Summary
Two 2026 JAVMA studies show commercial AI veterinary radiology platforms achieve 53-79% balanced accuracy on pathology-confirmed canine abdominal cases, with sensitivity ranging from 71% to 90%. The ACVR/ECVDI position statement requires radiologist oversight; no commercial AI platform meets the bar for autonomous primary reads. Radiologist-in-the-loop workflows are the evidence-aligned standard, not a conservative interpretation of the data.
The implication for your practice is clear: AI detection tools can augment your radiology workflow, but they cannot replace the radiologist's diagnostic judgment. The JAVMA evidence confirms what leading practices already know: AI works best as a second reader, a workflow accelerant, or a triage aid, not as a primary diagnostic engine.
What the JAVMA Studies Actually Measured
The two studies evaluated commercial AI platforms on their ability to detect pathology in canine abdominal radiographs. The key findings:
- Balanced accuracy ranged from 53-79%, depending on the platform and pathology type
- Sensitivity (true positive rate) ranged from 71% to 90% across platforms
- Specificity (true negative rate) was generally higher, 85-95%
- No platform achieved the clinical threshold for autonomous interpretation
These numbers matter because specificity and sensitivity tell different stories in a clinical setting. A tool can score well on one metric while still leaving meaningful diagnostic gaps, which is why radiologist oversight remains essential regardless of platform performance on any single measure.
Why Radiologist Oversight is Mandatory, Not Optional
The ACVR/ECVDI position statement is unambiguous: radiologist oversight of AI-assisted reads is a requirement, not a recommendation. This reflects a consensus across veterinary medicine that AI detection is a tool within a radiologist-led workflow, not a replacement for it.
The reason is straightforward. When an AI tool misses cases, the radiologist must catch what the AI missed. That requires the radiologist to read the study with fresh eyes, not simply review AI flags. The moment the radiologist becomes dependent on AI to direct attention, miss rates rise. This is called automation bias, and it is the silent killer of AI-assisted workflows in medical imaging.
What Radiologist-in-the-Loop Actually Means
Radiologist-in-the-loop means the radiologist remains the primary decision-maker, and AI serves a defined role within that decision process. Three evidence-aligned uses:
- AI as a second reader: The radiologist reads first, then the AI flags potential findings. The radiologist confirms or dismisses each flag.
- AI for triage and prioritization: High-volume imaging centers use AI to flag urgent cases (pneumothorax, free fluid) so critical cases reach the radiologist first.
- AI as a workflow accelerant: The AI pre-populates structured report templates or organizes images by body region, reducing clerical burden without changing diagnostic authority.
What radiologist-in-the-loop does not mean:
- The radiologist rubber-stamps AI findings
- The radiologist reads only cases the AI flagged as abnormal
- The radiologist skips cases where AI reported high confidence
These shortcuts undermine the entire evidence base for AI use. If your vendor is pitching AI as a way to reduce radiologist time by eliminating reads, they are selling a liability, not a tool.
How to Integrate AI into Your Workflow Safely
If you run your own radiology department or work with a teleradiology partner, here is how to structure the workflow:
- Define the AI role upfront: Before licensing any AI tool, write down exactly what it will do. Will it flag findings? Organize studies? Pre-populate reports? Be specific.
- Run a sensitivity/specificity audit: Test the AI on your own cases (or cases representative of your caseload). Do not rely on vendor validation. Your patient population may differ from the JAVMA study cohort.
- Maintain radiologist primacy: The radiologist's read is the gold standard. AI flags are suggestions, not conclusions.
- Use structured reporting or DICOM templates to formalize the workflow: If AI prepopulates a report template, the radiologist edits and signs off. The template becomes the audit trail for oversight.
- Monitor miss rates over time: Track cases where the AI missed pathology, even if the radiologist caught it. If miss rates are rising, the workflow has drifted toward automation bias.
Integrating AI safely requires a PACS or teleradiology platform designed for radiologist-controlled workflows. Not all systems enforce the boundary between AI suggestion and radiologist decision equally well.
The Bottom Line
The 2026 JAVMA studies are not a rejection of AI in veterinary radiology. They are a boundary-setting exercise: AI can assist radiology workflows when properly supervised; it cannot operate autonomously at the evidence threshold required for primary diagnostic authority.
If a vendor claims their AI can handle primary reads without radiologist oversight, they are misrepresenting both the technology and the evidence. The JAVMA data is publicly available. Use it as your standard.
Your radiologists remain the bottleneck in your workflow: not because AI is ineffective, but because diagnosis is a cognitive task that, today, still requires human judgment. AI can make that task faster, more organized, and less error-prone. It cannot replace it. The evidence says so, and your liability insurance agrees.