KC-UNet: Enhancing U-Net With KAN and CBAM for Medical Image Segmentation

Abstract

Medical image segmentation is a critical task in medical image analysis. However, traditional convolutional neural network (CNN) based methods are limited in modeling long-range dependencies, while Transformer-based segmentation models, though effective, suffer from high computational complexity due to quadratic attention operations. To address these challenges, this paper proposes an innovative U-Net variant, KC-UNet, which integrates Kolmogorov-Arnold Networks (KAN) and the Channel-Spatial Attention Module (CBAM). KC-UNet leverages the KAN representation theorem to represent features more efficiently, while CBAM enhances the model’s ability to adaptively capture both spatial and channel-wise dependencies, striking a balance between accuracy and computational efficiency. To validate the effectiveness of CBAM, this paper conducts comprehensive ablation experiments by replacing CBAM with Squeeze-and-Excitation (SE), and Efficient Channel Attention (ECA), as well as removing the attention module entirely. Results demonstrate that CBAM provides the most significant performance improvements in terms of segmentation accuracy, confirming its superior capability in enhancing feature representation. This study evaluates KC-UNet on four widely used benchmark datasets (BUSI, GLAS, CVC, and ISIC2017) and compare it against recent state-of-the-art models such as TransUNet, Swin-unet, and U-KAN. KC-UNet consistently achieves the best performance, with an IoU of 66.60% on BUSI, outperforming Swin-unet by 1.28%, and a Dice score of 80.46%, which improves upon the baseline U-Net by 7.44%. Similar advantages are observed on GLAS and ISIC2017, demonstrating the effectiveness and generalizability of our approach across different modalities. To the best of our knowledge, KC-UNet is the first framework to integrate KAN and CBAM for medical image segmentation.