A multi-task learning model for clinically interpretable sesamoiditis grading.

Area of Science:

• Veterinary Medicine
• Artificial Intelligence
• Medical Imaging

Background:

• Equine sesamoiditis is a prevalent condition impacting horse performance and increasing injury risk.
• Accurate grading of sesamoiditis is essential for effective treatment strategies.
• Current deep learning models for sesamoiditis lack clinical interpretability.

Purpose of the Study:

• To develop a novel, clinically interpretable multi-task learning model for equine sesamoiditis grading.
• To integrate clinical knowledge with machine learning for improved diagnostic accuracy.
• To enhance the transparency of AI-driven diagnostic decisions in veterinary medicine.

Main Methods:

• A dual-branch decoder architecture was employed for simultaneous sesamoiditis grading and vascular channel segmentation.
• Feature fusion was utilized to facilitate knowledge transfer between the two tasks.
• A diagnostic report and vascular channel mask were generated to explain grading decisions.

Main Results:

• The proposed model demonstrated superior performance compared to state-of-the-art methods on two independent datasets.
• The model achieved high accuracy in both sesamoiditis grading and vascular channel segmentation.
• The generated diagnostic reports provided clear explanations for the model's grading outcomes.

Conclusions:

• The developed model offers a clinically interpretable and accurate approach for grading equine sesamoiditis.
• This framework provides a foundation for interpretable AI in diagnosing similar veterinary diseases.
• The integration of feature fusion and multi-task learning enhances diagnostic capabilities and transparency.