DBG-Net: A Double-Branch Boundary Guidance Network for Polyp Segmentation

High-precision segmentation of polyps is crucial for timely screening and postoperative management of polyps. Although convolutional neural network (CNN)-based methods have made some progress in polyp segmentation research, they still face several bottlenecks in complex pathological environments: 1) insufficient feature extraction capability hinders the identification of polyps under different pathological morphologies and 2) inadequate boundary refinement ability, due to low differentiation between foreground and background, leads to inaccurate edge extraction. To address these issues, this article introduces a double-branch boundary guidance network (DBG-Net). Specifically, a dual-encoding path equipped with a dual-branch adaptive fusion (DAF) module is built to facilitate effective complementary fusion of both global and local features. Subsequently, a boundary guidance mechanism (BGM), consisting of a boundary learning module (BLM) and a boundary embedding module (BEM), is introduced to refine the edges of polyp regions. Additionally, we propose a multipath attention enhancement module (MAEM) that captures rich global context through different paths to enrich feature representation. Finally, a cross-scale feature aggregation decoder (CFAD) is constructed to fully integrate the efficacious information derived from the encoding part, enabling accurate feature reconstruction. To validate the effectiveness of the proposed DBG-Net, a series of experiments were conducted on multiple datasets, from which maximum Dice and mean intersection over union (mIoU) values as high as 93.48% and 93.30% were obtained, effectively substantiating the superior performance of DBG-Net and its potential as a polyp segmentation tool in complex pathological scenarios, thus bringing hope to postoperative management in clinical practice.