Skip to content
Menu

BlogGuides

AI for veterinarians: where it's actually used in clinics

A veterinarian in scrubs holds a Pomeranian dog, with an ultrasound machine switched on beside the exam table
Add as a preferred source on Google

One click, and Google will show my articles more often in your results.

In veterinary clinics, artificial intelligence is mostly used to support the reading of X-rays and ultrasounds, the field with the most independent research behind it. But the broadest independent review, published in 2026 on six commercial platforms, found that none is yet ready for clinical use on its own: sensitivity for individual findings dropped as low as 23% for intestinal obstruction, a surgical emergency.

The toughest verdict: none of the six platforms is ready

A research group at Murdoch University, in Australia, put six commercial AI platforms for veterinary radiology, Picoxia, Radimal, RapidRead, SignalRAY, Vetology AI, and X Caliber Vet AI, through an external validation on 53 canine abdominal radiographs with diagnoses already confirmed by surgery, necropsy, CT, ultrasound, cytology, or response to treatment. The cases, collected between September and December 2024 from general-practice clinics, hadn't been selected to make the AI's job easier: that's the difference between a lab test and the X-ray that actually comes through the door of a clinic.

The result, published online on 20 March 2026 in the Journal of the American Veterinary Medical Association (authors Doris Ma, Josephine Faulkner, Nerissa Stander, Anthea Raisis, and Stephen Joslyn), is blunt. When telling a pathological case apart from a normal one, the six platforms reached a balanced accuracy of only 60-65%. On individual radiographic findings, sensitivity ranged from 28% to 78% depending on the platform, and for small bowel obstruction, a condition that calls for rapid surgery, it fell as low as 23%: one case in four, on the worst-performing platform, goes unnoticed. The authors conclude that performance remains low to moderate, that even the best-performing algorithm has notable limitations, and that none of the six appears suitable for clinical use in its current form. They call for larger, independent external validation before these tools enter everyday practice.

Even the most favorable test shows a drop on hard cases

A year earlier, a more encouraging picture had circulated, but it's worth checking who paid for it. The study, published on 21 February 2025 in Frontiers in Veterinary Science, compared SignalRAY, SignalPET's software, against 11 board-certified veterinary radiologists on 50 radiographic studies of dogs and cats. Across all cases, the AI's overall accuracy, 90.2%, matched the best radiologist in the group, with higher specificity (94.4%) but lower sensitivity (68.8%) than the human average.

What matters for anyone evaluating the tool, though, is what happens on ambiguous cases: in that subgroup, the AI's accuracy drops to 70.1% and sensitivity to 44.4%, a sharp fall from the easy cases. The study was funded entirely by SignalPET, which covered the radiologists' fees and the processing of cases through its own software, while the conflict-of-interest section states there were no relevant commercial or financial relationships: the two statements, read side by side, sit uneasily together.

The Italian case: Padua's research on a different problem

In Italy, the most documented case concerns a narrower clinical problem. A group from the Radiology Unit at the University of Padua, led by Tommaso Banzato, trained a neural network to recognize the malignancy grade of meningioma, the most common brain tumor in dogs, from MRI scans. The study, published in The Veterinary Journal in 2018, used 56 cases of canine meningioma from two veterinary hospitals: 26 grade I, 22 grade II, 8 grade III. The better-performing of the two models tested correctly classified 8 out of 10 images in a test group kept separate from training. The final test group, though, was just ten images, too small to say how the model holds up on new cases, and a meningioma diagnosis is still only confirmed by histological examination after surgery or autopsy.

What Italy's veterinary federation is doing

In Italy the topic is also followed at the institutional level by FNOVI, the National Federation of Italian Veterinary Orders, which has a working group dedicated to artificial intelligence, coordinated by Orlando Paciello, professor of Veterinary Pathological Anatomy at the University of Naples Federico II. In 2025 the group ran a distance continuing education course, active from 3 March to 21 September, with speakers from Naples, Padua, and Bologna. Among them, Tommaso Banzato himself gave a talk titled "AI in veterinary diagnostic imaging, applications and limits": a title chosen by the federation itself, putting the limits on equal footing with the applications.

What it means for a clinic owner

Three independent sources, two scientific studies and one's own federation's training program, converge on the same caution: reading images automatically is the area of veterinary AI with the most serious research behind it, yet even the broadest independent validation didn't find a single commercial platform ready to replace a radiologist's report, and even the test most favorable to a single product gets noticeably worse exactly on the cases where a mistake costs the most. It's the same kind of caution that applies to any regulated intellectual profession facing a striking AI announcement: I wrote about this for law firms, where 55% of Italian lawyers already use AI but only 19.2% have a firm website, another case where adopting a new tool outpaces the basics of being found online.

If you run a clinic and want to know how to be found by people searching online for a vet in your area before you evaluate an advanced diagnostic tool, get in touch: let's talk it through.

Frequently asked questions

Where is AI actually used in veterinary clinics today?

Mainly as support for reading X-rays and ultrasounds: it's the field with the most independent studies behind it. The broadest review, from 2026, tested six commercial platforms and concluded that none is yet ready for clinical use on its own.

Do AI platforms for veterinary radiology work well?

Unevenly. In the 2026 external validation study (JAVMA, Murdoch University), sensitivity for individual findings ranged from 28% to 78% depending on the platform, and for small bowel obstruction, a surgical emergency, it dropped as low as 23%.

Can AI software replace the veterinarian in making a diagnosis?

No. Not even the most favorable study makes that claim: despite a high overall accuracy, its accuracy drops to 70.1% and sensitivity to 44.4% on the radiographically most ambiguous cases, exactly where expert judgment matters most.

What is the Padua study on canine brain tumors?

A 2018 study, published in The Veterinary Journal by a University of Padua group led by Tommaso Banzato, that trained a neural network on 56 cases of canine meningioma to predict malignancy grade from MRI scans, correctly classifying 8 out of 10 images in a separate test set.

Does the Italian federation of veterinarians work on artificial intelligence?

Yes. FNOVI, Italy's national federation of veterinary orders, has a dedicated working group, coordinated by a professor at the University of Naples Federico II, and in 2025 ran a continuing education course on AI in veterinary medicine, with a specific talk on the applications and limits of diagnostic imaging.

Should a clinic invest in an AI diagnostic tool right now?

The available data describes tools still being validated: the broadest independent review found none ready for clinical use, and even the most favorable tests, funded by whoever sells the product, show a sharp drop on difficult cases. Before a specialized investment, the more immediate priority is often being findable online.

More articles

Want to know where to start before investing in a specialized tool?

Tell me about your project