Federated learning-driven intelligent framework for multi-center radiotherapy dose distribution prediction oriented toward linear accelerators.

High-quality radiotherapy dose distribution prediction for linear accelerators remains a labor-intensive process constrained by inter-planner variability and institutional data silos. While deep learning has shown promise in automating dose distribution prediction, most existing methods are trained on single-center datasets, limiting their generalizability. Direct multi-center data pooling is hindered by stringent privacy regulations and heterogeneous clinical protocols. This paper proposes a federated learning-driven framework that enables collaborative model training across geographically distributed institutions without exchanging raw patient data. The framework comprises a multi-scale attention U-Net for three-dimensional dose prediction and an adaptive weighted federated aggregation strategy that dynamically balances data volume and local model quality to address non-independent and non-identically distributed data challenges. A layered privacy protection mechanism integrating gradient-clipped differential privacy with secure aggregation provides privacy-enhancing protections with quantifiable bounds during parameter exchange. Experiments conducted across four clinical centers on head-and-neck and abdominal IMRT cases demonstrate that the proposed approach achieves a mean Gamma pass rate of 96.8%, closely approaching the centralized training upper bound of 97.5% while significantly outperforming single-center models and standard federated averaging. Ablation studies confirm the individual contributions of adaptive weighting and dual attention modules, and robustness analyses validate fault tolerance under client dropout and adversarial conditions. The proposed framework offers a practical and privacy-preserving pathway for breaking data silos in AI-driven radiotherapy research.