<p>In the field of medical image segmentation, annotating data collection is crucial for model training, but it is also a formidable task. The mainstream algorithm to address this issue is semi-supervised learning based segmentation, with one of the key technologies being pseudo label generation. However, incorrect pseudo labels will lead to error accumulation effect during training, thereby decreasing the model’s segmentation performance. To alleviate this problem, we propose a simple but effective pseudo label generation mechanism called joint pseudo supervision (JPS). Our basic framework consists of two pairs of teacher-student architectures with the same structure but independent parameters. The parameters of the student model are obtained through training, and the parameters of the teacher model are updated by the parameters of the corresponding student model through exponential moving average. Inspired by the idea of ensemble learning, JPS uses two independent teacher models to jointly generate pseudo labels with high-confidence, and then supervises the two student models with it, thereby restricting the accumulation of errors. By imposing consistency constraints, JPS encourages the two student models to learn from each other, further reducing the negative impact of incorrect pseudo labels. In the inference phase, the two student models collaborate to generate segmentation results and improve the overall segmentation performance. Experimental results show that the our method performs competitively with several state-of-the-art methods when benchmarked on public datasets, i.e., Automated Cardiac Diagnosis Challenge, Multi-Modality Whole Heart Segmentation and CVC-ClinicDB.</p>