In recent years, deep learning-based methods have been wide- ly adopted in the field of medical image segmentation and have made great progress. However, when processing medical images with complex anatomical structures and blurred boundaries, existing CNN and Transformer architectures still suffer from a disconnect between local and global feature modeling, limited capacity for deep nonlinear semantic representation, and poor adaptability to structural heterogeneity. To address these issues, we propose a novel U-Net variant framework named Mamba-KANet, which integrates spatial state modeling capabilities and functional expression flexibility to enhance segmentation performance in challenging structures and blurry boundary regions. Specifically, we design a SwinMamba module that facilitates spatial semantic fusion from local details to global context through local feature perception and cross-window contextual interaction mechanisms. A state modeling strategy is further introduced to improve the understanding of dynamic structures. In the bottleneck layer, we introduce the Bottleneck KAN (B-KAN) module, which replaces conventional convolution with learnable one-dimensional spline functions to strengthen the model’s ability to capture and fit complex nonlinear semantic patterns. The synergy of these two modules results in a segmentation framework with strong expressiveness and generalization capability. Experimental results on the Synapse multi-organ CT and ACDC cardiac MRI datasets demonstrate that Mamba-KANet outperforms other methods in terms of performance, demonstrating its strong potential for practical application in clinical medical image analysis.