A New Bidirectional Unsupervised Domain Adaptation Segmentation Framework

Domain shift happens in cross-domain scenarios commonly because of the wide gaps between different domains: when applying a deep learning model well-trained one domain to another target domain, usually performs poorly. To tackle this problem, unsupervised adaptation (UDA) techniques are proposed bridge gap domains, for purpose improving performance without annotation domain. Particularly, UDA has great value multimodal medical image analysis, where difficulty is practical concern. However, most existing methods can only achieve satisfactory improvements direction (e.g., MRI CT), but often perform poorly other (CT MRI), limiting their usage. In paper, we propose bidirectional (BiUDA) framework based on disentangled representation equally competent two-way performances. This employs unified domain-aware pattern encoder which not adaptively encode images domains through controller, also improve efficiency by eliminating redundant parameters. Furthermore, avoid distortion contents and patterns input during process, content-pattern consistency loss introduced. Additionally, better segmentation performance, label strategy provide extra supervision recomposing target-domain-styled corresponding source-domain annotations. Comparison experiments ablation studies conducted two public datasets demonstrate superiority our BiUDA current state-of-the-art effectiveness its novel designs. By successfully addressing adaptations, offers flexible solution real-world scenario.