Boundary-Guided Global-Local Feature Fusion Network for Polyp Segmentation

Colon polyp segmentation can assist physicians in screening colonoscopy images, which is crucial in preventing colorectal cancer. Due to the limited area occupied by small object polyp objects in images, there is a high risk of overlooking them, which makes it one of the more challenging aspects to address. Additionally, the current segmentation of small object polyps also faces difficulties such as boundary blurring, diverse lesion shapes, and uneven image brightness. While deep learning methods based on convolutional neural networks (CNNs) have been successfully applied to polyp segmentation tasks, three significant challenges persist: 1) limited ability to extract boundary information; 2) inadequate robustness capture of global context information; and 3) insufficient ability for integrating global and local information. To address the issues mentioned above, we propose the boundary-guided global-local feature fusion network for small polyp segmentation (BGGL-Net), with the following contributions: 1) the local information encoder (LIE) and boundary feature extraction module leverage convolutional blocks and Laplacian operators to mine feature boundaries and fine-grained detail features; 2) we design a global information fusion module to enhance the model's representational capacity and acquire rich and accurate global information; and 3) boundary-guided module (BGM) using a cross-attention mechanism, capturing the inherent relationship between the feature and establishes long-range dependencies between global- and low-level features. We enhance boundary accuracy by employing local boundary features to guide the global features, facilitating the effective fusion of global and local information. In the experiment, we compared ten state-of-the-art (SOTA) networks with BGGL-Net. The BGGL-Net achieves the highest segmentation accuracy on small object polyp datasets. Concerning generalization performance, the BGGL-Net outperforms CaraNet by up to 12.4% in the mDice metric on the ETIS-LaribPolypDB* dataset. Our code is available at https://github.com/shenjoyao/BGGL-Net.