An Automated Scoring Method for Rheumatoid Arthritis X-ray Images Based on Global-Local Feature Fusion

• 1、 School of Information and Communication Engineering, Beijing University of Posts and Telecommunications, Beijing 100876

Abstract：To address the limitations of existing automated X-ray scoring methods for rheumatoid arthritis (RA) in balancing global context and joint-level interpretability, a dual-level automated evaluation framework fusing global and local features is proposed. This method combines global score regression with joint-by-joint classification through two designed strategies: a plug-and-play hybrid architecture and an end-to-end multi-task learning framework. By incorporating various backbone networks, the performance was comprehensively evaluated in joint space narrowing and bone erosion scoring tasks, demonstrating effective learning capabilities even with incomplete joint annotations. Experimental results indicate that, compared to single-task baseline models, the proposed hybrid framework under optimal configurations reduces the mean absolute error of global score predictions by up to 23\% and improves joint-level classification accuracy by over 30\%. The study concludes that this dual-level architecture provides accurate and highly interpretable RA X-ray scoring, maintains high robustness under incomplete annotation conditions, and exhibits significant potential for clinical auxiliary diagnosis.

Keywords： Medical Imaging Automated X-ray Scoring Deep Learning Global-Local Feature Fusion Multi-Task Learning