<p>Colonoscopy is an important method for the prevention and early detection of colorectal cancer, commonly used to detect polyps associated with colorectal cancer. However, accurate polyp segmentation still faces significant challenges: (1) polyps of the same type exhibit diversity in size, color, and texture; (2) the boundaries between polyps and the surrounding mucosa are often unclear. To address these issues, we propose a Deep Hybrid Attention Network (Deep-HybridUNet) that aims to improve the segmentation accuracy of polyps in colonoscopy images. Our method first employs a Hybrid Attention Module (HAM), which enhances segmentation performance by strengthening channel responses, extracting salient spatial features, and reinforcing boundary regions. Next, convolutional residual blocks are used in place of traditional double convolution layers to simplify the feature propagation path. Finally, by introducing a Depthwise Pooling Module (DPM), we excavate deeper information, thereby improving segmentation accuracy and detail restoration capabilities. The experimental results demonstrate that Deep-HybridUNet significantly outperforms existing mainstream methods across several key performance metrics. On the CVC-ClinicDB dataset, it tackles the dual challenges of accurately segmenting irregularly shaped polyps and mitigating artifact interference, achieving an IoU of 0.8592, accuracy of 0.9901 and F1 score of 0.9394. Notably, the model excels in cross-domain adaptability, with IoU and Dice scores improving to 0.8910 and 0.9286, respectively, on the ISIC2018 dataset. These results not only validate the effectiveness of the model architecture, but also highlight its powerful feature representation capabilities and cross-domain generalization, offering a novel technological pathway for the universality of medical image segmentation.</p>