EAAI Journal 2026 Journal Article
- Xiaojuan Liu
- Xuan Li
- Zhi Liu
- Ke Peng
- Shanxiong Chen
- Yijue Zhang
- Bo Hou
Colorectal cancer is a highly preventable disease. Colonoscopy allows for the detection and removal of polyps, which enables early diagnosis and timely intervention. In clinical practice, automatic polyp segmentation techniques based on colonoscopy images can improve both detection efficiency and accuracy, while helping physicians accurately locate polyps. However, existing methods still have limitations in coordinating global and local features, fusing multi-scale features, and handling ambiguous boundaries. To address these challenges, this paper proposes the Global-to-Local Deep Interaction and Boundary-Aware Transformer. This approach incorporates a Global–Local Aggregation module to coordinate fine-grained details with semantic information; employs a Multi-Scale Residual Decoder to enhance cross-layer feature fusion efficiency; and introduces a Dynamic Feature Fusion Module comprising Hierarchical Fusion and Error-Aware Refinement, for adaptive optimisation in boundary regions. We conducted extensive experiments and comparative analyses on five publicly available polyp datasets, evaluating our model against 15 state-of-the-art approaches. To further validate the model’s generalisation capabilities, we also designed experiments targeting small polyps. The experimental results show that our method performs well across multiple datasets, especially with mean Dice scores of 93. 9% on CVC-ClinicDB and 84. 7% on ETIS-LaribPolypDB.