Exploring refined boundaries and accurate pseudo-labels for semi-supervised medical image segmentation

Xiaochen Ma, Yanfeng Li, Jia Sun, Houjin Chen, Yihan Ren, Ziwei Chen

Abstract: Due to the tedious and expensive nature of annotating data for medical image segmentation tasks, semi-supervised learning (SSL) methods utilizing a small amount of labeled data have gained widespread attention. However, most existing methods overlook the importance of boundary regions in multi-class tasks and are sensitive to incorrect pseudo-label. In this paper, a novel Dual-Cycled Boundary Refine Network (DCBR-Net) is presented, consisting of two simple yet slightly different segmentation networks and a Boundary Residual Refine (BRR) module. First, a new boundary refine method for the semi-supervised medical image segmentation field is designed, that is a BRR module with a residual architecture. This module is capable of enhancing the representation of boundaries while ensuring the quality of inner regions. Besides, a dual-cycled pseudo-label scheme is designed to train the unlabeled data. Through providing diverse outputs and achieving double supervision for the result, the issue of ineffective guidance caused by model consistency after multiple iterations is alleviated. Furthermore, a novel dynamic loss function is developed based on the prediction disagreement between different models, which can suppress the influence of incorrect pseudo-labels. Extensive experiments conducted on four public medical datasets demonstrate that our network can achieve competitive results, especially with higher reliability in boundary regions. On the ACDC, LA, and Fundus datasets, with only a 20% labeled ratio, our network achieves DSC scores of 90.53%, 90.40%, and 89.19%, respectively, which are comparable to fully supervised performance.

External IDs:dblp:journals/apin/MaLSCRC25