Frequency-based boundary-guided attention network for domain generalizable polyp segmentation from colonoscopy images

Colonoscopy is the most effective method for detecting colorectal polyps and preventing colorectal cancer. The accurate segmentation of polyps in colonoscopy images is crucial for diagnosis and surgery, which remains a challenge in various clinical settings. In particular, poor boundary details can significantly degrade performance when segmenting polyps from colonoscopy images in an unobserved clinical setting, such as unexpected reflection by water, complex polyp shapes, and low contrast between the polyp and surrounding mucosa. To address this discrepancy, this study proposes a domain generalizable Frequency-based Boundary-guided Attention Network (FBGANet). The proposed model has high polyp detection ability in various domains, as we remove the noise occurring in the images from Out-domain data by applying a DCT-based component decomposition module (DCT CDM) to extract the feature maps. However, because the boundary information for detailed polyp segmentation mainly exists at high frequencies, we propose a Boundary-guided Attention Block (BGA Block), which preserves the polyp boundary details for higher confidence in unseen domains. Extensive experiments on various benchmark datasets showed that FBGANet exhibited a better domain generalization ability than other state-of-the-art methods with a reasonable inference speed (0.035 s/image). In particular, FBGANet exhibited higher Dice Score Coefficient (DSC) and mIoU than MADGNet and CFATransUNet in the In-domain and Out-domain datasets. This experiment results demonstrate that FBGANet enables precise detection, improving patient outcomes and advancing robotic-assisted healthcare. Our code is available in Github Link.