An improved Segformer Method for Polyp Segmentation in Digestive Endoscopy

Accurate and real-time segmentation of polyps in digestive endoscopy is critical for improving the diagnosis and treatment of gastrointestinal diseases. Although Transformer-based models like SegFormer achieve competitive performance in terms of accuracy and computational efficiency, they often fall short in capturing fine-grained polyp boundaries and handling complex morphological variations. This paper presents an enhanced SegFormer framework that integrates a cross-stage attention mechanism, and a UPerHead-based decoder. These improvements facilitate robust multi-scale feature fusion and refined edge localization. Extensive experiments on the Kvasir-Seg dataset demonstrate that the proposed model outperforms the baseline SegFormer in segmentation accuracy. The method also shows strong adaptability on unseen datasets such as CVC-300 and CVC-ClinicDB, indicating its potential for real-world clinical application.