A Robust Spatial‐Attention Kolmogorov–Arnold Network for Medical Image Segmentation

In recent years, medical image segmentation has witnessed marked progress driven by deep learning. Yet under high‐noise regimes, prevailing methods remain prone to boundary ambiguity and lesion miss‐detections, largely due to limited robustness to noise perturbations and an inadequate capacity to model spatial and channel‐wise dependencies. To address these limitations, we propose a robust segmentation framework that integrates spatial attention mechanisms with the Kolmogorov–Arnold Network: First, we incorporate a spatially adaptive feature modulation module, inspired by image super‐resolution, to dynamically fuse long‐range and local contextual cues, thereby improving resilience to input perturbations; Second, we devise a multi‐stage attention sampling architecture comprising channel‐split downsampling, multi‐dimensional attention upsampling, and refined skip connections to preserve salient lesion cues while suppressing noise amplification, thereby strengthening both spatial and semantic representations. Experiments on public datasets demonstrate that our method significantly outperforms state‐of‐the‐art counterparts on key metrics—including boundary delineation accuracy and lesion detection sensitivity—exhibiting superior robustness and practical utility.