摘要： 在半监督医学图像分割领域中，Mean Teacher是一个备受关注并且被广泛使用的框架之一。但是基于Mean Teacher的方法在训练时学生网络对教师网络的监督不做选择地接受，这就导致即使教师网络性能差于学生网络，学生网络依然受教师网络监督，从而导致了错误累计。而且这些方法都使用固定的伪标签阈值这种方式，从教师网络的预测值中寻找正确信息，这样虽然过滤掉了大部分错误信息，但也筛去了许多正确信息，这极大限制了伪标签的可用性。针对以上问题，我们提出一种选择监督和动态阈值的半监督医学图像分割模型（selection-supervised and dynamic threshold semi-supervised medical image segmentation model，SSDT）。该模型使得学生网络可以选择何时接受教师网络地监督，避免了教师网络在性能不足时依然监督学生网络。还通过新设计的动态阈值模块，网络可以选择适合当前训练阶段的伪标签阈值，更大限度地保留教师网络输出中的正确信息。在使用20%标注数据的LA，ACDC数据集上SSDT的Dice系数分别达到了90.94%，89.93%。在四个医学图像数据集上的大量实验结果表明，与几种最先进的方法相比，SSDT具有优越的分割性能。

Abstract: Mean Teacher is one of the highly regarded and widely used frameworks in semi-supervised medical image segmentation. However, the method based on Mean Teacher does not selectively accept the supervision of the student network over the teacher network during training. This leads to the fact that even if the performance of the teacher network is inferior to that of the student network, the student network is still supervised by the teacher network, resulting in the accumulation of errors. Moreover, these methods all use a fixed threshold for pseudo-labels to find the correct information from the teacher network's predictions. Although this filters out most of the incorrect information, it also eliminates many of the correct information, which greatly limits the availability of pseudo-labels. To respond to the above issues, we propose a selection-supervised and dynamic threshold semi-supervised medical image segmentation model (SSDT). This model allows the student network to choose when to accept supervision from the teacher network, preventing the teacher network from supervising the student network when its performance is insufficient. The network can select a pseudo-label threshold suitable for the current training stage through the newly designed dynamic threshold module, maximizing the retention of correct information in the teacher network output. On the LA and ACDC datasets with 20% labeled data, the SSDT achieves a Dice coefficient of 90.94% and 89.93%, respectively. Extensive experimental results on four medical image datasets demonstrate that SSDT has superior segmentation performance compared to several state-of-the-art methods.