Deep learning-based binary classification and probabilistic calibration of navicular bone pathology in equine radiographs

Radiographic assessment of the equine navicular bone is challenging because pathological changes may be subtle, surrounding anatomical structures can overlap, and image interpretation may vary among observers. This retrospective single-center study developed and evaluated deep learning models for the binary classification of equine navicular bone radiographs as radiographically normal or pathological. The dataset comprised 528 radiographs from 176 horses. ResNet-50 was used as the primary classifier, while Xception and VGG-16 were evaluated as comparator architectures. All data partitions were created at the horse level to prevent subject leakage. The development dataset included 140 horses and was evaluated using five-fold cross-validation, while a temporally separated internal hold-out set containing 36 horses and 108 radiographs was reserved for final testing. ResNet-50 achieved a mean AUROC of 0.893, an accuracy of 79.6%, and an F1-score of 85.5% across the cross-validation folds. On the hold-out test set, the model achieved an AUROC of 0.905, an accuracy of 82.4%, a sensitivity of 90.5%, a specificity of 71.1%, and an F1-score of 85.7%. Temperature scaling produced modest improvements in probability calibration, reducing the Brier score from 0.1582 to 0.1533 and the expected calibration error from 0.1639 to 0.1595 without altering discrimination. Qualitative Grad-CAM analysis was additionally used to illustrate image regions associated with model predictions. These findings support the feasibility of deep learning for automated classification of navicular bone radiographic appearances. However, the model does not diagnose clinical navicular syndrome, assess lesion severity, or replace veterinary interpretation. Prospective external validation using multi-center datasets is required before clinical implementation.