Radiomics Modeling of Catastrophic Proximal Sesamoid Bone Fractures in Thoroughbred Racehorses Using μCT
1
Clinical Sciences, Cornell University, Ithaca, NY 14853, USA
2
Biomedical Sciences, Cornell University, Ithaca, NY 14853, USA
3
Population Medicine and Diagnostic Sciences, Cornell University, Ithaca, NY 14853, USA
4
Clinical Sciences, Equine and Farm Animal Hospital & Population Medicine and Diagnostic Sciences, Cornell University, Ithaca, NY 14853, USA
*
Authors to whom correspondence should be addressed.
Animals 2022, 12(21), 3033; https://doi.org/10.3390/ani12213033
Submission received: 20 September 2022
/
Revised: 27 October 2022
/
Accepted: 1 November 2022
/
Published: 4 November 2022
(This article belongs to the Special Issue Injuries, Diagnosis and Treatment of Thoroughbred Racehorses)
Simple Summary
Mitigating the risk of catastrophic injuries in the horse racing industry remains a challenge. Non-invasive methods such as CT imaging in combination with machine learning could be used to screen horses at risk of injury, but there remain questions on the feasibility of such an approach. In this work, we investigated whether machine learning models could be developed from in vitro harvested μCT images of intact proximal sesamoid bones to predict whether the bone was from a horse that suffered a catastrophic injury or from a control group. The average accuracy in differentiating whether a sesamoid bone came from a case or control horse using our approach was 0.754. Our work suggests it may be possible to develop similar models using CT images of horses in the clinical setting.
Abstract
Proximal sesamoid bone (PSB) fractures are the most common musculoskeletal injury in race-horses. X-ray CT imaging can detect expressed radiological features in horses that experienced catastrophic fractures. Our objective was to assess whether expressed radiomic features in the PSBs of 50 horses can be used to develop machine learning models for predicting PSB fractures. The μCTs of intact contralateral PSBs from 50 horses, 30 of which suffered catastrophic fractures, and 20 controls were studied. From the 129 intact μCT images of PSBs, 102 radiomic features were computed using a variety of voxel resampling dimensions. Decision Trees and Wrapper methods were used to identify the 20 top expressed features, and six machine learning algorithms were developed to model the risk of fracture. The accuracy of all machine learning models ranged from 0.643 to 0.903 with an average of 0.754. On average, Support Vector Machine, Random Forest (RUS Boost), and Log-regression models had higher performance than K-means Nearest Neighbor, Neural Network, and Random Forest (Bagged Trees) models. Model accuracy peaked at 0.5 mm and decreased substantially when the resampling resolution was greater than or equal to 1 mm. We find that, for this in vitro dataset, it is possible to differentiate between unfractured PSBs from case and control horses using μCT images. It may be possible to extend these findings to the assessment of fracture risk in standing horses.
Keywords:
equine; machine learning; computed tomography; Thoroughbred; fetlock
