Language-Prompted Dynamic Learning for Semi-Supervised Medical Image Segmentation

Medical image segmentation is fundamental to clinical diagnosis and treatment planning, yet existing models are constrained by the scarcity of annotated data, which are costly and labor-intensive to obtain. Semi-supervised learning (SSL) mitigates this issue by leveraging large volumes of unlabeled data, but most SSL methods rely solely on visual cues and often fail to capture subtle structures or low-contrast regions common in medical imaging. To address this limitation, we present$\bf{LanDy}$, a Language-Prompted Dynamic Learning framework for semi-supervised medical image segmentation. LanDy introduces textual semantics from medical descriptions to enrich visual representations and reduce the ambiguity of pseudo-labels. Concretely, textual embeddings dynamically modulate convolutional filters to provide context-aware feature extraction, while a text-guided refinement mechanism improves the reliability of pseudo-labels on unlabeled data. Extensive experiments on benchmark datasets demonstrate that LanDy consistently outperforms state-of-the-art SSL methods, delivering more accurate and robust segmentation under annotation-efficient settings.