Improved unsupervised domain adaptation network based on category attention
Longhao Fan, Shouwei Gao, Fan Zhu, Z. A. Zhu, Chaozheng Zhou, Yaxin Peng · Journal of Physics Conference Series · 2021
Abstract Domain adaptation method can significantly reduce the distribution difference between images in variant domains, which plays an important role in unsupervised medical image segmentation. In this paper, an improved unsupervised domain adaptation framework is proposed based on category attention mechanism. The framework considers both image-level and feature-level alignment, and realizes semantic segmentation in different domains through a shared encoder-decoder. A category attention based classifier is proposed to compute the category attention feature and refine the semantic segmentation prediction. Sufficient experiments on MMWHS-2017 dataset indicate that the proposed method achieves the best segmentation performance among all comparasion algorithms.