A parallel UNet integrating KAN and mamba for medical image segmentation

Abstract Medical image segmentation is fundamental for delineating lesion and organ boundaries in clinical workflows. While UNet-based models remain widely used, CNN-dominant designs are limited in modeling long-range context, and Transformer-based variants often introduce substantial computational overhead due to quadratic attention. To address this issue, we propose KMP-UNet, a parallel U-shaped framework that combines a Mamba-based state-space branch for linear-complexity contextual modeling and a Kolmogorov–Arnold Network (KAN) branch for nonlinear feature representation. We further introduce a task-oriented fusion block and a skip refinement module to better exploit hierarchical encoder–decoder features. KMP-UNet has a compact model size (about 1.0M parameters in our implementation). We evaluate the proposed method on four public datasets (ISIC2017, ISIC2018, CVC-ClinicDB, and BUSI) using standard segmentation metrics. On ISIC2018, KMP-UNet achieves 0.9038 DSC and 0.9600 accuracy under our protocol. Extensive comparisons and targeted ablations are conducted to analyze the contribution of each component.