. JAVMA. 2025;263(6):773-776, doi:10.2460/javma.25.01.0027).
- The statement concluded that no commercially available veterinary AI product for diagnostic imaging meets the required standards for transparency, validation, or safety.
- The colleges call for a qualified veterinary professional to remain in the loop on every AI output, preferably a board-certified radiologist or radiation oncologist.
- The ACVR represents more than 600 accredited veterinary radiologists and radiation oncologists.
- The statement encourages veterinarians to disclose AI use to pet owners and to offer alternative diagnostic options.
What can AI actually do with a veterinary radiograph today?
Current tools are pattern detectors. Given a thoracic radiograph, a model trained on labelled images can return probability scores for findings such as cardiomegaly, pulmonary patterns, or effusion. Some products present this as a draft report. Others present it as a set of flags on the image.
Both are the same thing underneath. The software is telling you what it has seen before in images that looked similar. It is not reasoning about signalment, history, or the reason the patient came in. It has no access to the physical exam you just performed or the bloodwork sitting in the record.
That makes these tools useful for a narrow job: surfacing something worth a second look on a study you might otherwise have called normal. It does not make them a read.
What did the ACVR and ECVDI actually say?
The two colleges issued a joint position statement in 2025 setting out how AI should be developed and deployed in veterinary diagnostic imaging and radiation oncology.
The core requirements are specific. AI systems should follow good machine learning practices, meaning transparency about how they were built, mechanisms for reporting errors, and clinical experts involved throughout development rather than consulted at the end. They should handle patient data securely and be monitored after deployment, not validated once and left alone.
The finding that matters most for a practice evaluating vendors is blunter. The colleges assessed the commercial market and concluded that no available veterinary imaging AI product currently meets those standards. Not that the technology cannot get there. That as of the statement, nothing on sale had cleared the bar.
They also called for peer-reviewed research, independent third-party evaluation, and published benchmarks, none of which the veterinary market currently has in any organised form.
Why does "veterinarian in the loop" matter in practice?
The phrase sounds like a formality. It is not.
An AI flag on a radiograph is a hypothesis with no accountability attached. If the model misses a lesion, the model does not carry that. You do. The colleges' position is that a qualified professional, ideally board-certified in radiology, has to interpret the output rather than pass it along.
There is a practical version of this too. Automation bias is well documented in human radiology: when software says normal, readers become measurably less likely to find the abnormality themselves. A tool that returns a confident clean result on a study with a subtle lesion is not neutral. It has actively made the miss more likely.
Keeping a specialist in the loop is not about distrusting software. It is about who is answerable for the interpretation when it matters.
How should a practice evaluate a veterinary imaging AI product?
Ask the questions the position statement implies.
What data was the model trained on, and did it include the species, breeds, and body regions you actually image? What is the reported sensitivity and specificity, and who measured it, the vendor or an independent party? Has any of it been peer reviewed? How are errors reported and acted on after deployment? Is there a mechanism for you to flag a miss, and does anything happen when you do?
If a vendor cannot answer those in plain terms, that is your answer. The colleges asked for exactly this kind of scrutiny, and the market has not yet been asked it often enough.
The disclosure point is worth noting too. The statement encourages veterinarians to tell clients when AI has been used in their animal's diagnosis and to offer alternatives. That is a client communication decision most practices have not yet made.
What does this mean for how you set up imaging?
Whatever you conclude about AI, the operational requirement is the same: it has to be easy to get a study in front of a specialist.
If sending a case for an overread means burning a CD, uploading to a portal, and waiting on a file transfer, that friction will decide clinical behaviour more than any position statement does. Studies that should get a second read will not get one, because the process is slow enough that the borderline cases quietly do not go.
The imaging setup worth building is one where routing a study to a radiologist takes seconds and costs nothing in effort. That holds true whether you eventually add AI to the workflow or not. A specialist overread is the current standard of care for a difficult study, and the system around it should reflect that.
FAQ
Can AI diagnose a fracture on a dog X-ray?
Some AI tools can flag suspected fractures on canine radiographs with reasonable accuracy on straightforward cases. They are less reliable on subtle, incomplete, or unusually positioned fractures. The ACVR and ECVDI position is that a qualified veterinarian should interpret any such output rather than acting on it directly, which means AI can prompt a closer look but not confirm the diagnosis.
Is veterinary AI approved or regulated?
There is no equivalent of FDA clearance for most veterinary imaging AI. The ACVR and ECVDI have called for regulatory bodies to establish guidelines to prevent misuse, which indicates that the framework does not currently exist. Practices are effectively evaluating these products themselves, without independent validation to rely on.
Does AI replace a veterinary radiologist?
No. The 2025 ACVR and ECVDI position statement is explicit that a qualified veterinary professional, preferably a board-certified radiologist, should remain in the loop to interpret AI output and safeguard diagnostic quality. Current tools are positioned as an aid to the reading process, not a substitute for the specialist who signs the report.
How accurate is AI at reading veterinary X-rays?
Accuracy varies widely by product, by finding, and by species. The difficulty is that most reported figures come from vendors rather than independent evaluation. The ACVR and ECVDI specifically called for unbiased third-party assessment and published benchmarks, which tells you those figures are not yet available in a form the profession considers reliable.
Should I tell clients if AI was used on their animal's images?
The ACVR and ECVDI encourage veterinarians to disclose AI use to pet owners and to offer alternative diagnostic options where appropriate. There is no legal requirement in most jurisdictions, but disclosure is the position the profession's radiology colleges have taken, and it is worth deciding your practice policy before the question arrives from a client.
Where this leaves you
The AI conversation in veterinary imaging is moving faster than the evidence supporting it. The colleges have said plainly that the current products do not meet the bar and that a specialist stays in the loop.
The thing you can act on today is the workflow underneath. If getting a study to a radiologist is difficult, that is the constraint worth fixing first. Read how Keystone PACS handles study routing and sharing, or see the practice management systems it connects to.