<p>Fundus microvascular segmentation plays a crucial role in the intelligent diagnosis of diabetic retinopathy and glaucoma. However, existing methods often experience performance degradation due to complex vessel structures, fine-scale branches, low contrast in lesion regions, and significant scale variations. To address these challenges, we propose MFPANet, a multi-scale frequency-domain guided network for fundus microvessel segmentation. The proposed method integrates frequency–spatial collaborative modeling with a dynamic multi-scale feature fusion mechanism. Specifically, a depthwise separable convolution-based PAPEncoder is designed to extract hierarchical features while preserving high-frequency details, and a frequency-domain multiscale enhancement module is introduced to adaptively refine microvascular edge representations. In addition, multi-scale channel attention and hybrid attention mechanisms are employed to improve feature interaction across different semantic levels. Experimental results on the high-resolution FIVES dataset show that MFPANet achieves 74.66% mDice, 68.77% mIoU, and 70.21% F2-score under the overall evaluation setting. The corresponding sensitivity reaches 86.22% at the same scale, reflecting its performance in pixel-level microvascular detection. For small-scale vessels, the mDice increases to 87.21%, indicating improved segmentation performance on fine vascular structures. Ablation studies further indicate that the frequency-domain guided strategy reduces high-frequency information loss introduced by downsampling and yields a sensitivity improvement of 1.3 percentage points compared with the baseline configuration. These results suggest that the proposed method provides an effective solution for microvessel segmentation in complex fundus images.</p>