Domain Adaptation of Foreground and Scale Sensing for Gastric Polyp Detection

Abstract

Automated detection of gastric polyps has been proven crucial for improving diagnostic accuracy. However, when there is a domain shift in the data, deep learning-based detection methods may not perform well. Unsupervised domain adaptation has been demonstrated as a good approach to address this issue. However, existing unsupervised domain adaptation detection methods struggle to handle the problem of foreground–background similarity and the diverse appearances of polyps at different scales in gastric polyp images. In this paper, we propose a boundary-guided transferable attention module and a transferable prototype alignment module to address the foreground–background similarity issue, and a multi-scale enhanced alignment method to tackle the problem of information loss when aligning polyps at multiple scales. The boundary-guided transferable attention module fully explores spatial information of the image with a boundary-guided multi-field attention mechanism while considering the transferability of features to mine the easily transferable foreground regions. The transferable prototype alignment module adopts a prototype-based method to facilitate the transfer of difficult-to-align regions. The multi-scale enhanced alignment method prevents information loss across feature maps and scales with an attention filtering module, enhancing features at each scale. In experiments, this work outperforms advanced domain adaptation detection methods like SIGMA and CAT in polyp detection.