Accurate and reliable medical image segmentation is crucial for improving diagnostic precision in clinical practice and establishing trustworthiness in medical Artificial Intelligence (AI) systems. However, existing methods suffer from limited generalization, insufficient interpretability, and a strong dependence on large-scale annotated data. To address these challenges, we propose IKANet, an interpretable medical image segmentation framework that leverages Kolmogorov-Arnold Networks (KANs) to enhance both performance and interpretability. We design a Dual KAN Convolution (DKC) module for feature disentanglement and a Multi-Scale Attention Module (MSAM) that enforces two-dimensional semantic consistency. Together, these modules enable precise feature representation, robust channel feature learning, and refined spatial lesion localization. Furthermore, we introduce an Interpretable Analysis Module (IAM) to align the decision-making process with the patterns of human diagnostic reasoning. Experiments on public datasets show that IKANet outperforms state-of-the-art models, improving IoU and Dice coefficients by 3%–10%. The IAM further provides compelling visual evidence of its decision process. IKANet provides an interpretable AI solution for medical image segmentation with the potential for reliable clinical diagnostics. Code is available at: https://github.com/cconut/IKANet .