Unsupervised Domain Adaptation Network With Category-Centric Prototype Aligner for Biomedical Image Segmentation

With the widespread success of deep learning in biomedical image segmentation, domain shift becomes a critical and challenging problem, as gap between two domains can severely affect model performance when deployed to unseen data with heterogeneous features. To alleviate this we present novel unsupervised adaptation network, for generalizing models learned from labeled source unlabeled target cross-modality segmentation. Specifically, our approach consists key modules, conditional discriminator (CDD) category-centric prototype aligner (CCPA). The CDD, extended adversarial networks classifier tasks, is effective robust handling complex images. CCPA, improved graph-induced alignment mechanism cross-domain object detection, exploit precise instance-level features through an elaborate representation. In addition, it address negative effect class imbalance via entropy-based loss. Extensive experiments on public benchmark cardiac substructure segmentation task demonstrate that method significantly improves domain.