In medical image segmentation, semi-supervised learning addresses the challenge of high annotation costs. However, existing methods often overlook the inconsistent quality of pseudo-labels and the limited generalization of model constraints. To address these issues, this study proposes a conflict-driven cross-learning (CDCL) framework that seamlessly integrates cross supervision with a mean teacher model. The proposed framework leverages two different teacher-student structures to facilitate effective knowledge transfer and enhance segmentation performance. Knowledge is exchanged between student models via their respective teacher models, enabling mutual learning and capability improvement. In addition, a feature conflict loss (FCL) is introduced to promote the transfer of diverse and complementary features between models, thus enriching the learning dynamics. To further augment training data, a pairwise copy-paste (PCP) strategy is adopted to generate novel training samples. Experimental results on two public datasets demonstrate the effectiveness of the proposed CDCL framework. With only 10% labeled data on the ACDC dataset, average Dice and Jaccard coefficients reach 90.23% and 82.71%, outperforming the BCP model by 1.39% and 2.09%, respectively. On the PROMISE12 dataset, Dice coefficients of 78.9% and 80.09% are achieved with 20% and 30% labeled data, respectively.