BSFP-Net: Polyp Segmentation Based on Boundary Supervision and Feature Propagation

Colorectal cancer (CRC) is one of the most prevalent malignant tumors worldwide, and early screening with precise polyp segmentation is crucial for reducing CRC-related mortality. However, existing methods still exhibit limitations in leveraging boundary information and multi-scale feature fusion, which restricts segmentation accuracy. To address these challenges, this paper proposes a Boundary-Supervised Feature Propagation Network (BSFPNet) for polyp segmentation, aiming to enhance segmentation accuracy by precisely extracting and utilizing edge semantic information in conjunction with multi-scale feature fusion. The primary contributions of this work are as follows: (1) the improvement of skip connections in the inner U-shaped network by introducing a multi-scale feature propagation mechanism, thereby enhancing the model's attention to irregular polyp targets; and (2) the incorporation of global and local context fusion mechanisms within the inner U-shaped network to improve segmentation accuracy for polyps of varying sizes. Experimental results demonstrate that BSFPNet achieves superior segmentation performance on multiple polyp segmentation datasets (e.g., Kvasir-SEG and CVC-ClinicDB), particularly excelling in edge clarity and small target segmentation compared to existing methods.