ABSTRACT Robust cross‐domain segmentation is the key to achieving unlabeled large‐scale medical image analysis. However, the domain shift caused by modality differences weakens the model's generalization ability in the target domain significantly. In addition, the high cost of pixel‐level annotation greatly limits the applicability of existing methods in clinical practice. Although unsupervised domain adaptation (UDA) has been extensively studied, most existing methods focus on 2D style alignment or feature distribution matching, which makes it difficult to effectively model 3D anatomical structures. Meanwhile, although the teacher–student framework shows potential in addressing the lack of annotations in the target domain, the distribution discrepancy between the source and target domains renders its pseudo labels susceptible to noise and unreliable, which can lead to unstable training and error propagation. To address the above challenges, this paper proposes a unified 3D unsupervised segmentation framework, named 3D‐UDASeg, which achieves enhanced cross‐domain generalization, reliable 3D structural modeling, and improved pseudo‐label quality through collaborative optimization at the image, feature, and supervision levels. Specifically, the framework first employs an orthogonal multi‐head self‐attention (OMSA)‐driven style transfer module to generate cross‐domain transitional modalities, thereby mitigating domain shift while preserving the consistency of 3D anatomical structures. Secondly, during the encoder pretraining stage, a global–local boundary‐aware fusion (GLBA) module is introduced to enhance anatomical boundary modeling. Finally, a pseudo‐label optimization mechanism based on prediction entropy and dual‐branch consistency (PEDC) is designed to reduce the interference of low‐quality pseudo labels and improve the reliability of the supervisory signals. Systematic experiments were conducted on two publicly available cross‐domain cardiac segmentation datasets, M&Ms and MMWHS. The results demonstrate that the proposed method outperforms advanced approaches such as MAPSeg in terms of Dice and ASSD metrics, validating its effectiveness in mitigating domain shift, enhancing 3D boundary modeling, and improving pseudo‐label reliability, as well as its potential clinical applicability.
Zhang et al. (Sun,) studied this question.