Denoising Diffusion Probabilistic Models (DDPMs) have garnered significant interest due to their notable capabilities in image generation. The models could used for lesion segmentation in medical images. However, a prevailing challenge with these models is that they fail to capture global information. Our proposed MDADiff addresses this limitation. Drawing inspiration from the transformer paradigm, MDADiff employs a novel approach for global information capture. Furthermore, by integrating Multi-scale Dual Attention, MDADiff effectively handles semantic information, offering more detailed segmentation insights and accelerating convergence. Importantly, we also obtained the uncertain map for each test slice. Compared with the original diffusion models, our proposed model performs achieves superior performance on the brain tumor segmentation dataset Brats2021.