Prototype-aware channel-selective spatial interaction for semi-supervised medical image segmentation

Semi-supervised 2D medical image segmentation reduces the dependence on dense pixel-level annotations by exploiting unlabeled images, but reliable learning remains challenging when only limited labeled data are available. Existing methods mainly rely on prediction-level consistency or generic feature interaction, and often lack explicit semantic guidance for selecting foreground-background discriminative channels and refining selected features at spatial locations. To address this issue, we propose Prototype-aware Channel-Selective Spatial Interaction (PCSI), a teacher-student framework for semi-supervised 2D medical image segmentation. PCSI constructs decoupled labeled and unlabeled foreground-background prototype memories, where reliable labeled prototypes provide semantic anchors and mature unlabeled prototypes gradually incorporate complementary distributional information. Based on prototype responses, PCSI selects a compact set of task-relevant channels according to foreground-background discriminability. Within this selected channel subspace, spatial feature tokens interact with category prototypes to obtain semantic residuals, which are injected into bottleneck features through confidence-gated residual enhancement. Extensive experiments on four public 2D medical image segmentation datasets under different labeled ratios demonstrate that PCSI consistently outperforms representative semi-supervised segmentation methods. These results indicate that prototype-aware, category-discriminative, and spatially adaptive feature interaction is effective for reliable semi-supervised medical image segmentation under limited annotation.