Semi-supervised learning offers a viable solution to the scarcity of labeled data, particularly in the field of medical image segmentation. However, the performance enhancement of semi-supervised learning is often constrained by the presence of noise in pseudo-labels. Existing semi-supervised methods for medical image segmentation have not deeply explored the mechanism behind pseudo-label noise generation. This paper proposes a novel semi-supervised learning framework that effectively eliminates noise in pseudo-labels through label purification and reliable pixel learning, significantly enhancing the performance of gland histopathology image segmentation. Inspired by error-correction research, our semi-supervised framework employs two synergistic models to learn and correct noise in pseudo-labels. Initially, our cyclic iterative noisy pseudo-label generation strategy enables the segmentation model to produce noisy pseudo-labels that mimic the noise distribution of unlabeled data. Subsequently, the purification model learns the mapping relationship from these noisy pseudo-labels to ground truth, achieving denoising of pseudo-labels. Additionally, to mine reliable supervisory information within refined pseudo-labels, we introduce a weighted loss mechanism based on reliable pixel learning, enhancing the model’s learning quality. Experiments on two public gland segmentation datasets, GlaS and CRAG, demonstrate that our method achieves competitive performance compared to other advanced semi-supervised segmentation approaches.