<p>Medical image segmentation remains challenging due to the difficulty of simultaneously capturing long-range contextual dependencies and preserving fine structural boundaries, particularly in low-contrast and small-target scenarios. Although recent studies have explored the integration of foundation models and attention mechanisms, most approaches treat them as loosely coupled components, leading to limited performance gains. To address this issue, we propose Multi-scale Kernel Selection Fusion Attention Network (MKSFA-Net), a novel framework that introduces a structural prior-guided feature learning paradigm together with a cascaded multi-scale cooperative attention mechanism. Specifically, SAM-Med2D is employed to generate coarse segmentation masks, which are not merely used as auxiliary inputs but serve as explicit structural constraints to guide feature representation learning. To further enhance segmentation accuracy, we design a two-stage attention module composed of Multi-scale Fusion Attention (MFA) and Multi-scale Coordinate Attention (MCA). Unlike conventional attention designs, the proposed modules operate in a coarse-to-fine collaborative manner, where MFA aggregates multi-scale contextual information and MCA performs direction-aware spatial refinement, enabling progressive feature enhancement. Extensive experiments on private breast cancer CT/MR datasets and the public MSD Heart dataset demonstrate that MKSFA-Net consistently outperforms state-of-the-art CNN-based, Transformer-based, and hybrid segmentation models across multiple metrics, including Dice similarity coefficient (DSC), Intersection over Union (IoU), and 95% Haus Dorff Distance (HD95). The results validate the effectiveness of the proposed structural prior-guided learning strategy and the synergistic multi-scale attention design for robust and precise medical image segmentation.</p>