Accurate medical image segmentation is essential for clinical diagnosis, disease monitoring, and surgical planning. However, convolutional neural networks (CNNs) are constrained by limited receptive fields, making it difficult to capture long-range dependencies, while Transformer-based methods incur quadratic computational costs and substantial parameter overhead as the input resolution grows. To address these challenges, we propose AxialUNet, a lightweight U-shaped architecture that efficiently models global context while preserving the native 2D spatial structure. At its core, the Multi-scale Spatial Mixing (MSSM) module integrates axial mechanisms with multi-scale convolutions to capture fine-grained local details and long-range dependencies in parallel. Additionally, a Dynamic Heterogeneous Expert Attention (DHEA) module is introduced in the skip connections, leveraging a Mixture-of-Experts system that also utilizes axial decomposition for precise boundary refinement. Extensive experiments on multiple benchmarks show that AxialUNet, with only 5.14M parameters and 11.83 GFLOPs, achieves performance comparable to state-of-the-art methods, establishing it as an effective and efficient solution for medical image segmentation.