A retinal vessel segmentation algorithm based on multi-level semantic fusion and feature-coupled Transformer was proposed to address challenges such as the difficulty in extracting fine vessels, low imaging contrast, and interference from lesion information in retinal images. First, a column nonuniformity correction module was used to construct a dual-branch joint feature extraction module, which effectively preserved vessel texture information and enhanced the model’s ability to extract fine vessels. Then, a feature-coupled Transformer module was introduced at the encoder-decoder connection to enhance the representation of vessel features, enabling more accurate recognition of vascular semantic information. Finally, a multi-level semantic fusion module was added to the encoder to suppress background noise interference and focus on vessel-related features. Experiments were conducted on public datasets DRIVE, STARE, and CHASE_DB1. The sensitivity achieved was 80.30%, 80.84% and 82.43%, respectively, the accuracy reached 97.11%, 97.61% and 97.63%, respectively, and the F1-scores were 82.96%, 83.76% and 81.48%, respectively. The experimental results indicate that the proposed method achieves superior segmentation accuracy, preserves the integrity of fine vascular structures, and effectively handles complex lesion regions; overall performance surpasses that of most existing advanced methods and exhibits strong generalization and robustness, thereby providing more reliable technical support for intelligent auxiliary diagnosis of retinal vascular diseases.