A Semi-Supervised Semantic Segmentation Method Based on a Dual-Student Single-Teacher Network for Skin Lesions

近些年，在医学图像领域出现了很多半监督学习方法，这些方法大多数是基于学生–教师结构的一致性正则化(Consistency Regularization)的半监督学习范式，或者是基于伪标签(Pseudo-Label)的半监督学习范式，但将两种学习范式结合起来的研究较少。因此文章提出了一种双学生单教师网络的半监督语义分割方法，该方法同时结合了一致性正则化与伪标签的学习范式，通过引入额外的学生网络降低模型陷入局部最优解的几率，避免模型之间参数耦合的风险；与此同时，通过引入UMIX模块能够生成高质量的伪标签，提高模型对无标签数据中有效信息的获取，进而提升模型的分割表现。并在ISIC 2017与ISIC 2018数据集上做了大量的实验，结果表明本文提出的方法实现了更优秀的分割表现。In recent years, numerous semi-supervised learning methods have emerged in the field of medical image analysis. Most of these methods follow either the consistency regularization-based semi-supervised learning paradigm using the student-teacher framework or the pseudo-label-based semi-supervised learning paradigm. However, there is limited research on integrating both paradigms. Therefore, this paper proposes a semi-supervised semantic segmentation method based on a dual-student single-teacher network, which combines consistency regularization and pseudo-label learning paradigms. By introducing an additional student network, the proposed method reduces the likelihood of the model falling into local optima and mitigates the risk of parameter coupling between networks. Meanwhile, the incorporation of the UMIX module enables the generation of high-quality pseudo labels, enhancing the model’s ability to extract useful information from unlabeled data and ultimately improving segmentation performance. Extensive experiments on the ISIC 2017 and ISIC 2018 datasets demonstrate that the proposed method achieves superior segmentation performance.