TL;DR
We rebuilt all 94 of our AI classifiers on a new model architecture between June and the end of August 2026, releasing each to clinics as they cleared our internal benchmarks.
Ten performance measures are also available online for every classifier at vetology.net/ai-classifier-performance. The video below highlights how to use the information and features on that page.
Between June and the end of August 2026, we rebuilt all 94 of our AI classifiers on a new model architecture and re-released them. The library was updated in stages rather than in one switchover, so each rebuilt classifier was available to clinics as soon as it had cleared our internal benchmarks for production. There was nothing to install, and the clinic workflow wasn’t impacted.
Watch: How to use the AI classifier performance page
Watch this two-and-a-half-minute video to see how to find the classifier behind any conclusion on your AI Screening Report, and how to weigh that conclusion against what you know about the patient. It covers searching by a common word or a medical term, filtering by species or body region, plain-language explanations of each metric, and the full details for any classifier, including its confusion matrix.
What the new models look at
The new models interpret the images in a study as a whole rather than one at a time. That approach more closely matches how a radiologist reads a case.
We’ve been building and retraining classifiers since 2021. This release was different because every classifier was rebuilt on new architecture and held to a higher internal benchmark than the previous iterations.
Rebuilding all 94 in one stretch ensured the library is both internally consistent and ready for a new generation of report styles, order forms and integrations.
Ten metrics published per classifier
We publish at least ten performance measures for each classifier, including sensitivity, specificity, positive and negative predictive value, likelihood ratio, accuracy, prevalence, cases checked, and the positive and negative ground truth counts.
- Sensitivity and specificity describe how the classifier behaved on the set it was measured against.
- The predictive values answer a different question, which is what a positive or a negative result means for the patient on the table.
- Those values depend on sensitivity and specificity together with how common the condition was in that measured set, so we publish prevalence next to them.
Performance is not uniform across conditions, which is why we publish the numbers condition by condition rather than as a single headline figure.
Published numbers and professional context
In March 2025, the American College of Veterinary Radiology and the European College of Veterinary Diagnostic Imaging published a joint position statement on artificial intelligence in JAVMA. It named transparency, error reporting, post-implementation monitoring, and unbiased third-party evaluation as conditions for veterinary AI worth trusting. A pilot study in the same journal a year later evaluated six commercial veterinary radiology AI services on general practice canine abdominal radiographs and reported low to moderate performance.
Whether a given tool satisfies what the colleges named is their judgment to make, not ours. We publish our numbers so that judgment can be made on evidence. Every figure on the AI classifier performance page belongs to a model running on live cases today, and the page has been public since January 2026. We also have an open invitation to independent researchers to test those numbers.
New classifiers will be released on a monthly cadence going forward, and we’ll update the page to reflect our most up-to-date performance metrics. The AI screening report is one input into a diagnosis and treatment plan. It is automatically generated within minutes of autosending your radiographs. It does not have access to patient history, presentation or your clinical concerns.
Current figures for every classifier: vetology.net/ai-classifier-performance