VETNET Publishes Initial Multi-Institutional Veterinary Benchmark Datasets
VETNET supplies shared veterinary datasets and evaluation protocols to close the gap between human and animal AI development. Multi-institutional structure supports reproducible model testing on clinically relevant tasks. The framework enables measurable progress tracking for diagnostic applications in practice.
VETNET establishes live benchmarks across species including canines, equines, and bovines. The framework aggregates de-identified imaging, lab results, and outcome labels under unified schemas. Multi-site collection protocols enforce consistent annotation standards absent in prior veterinary collections.
Existing veterinary AI studies remain siloed with single-clinic cohorts under 5,000 cases. VETNET targets this fragmentation by requiring participating models to report performance on held-out test sets from external sites. Early baselines show radiology task AUCs between 0.78 and 0.91 depending on species and imaging modality.
Operational deployment requires models to handle variable image quality and species-specific anatomy. Standardized splits enable direct comparison of detection rates for conditions such as bovine respiratory disease and canine lymphoma. Institutions gain shared evaluation infrastructure without exchanging raw records.
Next releases will incorporate longitudinal treatment outcomes and genomic markers. Participating groups must submit code and weights for independent verification against the public test partitions.
VETNET: At least three external groups submit models exceeding 0.90 AUC on the bovine thoracic radiograph task within 18 months.
Sources (2)
- [1]Primary Source(https://arxiv.org/abs/2610.14149)
- [2]Supporting Source(https://arxiv.org/abs/2306.04512)