Semi-supervised medical image segmentation via pseudo-labeling refinement and dual-adaptive adjustment schemes

Bin Zheng, Wenyi Zhao, Wentao Liu, Ziyang He, Chuanbo Qin, Huihua Yang

Abstract: Segmentation of tissues, organs and lesions from 2D/3D medical images is crucial in clinical analysis, diagnosis and treatment. Fully-supervised segmentation methods are costly, since labeling extensive medical images is both laborious and time-consuming. Semi-supervised segmentation, which learns from both the labeled and unlabeled data, is a promising method to alleviate the requirement of annotations. However, most existing semi-supervised methods face challenges in filtering out pseudo-label noise, extracting useful information from hard-to-segment samples, and balancing the weights of supervised and unsupervised losses. In this work, we propose a novel pseudo-labeling-based semi-supervised framework for medical image segmentation tasks, comprising two training phases. The first process utilizes a small iteration to encourage the learning of reliable points in predictions, achieving the pseudo-labeling refinement. The second phase employs pseudo-labels to compute a novel loss based on the harmonic mean, which enhances focus on hard samples. Additionally, we introduce a new min-max strategy for adaptively adjusting the weights between supervised and unsupervised losses, improving the robustness of the segmentation model. Experimental results on four public multimodal medical datasets demonstrate that our proposed method outperforms other state-of-the-art semi-supervised segmentation methods.

External IDs:dblp:journals/pr/ZhengZLHQY26