Is veterinary ultrasound AI ready for practice?

Artificial intelligence could help veterinarians recognize abnormal fluid or air during point-of-care ultrasound (POCUS). A March 2026 study published in Frontiers in Veterinary Science evaluated the feasibility and diagnostic performance of models for detecting effusions and pneumothorax in dogs, with encouraging results on selected ultrasound images.[1] For clinicians, the next question is whether that performance would hold during everyday examinations. The study supports further development, but does not yet establish that these models are ready for routine clinical use.

Study objective

That distinction concerns the models evaluated in this publication. It is not a verdict on every AI application in veterinary imaging. Understanding the conditions under which these results were obtained helps identify what should be tested next.

Key findings

The best models achieved mean sensitivity and accuracy of 98% and 97% at the diaphragmatic-hepatic site, 87% and 85% at the pericardial sites, 81% and 88% at the chest tube sites, 83% and 78% for pooled splenorenal and hepatorenal views, and 84% and 77% at the cystocolic site, respectively. Sensitivity describes detection of positive images, while accuracy includes correct positive and negative classifications. These were results from selected images of dogs held out from training, not complete patient examinations.[1]

Clinical utility

The authors set development targets of at least 80% sensitivity and 75% accuracy. Meeting those targets was encouraging, but they were research benchmarks, not established criteria for clinical readiness. Performance also varied across scan sites and validation splits.[1]

The reported results describe classification of selected images, not the diagnostic accuracy of complete patient examinations. A practical evaluation should establish how image-level predictions translate into an examination-level conclusion. It should also examine what happens after an incorrect output. A false negative could provide reassurance when further assessment is needed, while a false positive could prompt unnecessary investigation. The study did not test these downstream effects or demonstrate improved patient outcomes.

Study limitations

The study enrolled five healthy military working dogs and twenty client-owned dogs. One client-owned dog underwent repeat imaging, producing an additional dataset without adding another patient. Individual scan-site datasets contained fewer dogs still. Thousands of frames therefore represented a limited number of animals, rather than thousands of independent clinical cases.[1]

This matters because repeated images from one dog cannot provide the anatomical and disease diversity of additional patients. Recruitment targeted dogs weighing 22–55 kg and excluded brachycephalic breeds. Performance in smaller dogs or breeds with different thoracic anatomy remains uncertain. Larger datasets should also include enough positive and negative cases to assess missed abnormalities and false alarms reliably.

An emergency and critical care specialist reviewed the recordings, excluded videos with insufficient detail and trimmed positive clips to retain periods when the abnormality was visible. Of 414 recorded videos, 220 entered the dataset used for frame extraction and model training.[1]

This curation helped establish interpretable examples, but it also defines the task the models faced. Classifying an image that already shows an abnormality is different from helping a clinician find it during an examination. In practice, the probe moves, views may be incomplete and image quality varies. A clinically useful system would need evaluation across that acquisition process, including how it handles images that are inadequate for interpretation.

The researchers used leave-one-subject-out cross-validation, holding a dog out for testing and assigning entire subjects to the training and validation groups. Keeping images from the same dog out of competing groups is an important safeguard against data leakage.[1]

Nevertheless, this remained an internal evaluation of the study dataset. All images were acquired with a Sonosite Edge system and C11 transducer. The study therefore does not establish performance with other machines, probes or clinical teams. External validation would test a fixed model on independently collected data from other hospitals, without using those cases to adjust it beforehand.

Practice takeaways

For these models, a useful next step would be prospective evaluation with new patients across several hospitals, ultrasound platforms and levels of operator experience. That evaluation should include routine acquisition difficulties and independently established clinical diagnoses. Comparing clinicians working with and without AI would help determine whether assistance improves decisions, how quickly useful information becomes available and whether errors increase in particular circumstances.

These are proposed next steps arising from the study design, not benefits already demonstrated. The publication offers a credible basis for further research. Before these models guide routine care, clinicians need evidence that their performance remains dependable beyond the selected images and conditions in which they were developed.