An AI classifier identified abnormal canine brain MRI studies with 83% sensitivity and 78% specificity in a held-out set of 53 examinations. The four-university study provides a measured starting point for automated scan screening across dogs with varied anatomy and imaging conditions.[1]
Published in Scientific Reports on 23 October, the study assembled 550 MRI examinations: 444 for training, 53 for validation and 53 for testing. Its task was to distinguish normal from abnormal imaging, rather than identify the disease responsible for an abnormality.[1]
The result and its uncertainty
Performance on the held-out MRI studies
| Measure | Reported result | Meaning |
|---|---|---|
| Sensitivity | 83%; confidence interval 67–91% | Proportion of abnormal studies flagged |
| Specificity | 78%; confidence interval 54–91% | Proportion of normal studies correctly cleared |
| Balanced accuracy | About 80% | Average of sensitivity and specificity |
| Test set | 53 MRI studies | Examinations held out from training and validation |
The confidence intervals are wide because the independent test set is small. They show why the headline percentages are promising estimates to test further, rather than precise forecasts of performance in another hospital.[1]
From MRI slices to a dog-level decision
MRI scanners produce many cross-sectional images, or slices, of a brain. One patient therefore contributes multiple pictures, but those pictures are not independent patients. The researchers first classified slices. At that level, the system recognised about 74 percent of normal slices and about 73 percent of abnormal ones. It then used the proportion of slices flagged as abnormal to make a judgement about the complete MRI study. The selected patient-level rule treated a study as abnormal when more than roughly 51 percent of its slices crossed that threshold.[1]
Aggregating slices into a study-level decision matters clinically. A small focal lesion may occupy few slices, while a widespread abnormality appears repeatedly. The chosen aggregation threshold therefore affects which kinds of disease the system is most likely to flag.[1]
The investigators analysed performance across institutions and canine characteristics, including breed-related size differences. Those checks matter because the algorithm could otherwise learn cues from a particular scanner, imaging protocol or population rather than general brain pathology. A model that relies partly on image appearance linked to a hospital may fail silently when transferred to another. Independent testing on truly new clinical sources is the strongest way to evaluate that risk.[1]
What a veterinary hospital might eventually gain
A possible role is helping a veterinary team prioritise scans for specialist review. The relevant question is whether the added screen helps clinicians find important abnormalities without an excessive alert burden. The present study measured classification, leaving that workflow benefit for a prospective test.[1]
A larger evaluation could compare veterinarians reading cases with and without assistance, including uncommon breeds, focal abnormalities and scans from new centres. Reporting errors alongside reading time would show whether the classifier improves the service as well as the numerical score.[1]
The useful next step is clear: test the study-level decision in the setting where a missed abnormality or an unnecessary alert changes a veterinarian’s work.
Sources & further reading
- Artificial intelligence for automated classification of canine brain magnetic resonance images Scientific Reports · 23 Oct 2026 · Accessed 24 Oct 2026