Harnessing deep learning for automated horse morphometry estimation

Body conformation plays a fundamental role in equine performance, biomechanics, and overall health. Traditional morphometric assessments, based on manual measurements and visual scoring, are time-consuming, subjective, and often impractical under field conditions. In this study, a deep learning-based framework was developed to automatically estimate horse morphometric traits from lateral view images. Two datasets were collected: Dataset A, consisting of 1000 standardized images used to train convolutional neural networks (CNNs), and Dataset B, comprising 106 field images with corresponding manual measurements used for testing. Four CNN architectures (MobileNetV3-small-100, EfficientNet-B0, ResNet50, and HRNetW48) were evaluated for keypoint estimation accuracy. Each network was trained to predict 16 anatomical landmarks, enabling automatic reconstruction of linear body measurements. Across models, mean Average Precision exceeded 0.9, demonstrating high keypoint localization accuracy. ResNet50 achieved the best overall performance, followed closely by EfficientNet-B0, which offered an optimal trade-off between accuracy and computational efficiency, making it suitable for mobile deployment. Automatically reconstructed body measurements yielded a mean absolute error of 3.5 to 4 cm, with central landmarks showing higher stability than distal ones. Procrustes analysis further confirmed that the CNN preserved the global body shape and morphological variability among individuals. To our knowledge, this work represents one of the first large-scale applications of CNNs to equine morphometry, providing a robust and scalable framework for objective conformation assessment. The proposed system enables automated, non-invasive, and field-applicable morphometric evaluations, supporting data-driven breeding, performance monitoring, and welfare assessment.