Adaptive Entropy Regularization for Unsupervised Domain Adaptation In Medical Image Segmentation

The unsupervised domain adaptation approach based on adversarial training has achieved promising performance in cross-modality medical image analysis tasks. However, deep learning models often produce overconfident but incorrect predictions, which is exacerbated in the presence of domain shifts. In this paper, we propose an adaptive entropy regularization framework for unsupervised domain adaptation in cross-modality medical image segmentation. Our framework consists of two key designs: pixel reliability assessment and entropy-based confidence regularization. We first assess pixel reliability based on the model's predictive consistency over a set of label-preserving randomly augmented image sets. We then propose an entropy-based confidence regularization strategy, which increases the confidence level by minimizing the information entropy of reliable pixels while maximizing the information entropy of unreliable pixels to diversify their predictions and alleviate the problem of overconfident but incorrect predictions. Extensive experiments on cross-modality cardiac structure segmentation tasks show that our approach outperforms other state-of-the-art UDA methods by a large margin. Our code will be released soon.