Swarm learning network for privacy-preserving and collaborative deep learning assisted diagnosis of fracture: a multi-center diagnostic study.

Background: While artificial intelligence (AI) has revolutionized medical diagnostics, conventional centralized AI models for medical image analysis raise critical concerns regarding data privacy and security. Swarm learning (SL), a decentralized machine learning framework, addresses these limitations by enabling collaborative model training through secure parameter aggregation while preserving data locality. However, no prior studies have specifically developed distributed learning models for fracture recognition due to challenges in multi-institutional data harmonization. We aimed to develop and validate a blockchain-based SL framework for privacy-preserving, multi-institutional fracture image analysis and compare its performance against centralized AI models and clinicians in real-world applications. Methods: We selected knee joint diseases in traumatic orthopedics as representatives to explore the AI imaging evaluation of fractures. The knee joint images were retrospectively obtained from patients diagnosed with knee injuries between December 2013 and July 2023 at 4 independent institutes hospitals in China. A total of 4,581 patients was included for retrospective study and establishment of the explainable and distributed SL model. An explainable object detection algorithm was proposed for the identification of fractures. Based on the architecture, a privacy-preserving SL system was established, and we further validated the performance of the model in external verification sets and clinical use. Finally, the SL system was appraised through a prospective cohort to aid 6 clinicians in the preoperative assessment of 112 patients with knee joint injuries. Results: The YOLOv8n-cls algorithm demonstrated superior performance in centralized experiments and was adapted for SL implementation. Our SL model achieved robust performance in both balanced (AUROC 0.991 ± 0.003, accuracy 0.960 ± 0.013) and unbalanced (AUROC 0.990 ± 0.005, accuracy 0.944 ± 0.021) datasets. External validation yielded an AUROC of 0.953 ± 0.016, matching centralized model performance while maintaining data privacy. Clinically, the SL system achieved 86.8% diagnostic accuracy and assisted treatment decisions in 91.5% of cases, outperforming junior clinicians and rivaling senior specialists in diagnostic efficiency. Conclusion: This study establishes blockchain-based SL as a secure, privacy-preserving paradigm for distributed AI training in medical imaging, with particular relevance for emergency orthopedic diagnostics. Our framework enables effective multi-center collaboration without compromising data security, addressing a critical need in modern healthcare AI. Clinical trial registration: [https://www.chictr.org.cn/showproj.html?proj=193847], identifier [ChiCTR2300070658].