摘要：

在半监督医学图像分割领域, Mean Teacher是一个备受关注并且被广泛使用的框架之一。现有基于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, 即使教师网络性能比学生网络差, 学生网络依然受教师网络监督, 从而导致了错误累积。此外, 这些方法都使用固定的伪标签阈值从教师网络的预测值中寻找正确信息, 这样虽然过滤了大部分错误信息, 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, 提出一种基于选择监督和动态阈值的半监督医学图像分割模型SSDT。该模型使得学生网络可以选择何时接受教师网络的监督, 避免了教师网络在性能不足时依然监督学生网络, 并且通过新设计的动态阈值模块, 使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%标注数据的左心房分割(LA)数据集和自动心脏诊断挑战(ACDC)数据集上, SSDT的Dice系数分别达到了90.94%和89.93%。在4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。

关键词: 医学图像分割, 半监督学习, 动态阈值, 教师网络, 学生网络

Abstract:

Mean Teacher is a highly regarded and widely used framework for semi-supervised medical image segmentation. However, methods based on the Mean Teacher do not selectively accept the supervision of the student network over the teacher network during training. This implies 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. This results in an accumulation of errors. Moreover, all these methods use a fixed threshold for pseudo-labels to obtain correct information from the predictions of the teacher network. Although this filters out most incorrect information, it also eliminates much of the correct information, which greatly limits the availability of pseudo-labels. To address these issues, a semi-supervised medical image segmentation model based on selective supervision and dynamic threshold, named SSDT, is proposed. 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 using the newly designed dynamic threshold module, thereby maximizing the retention of the correct information in the teacher network output. On the LA and ACDC datasets with 20% labeled data, SSDT achieves Dice coefficients of 90.94% and 89.93%, respectively. Extensive experiments on four medical image datasets demonstrate that SSDT has superior segmentation performance compared with several state-of-the-art methods.

Key words: medical image segmentation, semi-supervised learning, dynamic threshold, teacher network, student network