Diffusion models have recently demonstrated remarkable potential in reinforcement learning (RL) owing to their superior distribution modeling capability and inherent multimodality. Unlike conventional unimodal policies (e.g., Gaussian policies) that often suffer from mode collapse and insufficient exploration in complex tasks, diffusion models generate actions through an iterative noise injection and denoising process, enabling natural representation of multimodal structures in the policy space. This property facilitates broader coverage of high-value solutions and mitigates the risk of converging to suboptimal modes. In this work, we propose the Selection-Augmented Diffusion Policy (SADP), which leverages the sampling capability of diffusion policies to generate multiple candidate actions at each state, and selects the final action via a probability distribution constructed from value function estimates, granting higher selection probabilities to actions with larger values. This probabilistic selection mechanism not only improves decision quality but also enhances exploration. The selected actions are subsequently used in conjunction with policy gradient optimization to enable efficient and stable training of the diffusion policy, effectively balancing exploration and exploitation during policy improvement. Experimental results show that SADP achieves superior performance in continuous control benchmarks and exhibits strong potential for high-dimensional decision-making and complex real-world scenarios.