Accurate segmentation of retinal blood vessels plays a vital role in diagnosing various eye diseases. It not only aids in identifying conditions such as diabetes, arteriosclerosis, and cardiovascular diseases but also enhances clinicians′ ability to diagnose and treat patients effectively. While existing convolutional neural network (CNN) approaches excel at capturing local spatial features through convolutional operations, they face challenges in extracting global spatial information. Conversely, frequency domain methods can capture the overall spectral distribution and global structural features of images but struggle to precisely locate local details and preserve high-frequency information due to spatial information blurring during frequency transformation. To address these limitations, this study proposes a retinal blood vessel segmentation method based on spatial-frequency domain fusion, leveraging the strengths of both domains for local and global feature extraction. The approach features a dual-branch spatial-frequency feature extraction and fusion module in the encoding stage, designed to integrate frequency and spatial features and mitigate detail loss during downsampling. Additionally, a multi-scale Gaussian filter is incorporated to enhance the model′s capability in accurately locating vessel boundaries and preserving continuity of small vessels. Finally, an adaptive spatial-frequency fusion module dynamically calculates fusion weights across feature map regions, improving the precision of small vessel segmentation. Experiments conducted on two widely-used open-source datasets, DRIVE and CHASE_DB1, demonstrate accuracy rates of 96.9% and 97.81%, respectively. Results indicate that the proposed method achieves competitive performance in segmentation accuracy, consistency of small vessel detection, and robustness in handling lesions.