Medical image segmentation is essential for disease diagnosis and the formulation of medical therapy. However, obtaining substantial labeled data is difficult due to significant costs and reliance on experts. Semi-supervised Medical Image Segmentation (SSMIS) addresses this challenge by leveraging a combination of plentiful unlabeled data and scarce labeled data. However, current SSL techniques, such as pseudo-labeling and consistency regularization, may encounter obstacles such as model instability and confirmation bias. We have developed an Interactive Calibration Learning (ICL) architecture that employs discrepancy-guided consistency regularization to mitigate bias by utilizing two pairs of teacher-student networks. ICL facilitates cross-model calibration by utilizing complementary predictions from two sets of teacher-student networks and enhancing pseudo-labels through uncertainty-aware constraints. In addition, we provide a plug-and-play feature extraction component, the Astrous Pyramid Spatial-Channel Attention (APSA) module, which captures multi-level global and local context, improving sensitivity to intra-class differences. In the ISIC dataset, ICL achieves dice similarity coefficients (DSC) of 89.72% under 5% labeled data, surpassing our baseline model AIM++ and UCMT by 5.33% and 1.50%, respectively.