SCOPE-Med: Structure-Preserving and Contour-Oriented Pseudo-Label Reconstruction for mixed-domain semi-supervised medical image segmentation.

Mixed-domain semi-supervised medical image segmentation aims to learn segmentation models from a small labeled set in one domain and abundant unlabeled data from multiple unknown domains. Its central challenge is to construct reliable pseudo-supervision under substantial domain shift. Existing methods commonly infer pseudo-label reliability from confidence or predictive uncertainty at the sample, model, or complete-mask level, implicitly assuming uniform reliability across each prediction. Coarse anatomical support can remain relatively stable, whereas local contours and fine details are more sensitive to appearance variation and boundary ambiguity. This uneven sensitivity makes a single trust decision inadequate, because it may either retain corrupted local details or discard useful structural information. To address this issue, this paper proposes SCOPE-Med, a Structure-preserving and Contour-Oriented Pseudo-label rEconstruction framework for mixed-domain semi-supervised medical image segmentation. SCOPE-Med organizes heterogeneous teacher predictions into coarse structural components and detail-sensitive residuals through Gaussian decomposition. The structural components retain spatially coherent region-level information, while the residuals preserve local variation that may contain both useful corrections and domain-induced noise. Predictive uncertainty then regulates local residual correction and suppresses unstable detail injection in ambiguous regions. An auxiliary boundary prediction head and contour-constrained objective provide complementary geometric supervision during student optimization. Through decomposition, regulation, and reconstruction, SCOPE-Med refines pseudo-supervision at the component level instead of assigning a single confidence score to an entire prediction. Experiments are conducted on two public two-dimensional binary segmentation benchmarks. Across overlap- and boundary-sensitive metrics, SCOPE-Med achieves strong average performance under the evaluated mixed-domain settings.