TwinFedSeg: Anatomical Twin State-Guided Federated Learning for Multimodal Medical Image Segmentation

Federated learning enables privacy-preserving collaborative healthcare, while digital twins can represent institution-specific anatomical and modality states. However, federated multimodal medical image segmentation remains challenged by cross-center heterogeneity and missing modalities, while existing methods largely overlook the anatomical and modality complementarity of client updates. To address this problem, we propose TwinFedSeg, a federated learning framework guided by an anatomical twin state for multimodal medical image segmentation. Each client maintains a lightweight state comprising dynamic class prototypes, structural statistics, modality availability, and reliability. These states guide client selection, model aggregation, and global state updating, while the updated global state is fed back through a lightweight feature modulation adapter for cross-center adaptation under missing modalities. Extensive experiments on BraTS 2023-GLi and ISLES 2022 show that TwinFedSeg consistently outperforms representative federated methods, achieving average Dice scores of 80.77% and 72.58%, respectively, with gains of 3.52 and 4.06 percentage points over FedAvg.