Abstract: To address the insufficiency of local and global context feature modeling capabilities and the weak learning ability of target morphological features in the existing improved U-Net for pathological image segmentation tasks, we propose an enhanced morphological U-Net (MKU-Net) for pathological image segmentation. In the downsampling stage, we propose a multi-scale channel fusion module, a dual-path channel and spatial attention enhancement module, and build a hierarchical structure by using a modular approach which enhances the network's local feature expression ability and global information understanding ability without significantly increasing the number of parameters. In the skip connection stage, we propose a morphological feature enhancement module integrating the Kolmogorov-Arnold representation theorem. The module designs the structural element as a learnable parameter and combines the B-spline method in the representation theorem to simulate morphological operations, thereby enhancing the learning ability of complex morphological features. In the upsampling stage, we introduce content-aware feature rearrangement upsampling, and propose an edge enhancement module based on the Euclidean distance threshold classification method. The module fully utilizes the complementarity of segmentation information and edge information to enhance the edge features of the feature map, thereby improving the completeness of edge segmentation. Experiments show that the performance of MKU-Net on multiple pathological image datasets is superior to most competing networks, with Dice coefficients of 92.85% and 87.16% on the GlaS dataset and CoCaHis dataset, respectively.