Deep learning model shows promise for detecting and grading sesamoiditis in horse radiographs.

The research paper discusses the development of a deep learning model for the diagnosis and grading of sesamoiditis in horses using radiographs.

Objective and Methods

• The researchers focused on creating an efficient machine learning tool, specifically designed for use under data-limited conditions.
• A dataset of 255 equine radiographs, both dorsolateral-palmaromedial oblique (DLPMO) and dorsomedial-palmarolateral oblique (DMPLO), was utilized for this study. The images, obtained from the Hagyard Equine Medical Institute, were anonymized and categorized into three levels of sesamoiditis severity – normal, mild, and moderate.
• Two RetinaNet models were employed consecutively, each performing a specific task. The first model’s task was to localize the sesamoid bone in the radiographs, whereas the second model graded the severity of sesamoiditis based on the localized region. The second RetinaNet’s classification subnetwork incorporated a self-attention module.
• The performance of the model was evaluated using a confusion matrix and average precision (AP), to determine its overall accuracy and precision in detecting sesamoiditis and its severity.

Results and Conclusions

• The deep learning model had a marked performance, with an accuracy rate of 92.7%, surpassing the base RetinaNet model’s level of accuracy.
• The model also had a mean average precision (mAP) of 81.8%, indicating its superior ability to detect objects or issues within the radiographs.
• Notably, the model showed significant improvements in the accuracy of results in each of the severity categories.
• Its ability to accurately localize sesamoid bones and grade sesamoiditis severity makes it a potentially valuable tool that could be integrated into a clinical environment to improve the interpretation of metacarpophalangeal (fetlock) joint radiographs in horses.
• Lastly, by increasing the size of the training dataset, the model could potentially learn to assist in diagnosing pathologies in other skeletal regions of a horse, thereby expanding its usability.