MRI stroke lesion segmentation is crucial for accurate clinical diagnosis, however, it faces challenges such as low contrast between stroke lesions and normal tissues, unclear boundaries, and high inter-patient variability and so on. Denoising diffusion probabilistic models can capture fine details through iterative denoising, but they struggle with small targets and boundary clarity. In addition, existing network architectures often fail to effectively extract contextual and multi-scale information during the encoding process. This paper proposes a diffusion model enhanced with attention and contextual features for stroke lesion segmentation (DiffAC-Seg). Specifically, we propose a new U-Net-based denoising network. First, a three-pathway encoder is designed, which uses the original image boundary labels obtained by the Canny operator as explicit supervision. This integrates boundary, noise, and semantic information into the network, enhancing the guidance integration throughout the denoising process. Second, a feature fusion module is designed between the encoder and decoder, aiming to provide hierarchical global contextual information to the decoder by reconstructing skip connections. Additionally, a new efficient attention mechanism is proposed based on the quartet (Q, A, K, V) agent attention framework, where a composite function is used to replace the Softmax operator for computing the agent attention matrix. This composite function consists of two sub-functions, which are used to achieve non-negativity and nonlinear reweighting, respectively, thereby improving the internal processing speed of the model. Our model achieves Dice coefficients of 89.63% on the ISLES2022 dataset and 82.75% on the ATLAS dataset, demonstrating its effectiveness in stroke lesion segmentation.