MedKamba: A Novel Approach Integrating State-Space Models and Fractional Kolmogorov–Arnold Networks for Medical Image Segmentation

Abstract

Medical image segmentation plays a crucial role in healthcare, serving as a key step in disease diagnosis and treatment planning. Convolutional neural networks (CNNs) are limited by their restricted receptive fields, whereas Transformer-based models suffer from quadratic computational cost. Recent advances such as Mamba, a selective state-space model with linear complexity, and its vision-oriented variant, the Visual State Space (VSS) models, have shown strong ability to capture long-range dependencies efficiently. However, they still exhibit shortcomings in segmentation tasks, including loss of pixel-level structural information and inefficient channel utilization. To address this, we introduce VSSM-based Local Aware Channel Enhancement (LACE) block, which incorporates local enhancement and channel attention to better preserve spatial detail. To this end, we proposed MedKamba, a novel U-shaped segmentation approach that employs a hybrid encoder with CNNs and LACE blocks to effectively capture both local and global contextual information. While the U-Net backbone remains highly efficient, its traditional skip connections rely on simple scale-matched fusion, limiting cross-scale interaction. To overcome this, we redesign the skip connections using Fractional Kolmogorov–Arnold Networks (f-KANs) to generate channel-wise attention weights from features aggregated across multiple stages. Experiments on two benchmark datasets demonstrate that MedKamba consistently outperforms competing approaches and produces more visually accurate segmentation results.