Abstract: Stroke MRI images are difficult to be segmented due to small lesion areas and blurred boundaries between lesions and healthy tissues. Therefore, this paper proposes an optimized encoder-decoder structure network. Firstly, in order to extract richer contextual information for the network, a dual-attention convolution fusion coding module is established, which combines 2D convolution with 3D convolution on the contraction path of the encoding end, and builds global correlation of the features from both space and channel dimensions. In addition, a residual-attention gate hybrid decoding module is proposed to better fuse low-level and high-level features and focus on the target area, thereby improving the segmentation fineness of the edges of the small lesions. The experimental results in the open source dataset anatomical tracings of lesions after stroke (ATLAS) show that the DSC index of the algorithm reaches 0.62. Compared with models such as UNet, D-UNet, 3D-UNet and attention-UNet, it demonstrates a much improved segmentation performance.