Research on Polyp Segmentation Method Using PANet Based on Contrastive Learning

Abstract The early screening of colorectal cancer primarily relies on the identification and removal of polyps during colonoscopy. However, the high similarity between polyps and surrounding tissues, as well as the variability in their morphology and size, poses significant challenges to automated and accurate segmentation. This paper proposes a polyp-aware network named PANet based on contrastive learning to improve the accuracy and robustness of polyp segmentation. The key innovations of PANet include: 1) A contrast-learning-based backbone network ContFormer, which integrates the global dependency modeling capability of Transformers with a supervised contrastive learning strategy to effectively learn a highly structured feature space for accurately locating camouflaged polyps; 2) A scale-aware enhancement module SAEM that adaptively aggregates multi-scale features to handle polyps of different sizes; 3) A semantic-detail fusion module SDFM that progressively injects high-level semantic information into low-level features, generating high-resolution, semantically rich feature maps for precise segmentation. Extensive experiments were conducted on five public polyp segmentation datasets, including Kvasir-SEG and CVC-ClinicDB. The results show that PANet achieves excellent performance in leave-one-dataset-out generalization tests. For instance, on the highly challenging ETIS-LaribPolypDB dataset, it attains an mDice score of 0.807, significantly outperforming existing state-of-the-art methods such as UNet, C2FNet, and ColonFormer. Visualizations further confirm the robustness of PANet in handling small polyps and complex backgrounds. This study provides an efficient and reliable polyp segmentation solution for computer-aided diagnosis systems in colonoscopy.