Veterinary AI Validation Lags Human Radiology by a Decade
“A peer-reviewed Frontiers commentary published in June 2025 by four veterinary AI researchers — including the lead author of the ACVR/ECVDI’s official AI position statement — methodically dismantled the only published external validation study of a major veterinary AI radiology product. Circular ground truth, severe class imbalance, sensitivity of 0.444 in difficult cases, the wrong statistical test, no version traceability. That is the state of validation in commercial veterinary AI. On the human side, by contrast, a model called CheXNet was trained on 112,120 publicly released chest radiographs in 2017, validated against three independent cardiothoracic specialists, published in PLOS Medicine, and then beaten on the public leaderboard by hundreds of subsequent teams. That is what the scientific method looks like in medical AI. The veterinary industry skipped it.”