Abstract: Although 3D medical image segmentation has achieved remarkable progress, its heavy reliance on large-scale annotated data remains a critical limitation. Semi-supervised learning can alleviate this issue. However, existing semi-supervised methods usually struggle to dynamically adjust the supervision strength of pseudo-labels during training and typically enforce prediction consistency at a single scale only. To address these limitations, a dynamic uncertainty-aware loss function is first proposed to evaluate model predictions and generate uncertainty-aware weights, thereby adaptively regulating the supervision strength of pseudo-labels. In addition, a novel dual perturbation module at both pixel and region levels is introduced to impose perturbations from both fine-grained appearance variations and coarse-grained structural changes, which work together to enhance consistency constraints from multiple perspectives and improve the model’s adaptability to local intensity variations and structural discrepancies. To validate the effectiveness and generalizability of the proposed method, extensive experiments are conducted on three public 3D medical image segmentation datasets, including LA (left atrium), PA (pancreas-CT), and BR (brain tumor segmentation). Both quantitative and qualitative experimental results demonstrate that the proposed method achieves superior segmentation performance across multiple evaluation metrics and consistently outperforms several state-of-the-art semi-supervised segmentation methods, confirming its effectiveness and practical applicability. Code is available at https://github.com/SSMIS-jiutian/DUAMC.