FeTTL：多机构医学影像的联邦模板与任务学习

Federated learning enables collaborative model training across geographically distributed medical centers while preserving data privacy. However, domain shifts and heterogeneity in data often lead to a degradation in model performance. Medical imaging applications are particularly affected by variations in acquisition protocols, scanner types, and patient populations. To address these issues, Federated Template and Task Learning (FeTTL) is introduced. FeTTL proposes a novel federated learning framework that aims to extract generalizable representations from shared knowledge across different institutions while adapting to institution-specific data characteristics through task learning. The core idea of this method is to decompose the model into a sharable template component and institution-specific task components. The template component is federatedly trained across all participating institutions to capture common patterns and features, thereby enhancing the model's generalization capabilities. The task components are trained locally at each institution to learn and adapt to its unique data distribution and specific task requirements, effectively addressing domain shifts and heterogeneity. This decomposition strategy ensures that the model can leverage rich data from multiple institutions while maintaining privacy, improving robustness and accuracy in complex and varied medical imaging environments. FeTTL optimizes the federated aggregation mechanism to ensure that the shared template effectively integrates data insights from diverse sources while preventing the leakage of sensitive information. Experimental results demonstrate that FeTTL significantly outperforms traditional federated learning methods in handling multi-institutional medical imaging data, particularly showcasing superior adaptability and generalization when dealing with datasets exhibiting significant domain differences.