Kolmogorov-Arnold Guided Local-Global Attention for Medical Image Classification

Medical image classification relies on both fine lesion details and broader anatomical context to support reliable clinical decisions. Existing attention mechanisms often capture only one of these aspects, resulting in unstable and incomplete representations. We propose KA, a lightweight, attention-inspired nonlinear interaction module built on Kolmogorov-Arnold operators, in which two complementary pathways work together to balance local and global feature modeling. Specifically, KAN Local Attention Module (KLAM) enhances local structures through nonlinear modeling within grouped windows, while KAN Adaptive Mixer (KAM) integrates global semantics using spline-based adaptive fusion. Together, these components provide a balanced and efficient mechanism for local and global feature extraction and can be applied to CNN, Transformer, and Mamba architectures. Experiments on three public clinical datasets show that KA consistently improves performance with limited parameter and FLOP overhead, demonstrating its effectiveness for medical image classification.