Med-SSFWT: A Self-supervised Federated Weight Transfer Framework for Medical Model Fusion

Abstract

Abstract Artificial Intelligence (AI) holds great potential to revolutionize healthcare by integrating and analyzing diverse multi-source medical data to drive advancements in disease diagnosis, treatment strategies, and patient management. However, deploying AI in distributed medical environments presents critical challenges, including data silos, label deficiency, and data heterogeneity. To address these challenges and enable effective and privacy-preserving distributed medical AI models, we propose Med-SSFWT, a Self-Supervised Federated Weight Transfer framework designed for medical data fusion. Firstly, Med-SSFWT employs a fine-tuned Large Language Model (LLM) to extract structured features from each client’s medical data, followed by feature alignment across clients via a shared global schema. Subsequently, an information gain-based gradient filtering mechanism is introduced to federated aggregation by filtering out ineffective gradients, thereby improving the robustness of global model. Furthermore, Med-SSFWT leverages a novel federated model fusion frame, consisting of self-supervised pre-training and fine-tuning through weight transfer to balance global optimization with client-specific personalization. Finally, extensive experiments show that Med-SSFWT consistently outperforms federated learning approaches in both performance and adaptability under diverse non-IID conditions, highlighting its effectiveness within distributed medical environments and establishing a foundation for the development of privacy-preserving and scalable AI-driven healthcare solutions.

Related Publications