Unsupervised Domain Adaptation for Medical Image Segmentation by Selective Entropy Constraints and Adaptive Semantic Alignment

Generalizing a deep learning model to new domains is crucial for computer-aided medical diagnosis systems. Most existing unsupervised domain adaptation methods have made significant progress in reducing the distribution gap through adversarial training. However, these may still produce overconfident but erroneous results on unseen target images. This paper proposes framework cross-modality image segmentation. Specifically, We first introduce two data augmentation approaches generate sets of semantics-preserving augmented Based model's predictive consistency images, we identify reliable and unreliable pixels. then perform selective entropy constraint: minimize pixels increase their confidence while maximizing reduce confidence. identified pixels, further propose an adaptive semantic alignment module which performs class-level by minimizing distance between same class prototypes domains, where are removed derive more accurate prototypes. conducted extensive experiments cardiac structure segmentation task. The experimental show that proposed method significantly outperforms state-of-the-art comparison algorithms. Our code available at https://github.com/fengweie/SE_ASA.