A multi-task learning model for clinically interpretable sesamoiditis grading

• Published in: Computers in biology and medicine Vol. 182; p. 109179
• Main Authors: Guo, Li, Tahir, Anas M., Hore, Michael, Collins, Andrew, Rideout, Andrew, Wang, Z. Jane
• Format: Journal Article
• Language: English
• Published: United States Elsevier Ltd 01.11.2024
• Subjects: Cross-attention; Interpretability; Radiograph; Sesamoiditis; Visual transformer; Sesamoiditis; Radiograph; Visual transformer; Cross-attention; Interpretability
• ISSN: 0010-4825; 1879-0534
• DOI: 10.1016/j.compbiomed.2024.109179

Sesamoiditis is a common equine disease with varying severity, leading to increased injury risks and performance degradation in horses. Accurate grading of sesamoiditis is crucial for effective treatment. Although deep learning-based approaches for grading sesamoiditis show promise, they remain underexplored and often lack clinical interpretability. To address this issue, we propose a novel, clinically interpretable multi-task learning model that integrates clinical knowledge with machine learning. The proposed model employs a dual-branch decoder to simultaneously perform sesamoiditis grading and vascular channel segmentation. Feature fusion is utilized to transfer knowledge between these tasks, enabling the identification of subtle radiographic variations. Additionally, our model generates a diagnostic report that, along with the vascular channel mask, serves as an explanation of the model’s grading decisions, thereby increasing the transparency of the decision-making process. We validate our model on two datasets, demonstrating its superior performance compared to state-of-the-art models in terms of accuracy and generalization. This study provides a foundational framework for the interpretable grading of similar diseases. •Innovative multi-task learning model grades sesamoiditis in equine radiographs with clinical interpretability.•Dual-branch decoder and feature fusion improve grading accuracy by focusing on vascular channels.•Visual and quantitative explanations make model decisions intuitive to clinicians.•Superior performance validated on both development and independent datasets.