Existing semi-supervised learning methods use uncertainty estimation to perform well in 3D medical image segmentation. However, these methods usually use fixed thresholds to filter high-confidence regions to reduce the impact of noise in pseudo-labels, but do not fully consider the complexity of the distribution of different labeled data and unlabeled data, resulting in the loss of effective information. In addition, most methods usually require multiple forward passes to obtain the uncertainty of the prediction, which greatly increases the computational cost. To this end, this paper proposes a semi-supervised medical image segmentation framework based on uncertainty-driven dynamic correction and multi-scale consistency learning. First, a dynamic correction strategy is designed to effectively reduce the impact of pseudo-label noise on the model through uncertainty-driven dynamic weight adjustment and dynamic scale screening combined with consistency regularization, thereby significantly improving the segmentation performance and robustness. Secondly, an uncertainty estimation method based on multi-scale prediction is designed, which can generate reliable uncertainty estimates with only a single forward pass, reducing the computational cost, and introduces uncertainty-based multi-scale consistency learning to gradually learn meaningful feature regions. Experimental results on two public medical image datasets (LA and BraTS) show that our method achieves superior performance compared to several state-of-the-art methods.