Dynamic Context-guided Feature Fusion Network for gastrointestinal tumor image segmentation

Precise segmentation of gastrointestinal tumor images holds significant importance in clinical diagnosis and prognostic assessment. However, it faces two major challenges. First, the blurred boundaries between lesion regions and normal tissues often lead to over-segmentation or missed detections by models. Second, the subtle variations in color and the minimal foreground–background contrast hinder the ability of models to accurately differentiate lesions from the surrounding tissues. To overcome these challenges, we propose a two-stage gastrointestinal tumor image segmentation algorithm, dubbed Dynamic Context-guided Feature Fusion Network, which incorporates two key modules. Specifically, a boundary-aware semantic decoupling module employs a cascaded design to explicitly decompose deep features into foreground, background, and uncertain boundary regions, while introducing a boundary-aware loss function to improve boundary recognition in uncertain regions. Moreover, a dynamic context-guided feature fusion module leverages global contexts to adaptively emphasize foreground features and reduce false positives from surrounding tissues, significantly improving segmentation accuracy, thereby assisting clinicians in making more accurate diagnostic decisions. Experimental results on SEED, CVC-ClinicDB, and Kvasir-SEG datasets show that the proposed model delivers state-of-the-art performance in both quantitative and qualitative aspects. • BSDM combines feature decoupling and boundary loss to enhance edge perception. • DCFFM adapts global context to stress foreground and reduce background noise. • DCFF-Net achieves state-of-the-art segmentation on gastrointestinal tumor datasets.