DSMT-Net: Dual-student mean teacher network with pixel-level pseudo-label optimization for semi-supervised medical image segmentation

Jun Su; Wenlong Sun; Bogdan Adamyk

doi:10.1016/j.compbiolchem.2025.108579

DSMT-Net: Dual-student mean teacher network with pixel-level pseudo-label optimization for semi-supervised medical image segmentation

Comput Biol Chem. 2025 Jul 9:119:108579. doi: 10.1016/j.compbiolchem.2025.108579. Online ahead of print.

Authors

Jun Su¹, Wenlong Sun², Bogdan Adamyk³

Affiliations

¹ School of Computer Science, Hubei University of Technology, Wuhan, 430068, China. Electronic address: sujuncs@hbut.edu.cn.
² School of Computer Science, Hubei University of Technology, Wuhan, 430068, China. Electronic address: sunwenlong@hbut.edu.cn.
³ Aston Business School, Aston University, Birmingham, B4 7ET, UK. Electronic address: b.adamyk@aston.ac.uk.

PMID: 40651267
DOI: 10.1016/j.compbiolchem.2025.108579

Abstract

Medical imaging technologies, as essential tools for the precise visualization of internal anatomical structures, play a crucial role in early disease detection and ensuring accurate diagnosis. Recently, semi-supervised learning has become a key strategy in medical image segmentation to reduce reliance on scarce annotated data. However, existing frameworks, such as the Mean Teacher (MT), often suffer from low-quality pseudo-labels and limited robustness due to structural homogeneity and noise amplification in complex medical scenarios. To address these issues, this study presents a novel Dual-Student Mean Teacher Network (DSMT-Net), which enhances performance through a collaborative complementary architecture and pixel-level pseudo-label optimization. First, DSMT-Net combines U-Net and Mamba-UNet as dual students, utilizing the former's local boundary accuracy and the latter's global dependency modeling via a state-space model. Second, a pixel-level pseudo-label enhancement mechanism is introduced, combining pixel-level similarity analysis, adaptive confidence threshold setting, and iterative propagation to improve pseudo-label quality while maintaining structural consistency. Third, a self-supervised contrastive loss is adopted to enforce feature consistency between the dual students, alleviating noise propagation and improving the efficiency of unsupervised learning. Comprehensive evaluations on the ACDC and LA datasets confirm the effectiveness of DSMT-Net, highlighting its substantial capability to lower annotation requirements in medical image segmentation tasks. This provides a robust and scalable framework for semi-supervised learning in medical image segmentation, advancing clinical diagnostic efficiency and accuracy. Our code is available at https://github.com/sunwenlong1/DSMT.git.

Keywords: Dual student network; Mamba architecture; Mean teacher; Medical image segmentation; Pseudo labels optimization; Semi-supervised learning.