External Validation of a Custom-Trained Artificial Intelligence Model for Canine Ear Canal Disease and Comparison With Veterinary Professionals.

AI otoscopy models show promise as decision-support tools but currently do not outperform veterinary professionals.

This study externally validated four custom-trained YOLOv8n artificial intelligence (AI) models designed to classify canine otoscopic images into three categories: healthy, otitis externa (OE), and ear canal masses. Four models (A–D) were trained on datasets of varying size and deduplication status and tested on 300 previously unseen images. Ground truth labels were assigned by a board-certified veterinary dermatologist. Six veterinary professionals of varying experience levels independently classified the same images for comparison. Model D, trained on the largest deduplicated dataset, achieved the best external performance with 66.3% overall agreement and a Cohen's κ of 0.508 at a 0.25 confidence threshold. Class-specific sensitivity and specificity for Model D were strongest for healthy ears (90.0%/69.0%) and ear canal masses (50.0%/93.5%), while OE detection was moderate (59.0%/75.5%). All six veterinary professionals outperformed the AI models, achieving a median agreement of 77.2% and κ of 0.658, with the board-certified referral clinician attaining the highest concordance. The study confirms that larger, deduplicated training datasets improve model generalizability. However, the AI system did not surpass veterinary professionals, including non-specialists. The authors suggest the model is best suited as a clinical decision-support tool to assist practitioners rather than replace expert judgment. Future work should focus on expanding and diversifying training datasets to further improve sensitivity for OE and mass detection.

This summary was distilled by AI and may occasionally misinterpret data. Confirm critical details with the primary literature before clinical application.