SKANet: Spatial-Kernel Attention Network for Efficient and Interpretable Medical Image Segmentation

Accurate polyp segmentation in colonoscopy images remains a critical yet challenging task, driven by the need to handle diverse polyp morphologies, subtle boundary transitions, and real-time clinical constraints. We propose SKANet, a hybrid architecture that synergizes the spatial encoding strengths of PVTv2 with nonlinear attention modeling via Kolmogorov-Arnold Networks (KANs). SKANet incorporates two key innovations: (1) a multi-scale Spatial-Domain Interaction (SDI) module that enhances semantic coherence through attention-guided feature fusion, and (2) a lightweight KAN bottleneck leveraging depthwise separable convolutions and spline activations to efficiently capture long-range dependencies. To assess its generalization and clinical relevance, we conduct extensive experiments on the widely used Kvasir-SEG, ETIS-Larib, CVC-ClinicDB, CVC-ColonDB and CVC-300 datasets, demonstrating SKANet’s consistent superiority over both CNN- and transformer-based baselines. SKANet achieves Dice scores up to 94.5% and IoU scores up to 89.8%, with an average Dice of 88.5% and IoU of 82.1% across all datasets. It maintains high computational efficiency (121 FPS on RTX 3090) and a compact footprint (7.2M parameters), underscoring its suitability for real-time, interpretable polyp segmentation in clinical endoscopic workflows.