Unsupervised domain adaptive medical segmentation network based on contrastive learning

Accurate organ segmentation from magnetic resonance imaging (MRI) or computed tomography (CT) images is essential for surgical planning and decision-making. Traditional fully supervised deep learning methods often exhibit a significant decline in performance when applied to datasets that differ from the training data, thus limiting their clinical applicability. This study proposes a novel segmentation method based on unsupervised domain adaptation, aiming to improve cross-domain segmentation performance without the need for ground truth labels in the target domain. Specifically, our method trains the network with labeled source images and unlabeled target images, introducing a bidirectional feature-prototype contrastive loss to align features across domains, minimizing within-class variations and maximizing between-class variations. To further improve model performance, we propose a prototype-guided pseudo-label fusion module that generates high-quality pseudo-labels for the unlabeled target images between domain prototypes. Experimental results show that our method outperforms other unsupervised domain adaptation segmentation approaches, achieving state-of-the-art performance. Code is available at: https://github.com/WANGSIQII/UDA.git.