The study of classification methods for class imbalanced medical images not only helps to improve the performance of medical image processing methods and the accuracy of medical diagnosis, but also has important significance for promoting medical research and improving medical standards. Aiming at the problem of how to correct the bias between the majority class and the minority class in the method, this paper proposes a class imbalanced medical image semi-supervised classification framework BCMatch based on bias correction. First, to address the problem of low utilization of unlabeled data for minority classes under semi-supervised learning, This paper proposes a dynamic pseudo-label threshold adjustment module called LTCPL based on the idea of curriculum learning. It enables adaptive adjustment of pseudo-label thresholds for each category, facilitating better use of unlabeled data and increasing attention to minority classes, thereby correcting the encoder’s bias caused by class imbalance. In addition, to address the classifier’s bias of predicting all data as majority classes, this paper proposes a data selection module CSS based on feature cosine similarity, which freezes the encoder’s parameters and allows the classifier to perform two-stage training on the balanced data alone, thereby correcting the method classifier’s bias towards majority and minority classes on the decision boundary. Experimental results on the ISIC2019 Dermatology Dataset and ChestXRay Chest X-ray Dataset show that BCMatch is superior to the most advanced semi-supervised classification methods, and can further improve performance when combined with two-stage training.