Radiomics Modeling of Catastrophic Proximal Sesamoid Bone Fractures in Thoroughbred Racehorses Using μCT.
Authors: Basran Parminder S, McDonough Sean, Palmer Scott, Reesink Heidi L
Journal: Animals : an open access journal from MDPI
Summary
# Editorial Summary: Radiomics and Machine Learning for Sesamoid Fracture Prediction Proximal sesamoid bone fractures represent the leading musculoskeletal catastrophe in racing Thoroughbreds, yet current imaging strategies provide limited prospective risk assessment. Researchers at this institution used micro-CT scanning and radiomic analysis—extracting quantitative textural and structural features from bone tissue—to develop machine learning models capable of distinguishing between bones that subsequently fractured and those that remained intact, examining 129 intact contralateral sesamoids from 30 horses with documented catastrophic fractures and 20 control animals. Six different algorithms achieved accuracy ranging from 64–90%, with Support Vector Machine, Random Forest, and logistic regression models performing substantially better than other approaches; notably, prediction accuracy peaked at the finest scanning resolution (0.5 mm voxels) and declined sharply at coarser resolutions, suggesting that subtle microarchitectural changes below conventional radiographic detection thresholds may precede fracture. Whilst these findings currently apply only to post-mortem imaging analysis, the capacity to identify high-risk sesamoid bone architecture offers genuine potential for in-vivo screening of racing stock, potentially enabling targeted intervention strategies—whether through modified training protocols, farriery adjustments, or selective breeding—to reduce one of the sport's most significant welfare and economic challenges.
Read the full abstract on PubMed
Practical Takeaways
- •This proof-of-concept suggests µCT radiomics may eventually allow identification of horses at high fracture risk before catastrophic injury occurs, potentially enabling preventive interventions
- •The optimal imaging resolution (0.5 mm) should be prioritized if translating this approach to clinical standing horse examinations
- •Current findings are limited to laboratory analysis of excised bones; further research is needed to validate whether these radiomic patterns are detectable and predictive in living horses
Key Findings
- •Machine learning models achieved 64.3–90.3% accuracy (mean 75.4%) in predicting PSB fractures from µCT radiomic features of intact contralateral bones
- •Support Vector Machine, Random Forest (RUS Boost), and Logistic Regression models outperformed K-means Nearest Neighbor, Neural Network, and Bagged Trees models
- •Model accuracy peaked at 0.5 mm voxel resampling resolution and decreased substantially at ≥1 mm resolution
- •Radiomics successfully differentiated unfractured PSBs from horses with catastrophic fractures versus control horses in this in vitro dataset