Tubular structure segmentation is essential across various domains, including medical imaging, remote sensing, and industrial inspection. While traditional models rely on expert knowledge for feature extraction, deep learning offers a more efficient approach. However, existing deep neural networks still have room for enhancement. This paper introduces an enhanced U-Net with a cascaded shared dilated convolution module, which broadens the receptive field and retains finer image details. Additionally, a cascaded multi-scale dice loss fusion module is incorporated to leverage diverse scales of structural information, enhancing prediction accuracy. Experiments on public datasets and comparisons with mainstream models demonstrate our model’s superiority, particularly in perceiving tubular structures. Our model achieves mIoU of 84.3 and 81.4 on the DRIVE dataset and the Massachusetts Roads dataset, respectively. The research advances applications like retinal vessel detection, urban road mapping, and industrial defect identification.