Multimodal medical image fusion has demonstrated significant potential in improving diagnostic and segmentation accuracy. However, modality heterogeneity and privacy protection requirements pose challenges to joint modeling. Existing multimodal federated learning approaches often struggle to balance communication efficiency with model generalization, limiting their practical effectiveness. To address this, we propose a wavelet decomposition-based multimodal federated learning framework, WaveFed. This method extracts key low-frequency features on the client side through wavelet decomposition, constructing compact local-global representations to reduce communication overhead. During training, a feature similarity modeling and fusion mechanism is employed to enhance structural alignment and information integration across heterogeneous modalities. In addition, a cross-modal attention module is introduced to effectively improve feature interaction and fusion at both local and global scales. Experiments on multiple joint CT and MRI segmentation tasks validate that WaveFed achieves superior segmentation performance compared to existing methods while maintaining high communication efficiency, demonstrating strong generalization and deployment potential.