In the field of medical image analysis, deep learning, especially convolutional neural networks (CNNs), has developed rapidly. However, it still has limitations in capturing long - range dependencies. To obtain more accurate boundaries and finer - structured segmentation results, this paper proposes a powerful boundary - aware segmentation network, BEA - UNet. This model optimizes the boundary feature extraction pattern using convolutions with multiple dilation rates and addresses the long - range dependency problem with the help of a dual attention mechanism. On this basis, the boundary loss and the segmentation loss are integrated to obtain a new loss function for optimizing the model parameters. Experimental results on multiple datasets such as ACDC, BUS, and BUSI show that BEA - UNet performs excellently in various medical image segmentation tasks. This research provides a new and powerful framework for medical image segmentation tasks, and is expected to strongly support and contribute significantly to clinical medical image analysis. The code is available at https://github.com/ychAlbert/BEA-UNet for academic exchanges and applied research.