MPKU-Net: A U-Shaped Medical Image Segmentation Network Based on MLP and KAN

Abstract

The UNET architecture has been widely adopted for image segmentation across various domains, owing to its efficient and powerful performance in recent years. Its application and enhancement in medical image segmentation primarily involve convolutional neural network (CNN) and Transformer. However, both methods have fundamental limitations. CNN struggle to capture global features, which greatly reduces the computational complexity but compromises its effectiveness. Transformers excel at capturing global features but demand substantial parameters and computations and fail to effectively extract the local features. To address these challenges, we propose a U-shaped network model, MPKU-NET, which integrates a multilayer perception (MLP) with a Knowledge-Aware Networks (KAN) network architecture, aiming to effectively extract both local and global characteristics in a coordinated manner. MPKU-NET features the flexible rolling Flip operation that, along with MLP and Knowledge-Aware Network (KAN), creates the WE-MPK modules for thorough learning of global and local features. Its effectiveness is proven by extensive testing on the BUSI, CVC, and GlaS datasets. The results demonstrate that MPKU-Net consistently outperforms several widely used segmentation networks, including U-KAN, Rolling-U-net, U-Net ++, in terms of both model parameters and segmentation accuracy, highlighting its effectiveness as a scalable solution for medical image segmentation. The network model code has been uploaded: https://github.com/cp668688/MPKU-Net.