A domain adaptation segmentation network based on pyramid prototypes and contrastive learning

Accurate organ segmentation from MRI or CT images is essential for surgical planning and decision-making. Traditional fully supervised deep learning methods often experience performance degradation when applied to datasets that differ from the training data, limiting their clinical applicability. Unsupervised domain adaptation is an effective solution to this challenge, which alleviates domain distribution shifts and significantly enhances performance. This study proposes a novel segmentation method based on unsupervised domain adaptation to improve cross-domain segmentation without needing ground truth labels in the target domain. Specifically, our approach trains the network using labeled source images and unlabeled target images. First, a ResNet-50 encoder is used to extract pyramid features from both the source and the target domains. Based on these extracted features, pyramid prototypes for both domains are generated. Furthermore, we introduce contrastive loss between multiple prototypes to align features across domains, minimizing within-class variations and maximizing between-class variations. Experimental results demonstrate that our method outperforms existing unsupervised domain adaptation segmentation techniques, achieving state-of-the-art performance.