DDKVM-UNet: Dual-Domain KAN with vision mamba U-Net for medical image segmentation

Medical image segmentation plays a core role in clinical applications such as robotic-assisted surgery, disease diagnosis, and treatment planning. Although convolutional neural networks (CNNs) and Transformers have achieved significant progress, both have limitations: CNNs struggle to model long-range dependencies, while Transformers incur prohibitive computational costs due to the self-attention mechanism. Recently, the vision Mamba architecture integrated with KANs has provided new insights for this field with its efficient global modeling capability. This study proposes a Dual-Domain KAN with Vision Mamba U-Net (DDKVM-UNet), introducing the Visual State Space (VSS) module and a Dual-Domain KAN fusion (DDKF) module to achieve complementary learning of global context and nonlinear details. The VSS module adopts the 2-D Selective Scan (SS2D) strategy to capture long-range dependencies; the DDKF module deploys Tokenized-KAN in dual domains to accurately extract lesion edges and texture features. In addition, this paper proposes the Direction-Aware Pixel Attention (DPA) mechanism, which guides pixel-level feature calibration through direction-sensitive priors to achieve efficient modeling of direction-dependent structures. Evaluations on four benchmark datasets demonstrate that DDKVM-UNet improves the Dice Similarity Coefficient (DSC) and Intersection over Union (IoU) over U-Net by 3.88% and 4.49% on the ISIC2017 dataset, respectively, validating its superior performance in medical image segmentation.