In order to solve the problem of limited model generalization ability caused by imbalanced distribution of target semantic categories in some scarce medical image segmentation tasks, this paper proposes a semi-supervised learning medical image segmentation model CDCL-SSLNet, which achieves feature complementarity through cross-learning of two segmentation submodels with different attributes, namely, UNet and Swin-UNet. The introduction of class distribution fairness factor and class learning fairness factor reasonably weights the loss function, dynamically guides the model to learn the unbalanced data of semantic categories, effectively reduces the learning bias, and then improves the model generalization ability. In the experiment, 5% and 10% of the data in Synapse multi-organ segmentation dataset are selected to simulate labeled data to train the model. When only 5% and 10% of the label data participated in the training, the Dice coefficients of CDCL-SSLNet reached 65.71% and 77.49%, respectively, and the performance of HD95 was 28.97 and 22.07, respectively, and the performance of these two indicators was improved by 17%. The results show that CDCL-SSLNet is able to take into account the accurate segmentation of large and small targets, effectively solves the problem of insufficient model generalization ability caused by the imbalance of class distribution in scarce data, and effectively improves the efficiency and accuracy of medical image segmentation.