Unsupervised Domain Adaptation Using Soft-Labeled Contrastive Learning with Reversed Monte Carlo Method for Cardiac Image Segmentation

Recent unsupervised domain adaptation methods in medical image segmentation adopt centroid/prototypical contrastive learning (CL) to match the source and target features for their excellent ability of representation learning and semantic feature alignment. Of these CL methods, most works extract features with a binary mask generated by similarity measure or thresholding the prediction. However, this hard-threshold (HT) strategy may induce sparse features and incorrect label assignments. Conversely, while the soft-labeling technique has proven effective in addressing the limitations of the HT strategy by assigning importance factors to pixel features, it remains unexplored in CL algorithms. Thus, in this work, we present a novel CL approach leveraging soft pseudo labels for category-wise target centroid generation, complemented by a reversed Monte Carlo method to achieve a more compact target feature space. Additionally, we propose a centroid norm regularizer as an extra magnitude constraint to bolster the model's robustness. Extensive experiments and ablation studies on two cardiac data sets underscore the effectiveness of each component and reveal a significant enhancement in segmentation results in Dice Similarity Score and Hausdorff Distance 95 compared with a wide range of state-of-the-art methods.