Coronary artery segmentation is essential for diagnosing coronary artery diseases, yet varying branch thickness and confusing noises pose significant challenges. This paper presents SF-UNet, a progressive segmentation approach featuring a Global Suppression Gate (GSG) to spatially cluster potential vessel regions and suppress distant noises. Each decoding layer comprises Selective Fusion Gates (SFG) and a Gate Compensation Mechanism (GCM) for vessel refinement and efficient training. The SFG enhances detail extraction and minimizes localized noise, dynamically refining regions from the GSG. Experiments on the DCA1 dataset show SF-UNet achieved top scores in IoU(66.51%), DSC(79.77%). On the CHUAC dataset, it outperformed most methods across all metrics, including IoU(64.89%), and DSC(78.57%). These results demonstrate SF-UNet’s performance, offering an effective solution for coronary artery segmentation.