Detection of mare parturition through balanced multi-scale feature fusion based on improved Libra RCNN.
Authors: Wang Buyu, Duan Weijun, Zhao Jian, Bai Dongyi
Journal: PloS one
Summary
# Editorial Summary: AI-Based Mare Parturition Detection Monitoring mares for signs of impending labour remains labour-intensive and imperfectly reliable in large stud facilities, yet early intervention during dystocia can be the difference between a live foal and a stillborn. Wang and colleagues have developed an artificial intelligence algorithm based on improved Libra RCNN architecture that can automatically detect parturition onset through video analysis, utilising enhanced feature extraction via ResNet101 backbone with attention modules, balanced multi-scale feature fusion (CARAFE-BFP), and geometric region-of-interest modules to handle the subtle physiological changes and variable data distributions inherent in equine parturition. Testing across imbalanced datasets achieved 86.26% mean average precision with 98.17% recall rate at 15 frames per second processing speed; when applied to continuous video stream monitoring with statistical smoothing, the system correctly identified parturition onset in 92.75% of cases. For breeding operations managing multiple mares simultaneously, this non-contact, stress-free technology offers potential to automate intensive labour monitoring, enabling staff to prioritise intervention decisions and respond more rapidly to complications—though practical implementation would require standardised installation protocols, validation across diverse facility conditions and mare phenotypes, and integration with existing stud management systems.
Read the full abstract on PubMed
Practical Takeaways
- •Automated parturition detection systems could enable 24/7 monitoring of mares without observer fatigue, allowing rapid intervention during dystocia when outcomes depend on timing
- •Non-contact, stress-free monitoring via computer vision eliminates human disturbance that could interfere with natural parturition progression
- •Implementation in large-scale breeding operations could reduce stillbirth rates and mare mortality by enabling timely veterinary intervention during abnormal presentations
Key Findings
- •Improved Libra RCNN algorithm with CBAM attention module achieved 86.26% mean average precision in detecting mare parturition from imbalanced datasets
- •Video stream monitoring using sliding window mechanism achieved 92.75% accuracy for real-time mare parturition detection
- •Algorithm demonstrated 98.17% average recall rate with processing speed of 15.06 images per second, enabling automated non-contact monitoring