Biomedical image segmentation plays a crucial role in aiding diagnosis and treatment planning. However, constructing effective frameworks remains challenging due to the variable size and irregular shape of target structures. U-Net has become a cornerstone in this field, but integrating it with Transformer or Multilayer Perceptron (MLP) models faces limitations such as quadratic computational complexity and insufficient interpretability. To address these challenges, we propose KM-UNet, a novel structure inspired by state-space models (SSMs) (e. g. , Mamba) and the Kolmogorov-Arnold network (KAN). KM-UNet leverages nonlinear, learnable activation functions, rooted in the Kolmogorov-Arnold representation theorem, to enhance interpretability and efficiency. By combining the strengths of state-space models and Kolmogorov-Arnold networks, KM-UNet achieves a balance between accuracy and computational performance. Experiments on five public datasets demonstrate its superiority. On the BUSI dataset, KM-UNet achieved an Intersection over Union (IoU) of 65. 21% and an F1-score (F1) of 78. 43%, improving IoU by 1. 83% over state-of-the-art methods. It also achieved the highest IoUs on the Glas (87. 31%) and CVC (85. 22%) datasets and delivered the best overall performance on the ISIC series datasets. These results highlight KM-UNet’s ability to effectively integrate global and local information while maintaining interpretability. With its powerful feature extraction capabilities and computational efficiency, KM-UNet emerges as a versatile and reliable framework for medical image segmentation across diverse biomedical applications. The source code is available at https: //github. com/2760613195/KMUNet.
Zhang et al. (Tue,) studied this question.