Unsupervised domain adaptation via style adaptation and boundary enhancement for medical semantic segmentation

Abstract

The objective of semantic segmentation in cross-modal medicine is to align the distribution among different domains. The images from different domains contain various styles and boundary information, which the previous method ignores. Inspired by this observation, we considered employing the style and boundary information from features. We proposed a simple but effective way containing a style adaptation module and boundary enhancement module to facilitate the medical semantic segmentation on an unlabeled domain. In particular, the style adaptation module highlights the style information from the deep features through Fast Fourier Transform and low pass filter to assist in aligning the domain distributions. The boundary enhancement module utilizes the phase spectrum of features to obtain the boundary information and improve the segmentation ability by aggregating the boundary information. Results from experiments on two public datasets show that our proposed method effectively improves the segmentation performance of unlabeled target images and outperforms the most advanced domain adaptation approaches. Both qualitative and quantitative results evaluate the effectiveness of style features and boundary information in domain alignment.