SGFNet: An Attention-Enhanced Network for Colonic Polyp Segmentation in Endoscopic Images

Abstract

Accurate segmentation of colonic polyps in endoscopic images is vital for early diagnosis and prevention of colorectal cancer (CRC). However, this task remains challenging due to the high variability in polyp morphology, low contrast with surrounding mucosa, and frequent boundary ambiguity. To address these issues, we propose a selective-gated fusion network (SGFNet), a lightweight semantic segmentation network based on the U-Net architecture. SGFNet incorporates two targeted modules: the SSE-Encoder, which integrates selective kernel convolution (SKConv) and squeeze-and-excitation (SE) attention to enhance multi-scale representation and channel recalibration; the PGF-Unit, which employs a gated parallel polarized self-attention (PPSA) mechanism to improve boundary-sensitive feature fusion. The model is evaluated on a unified dataset of 1 612 annotated images from Kvasir-SEG and CVC-ClinicDB. Experimental results demonstrate that SGFNet achieves superior performance in mean intersection over union (mIoU), pixel accuracy (PA), and F1 score, outperforming several representative baseline models. With its balanced design and strong generalization, SGFNet offers a practical and interpretable solution for real-world medical image segmentation and provides insights into attention-based module integration for future computer-aided diagnosis systems.

Keywords

colonic polyp segmentation / U-Net / multi-scale feature encoding / attention mechanism / SGFNet / medical image analysis