Multiscale Hybrid Attention Network for Automated Segmentation of Colorectal Polyps

Accurate and reliable segmentation of colorectal polyps plays a critical role in the early diagnosis of colorectal cancer. Although deep learning–based methods have shown significant progress in recent years, the variations in polyp size, shape, color, and texture can adversely affect segmentation performance. In this study, a hybrid network model is proposed for automatic polyp segmentation. The proposed approach utilizes multi-level single-layer feature maps extracted from a DAF3D-based structure, which has demonstrated strong performance in deep feature extraction. These feature maps are first processed with a channel attention module and then with a reverse attention module to generate a multi-layer feature representation. While the channel attention module emphasizes the most informative channels, the reverse attention module enhances boundary regions, thereby improving segmentation accuracy. To evaluate the effect of these modules, scenarios using only channel attention, only reverse attention, and both modules together were compared, revealing that the integrated use of both modules achieved the highest performance. Additionally, the proposed model was compared with several polyp segmentation networks from the literature under the same backbone architecture and identical training conditions. Experiments conducted on multiple datasets demonstrate that the hybrid architecture provides significantly superior performance in mean Dice and mean IoU metrics compared to other methods. The findings indicate that the proposed approach is an effective, stable, and highly accurate method for polyp segmentation.