For the endoscopic surgical instrument segmentation task, existing methods have failed to address the semantic gap caused by the mismatch between high-frequency spatial details and low-frequency semantic features in the U-shaped network (U-Net) architecture. This study proposes a U-Net algorithm based on frequency-domain adaptive feature decomposition and visual Mamba, namely frequency-domain decoupling Mamba U-Net (FDMUNet), for surgical instrument segmentation. This algorithm embeds a frequency-domain adaptive feature enhancement module into the skip connections, and decomposes features into high-frequency and low-frequency components through the Fourier transform and a learnable filter, followed by channel weighting and fusion, so as to enhance surgical instrument edge information and bridge the semantic gap between encoder and decoder features. On the Endoscopic Vision Challenge 2017 public dataset, FDMUNet achieved Intersection over Union, mean Intersection over Union, and mean class Intersection over Union scores of 70.79%, 74.25%, and 69.50%, respectively. In addition, the ablation experiment further verified the effectiveness of the proposed module. This method not only provides a new solution for instrument segmentation in complex scenes, but also provides a new research idea for the application of frequency-domain information in medical image segmentation.