To address the influences of factors such as different sizes, diverse shapes, and complex anatomical structures in multi-organ segmentation, a novel multi-organ segmentation method based on MFSA-Net (multi-scale frequency spatial attention network) is proposed. The network utilizes multi-level and multi-directional frequency decomposition to extract frequency-domain features of organs at different scales, which effectively expands the receptive field and enhances the discriminability of shallow semantic features. Moreover, a multi-level gated attention mechanism is introduced to achieve the integration of local fine-grained features and long-range dependencies, enabling the network to focus on critical regions while suppressing background noise. To address the structural diversity and variation among organs, a direction-enhanced dual-branch spatial attention module is designed to deeply integrate the spatial position and gray-scale distribution features of edge pixels, thereby improving the model’s spatial representation capability. Experimental results demonstrate that the proposed method can effectively segment multiple organs with large scale variations and complex structures, achieving average Dice similarity coefficient (DSC) scores of 81.66% and 91.61% on the Synapse and ACDC datasets respectively, which outperforms existing mainstream methods.