Automated detection and classification of transverse versus non-transverse fractures in tibia of cats and dogs using artificial intelligence

Fracture classification is central to treatment planning in veterinary orthopedics. This study aimed to evaluate the performance of You Only Look Once (YOLO) v11 for the rapid and accurate detection of transverse tibial fractures in cats and dogs and to compare its performance with an EfficientNet image classifier using clinically relevant performance metrics. We developed and compared two deep learning approaches for automated detection/classification of transverse vs. non-transverse tibial fractures on radiographs: an object detector (YOLOv11) and an image classifier (EfficientNet). From an initial dataset of 744 radiographs, 722 radiographs with corresponding Extensible Markup Language (XML) annotations (684 unique cases) were included after data curation. We performed case-level stratified splitting (train/validation/test = 480/102/102 cases). On the held-out test set, YOLOv11 achieved a Receiver Operating Characteristic (ROC)-Area Under the Curve (AUC) of 0.7265, a Precision-Recall (PR)-AUC of 0.7391, a precision of 0.5750, a recall of 0.902, and an F1-score of 0.7023 at the threshold that maximized the F1-score, while EfficientNet achieved a ROC-AUC of 0.642, a PR-AUC of 0.601, a precision of 0.562, a recall of 0.804, and an F1-score of 0.661 at its F1-optimal threshold. The results indicate that localization-aware detection yields higher sensitivity and overall F1-score than global classification, highlighting the clinical promise of modern object detection models for automated classification of tibial fracture types in cats and dogs while underscoring the need to reduce false-positive predictions.