In medical domain, the scarcity and sensitivity of imaging data present significant challenges to the development of accurate and privacy-preserving segmentation models. In this paper, an Optimized Training framework with Trustworthy Enhanced Replication (OTTER) via diffusion and federated VMUNet is proposed to solve these problems. OTTER leverages diffusion-based generative models to synthesize diverse and high-fidelity medical images that reflect the statistical distribution of real data while avoiding direct exposure of patient information. These synthetic samples are further filtered using a Bayesian quality estimator to ensure training on only reliable and representative data. These trusted replicas are integrated into a federated learning pipeline built on VMUNet, a state-space-based segmentation architecture that enables decentralized training across medical centers without sharing raw data. Experiments on multiple public medical image segmentation datasets have demonstrated that OTTER achieves competitive or superior performance in terms of accuracy, robustness, and generalization, while maintaining strong privacy guarantees. Our results highlight the potential of combining generative data augmentation with federated architectures to advance privacy-aware medical AI.