SpaSA-Net: Multi-Scale Sparse Attention-Guided Spline-Activated Network for Efficient Medical Segmentation

Medical image segmentation is a key task in computer-aided diagnosis; however, traditional convolutional neural networks (CNNs) are limited in modeling long-range dependencies and complex spatial relationships in high-resolution medical images. Although vision Transformers partially alleviate the challenge of modeling long-range dependencies, their quadratic computational complexity and potential redundancy in the global attention mechanism significantly limit their practical efficiency. To address these limitations, we propose SpaSANet. Its KAN-Conv Parallel (KConvP) module integrates Kolmogorov-Arnold Networks (KAN) and convolutional layers in parallel during the U-Net encoder stage, utilizing KAN's spline-based learnable activation mechanisms to replace fixed activation functions, thereby enabling more flexible and interpretable nonlinear mappings. Meanwhile, the KAN-based Multi-Scale Sparse Attention (KMSSA) combines dilated convolution expansion and multi-scale window mechanisms to efficiently capture local contextual information and model spatial dependencies in a sparse manner, thereby reducing redundant computation and focusing attention on the most informative regions. Experimental results on the ISIC2017 and ISIC2018 datasets demonstrate that, compared to the baseline U-Net and its variants, the proposed model achieves significant improvements in segmentation performance.