Semi-supervised learning (SSL) is an effective strategy for extracting useful information from unlabeled datasets to improve deep model performance. SSL is widely used in medical image segmentation to alleviate the burden of expensive pixel-level labeling. Most existing SSL-based medical image segmentation methods use all unlabeled data equally to update the model via unsupervised loss. However, not all unlabeled data are applicable to the same model training process due to differences in data quality and importance. In this study, we propose a dual-decoder data decoupling training-based semi-supervised medical image segmentation network (DD-Net) that enables the model to focus on challenging regions. DD-Net decouples the prediction of unlabeled data into data with different functions based on the degree of confidence matching from two student decoders and adopts different optimization strategies for different functional data. Specifically, for high- reliability part, we employ cross pseudo supervision learning to improve the reliability of model prediction. For divergent predictions, we propose mutual matching learning to guide the model to learn richer information in high uncertainty data by assigning high entropy to the prediction data. For low-confidence prediction, we employ reinforcement consistency learning to enhance the context representations of the model to effectively extract important details, such as edges and contours, from unlabeled data. Experiments on four medical challenging image datasets demonstrate that our method outperforms existing state-of-the-art methods. The code is available at https://github.com/TaifengHuang/DD-Net.