Semi-supervised learning offers a beneficial approach for segmenting medical images when limited annotations are available. Recent prominent techniques focus on using essential elements from a dual-view perspective to create pseudo-labels for unlabeled data. However, they still suffer from the potential loss of important data and the risk of creating inaccurate labels. To overcome these issues, we present an innovative training framework with dynamic pseudo-label refinement that thoroughly explores the potential of the dual-view method. Our proposed training framework includes two complementary modules: Hierarchical Pseudo Label Generation (HPLG) and Dynamic Pseudo Label Correction (DPLC). The HPLG module employs a dynamic approach to generate hierarchical pixel-level pseudo-labels, which are stratified based on their reliability, leveraging both the agreement and discrepancies within the …
本文提出一种基于动态伪标签优化的半监督医学图像分割框架。针对双视角方法易丢失重要数据且伪标签不准确的问题，设计分层伪标签生成（HPLG）与动态伪标签校正（DPLC）两个互补模块，按可靠性生成分层像素级伪标签，并利用双视角的一致性与差异进行动态修正，从而提升分割性能与标签质量。