GSFDoubleU ‐Net: A Dual‐Branch Architecture Incorporating Gated Self‐Attention Fusion for Enhanced Gastric Polyp Image Segmentation

Abstract

ABSTRACT Gastric polyps are clinically important lesions whose accurate delineation in endoscopic images can support lesion assessment and treatment planning. However, small size, weak contrast, irregular morphology, indistinct boundaries, and interference from surrounding gastric tissues make automatic segmentation difficult. To address these challenges, this study proposes GSFDoubleU‐Net, a lightweight dual‐branch architecture for gastric polyp segmentation. The first branch employs a pretrained VGG19 encoder and a Lightweight Atrous Spatial Pyramid Pooling (L‐ASPP) module to capture multi‐scale contextual information, whereas the second lightweight encoder–decoder branch focuses on contextual perception and fine‐grained feature recovery with limited computational overhead. A Gated Self‐Attention Feature Fusion (GSFF) module is further introduced to integrate complementary dual‐branch features, model long‐range dependencies, suppress irrelevant background responses, and refine lesion boundaries. On CVC‐ClinicDB, GSFDoubleU‐Net achieved a Dice score of 0.957 and a Precision of 0.970; on Kvasir‐SEG, it obtained a Dice score of 0.949, an mIoU of 0.898, and a Precision of 0.953. When evaluated on the independent internal clinical dataset, the model achieved a Dice score of 0.857 and an mIoU of 0.832, while obtaining lower HD and MSD values than the compared methods. These results indicate that GSFDoubleU‐Net provides a favorable balance between segmentation accuracy, boundary delineation, and computational efficiency, and may serve as a useful framework for computer‐assisted endoscopic image analysis.