CPLMT-Net: Consistency and Prototype Learning Mean Teacher Network for Semi-Supervised Medical Image Segmentation

Fine-grained measurement of the organs with diverse shapes and arbitrary sizes in medical images has significant clinical value, which helps to accurately identify the key lesions. Semi-supervised medical image segmentation (SSMIS), leveraging a small amount of labeled data along with a large amount of unlabeled data, can fulfill this goal well. SSMIS depends on pseudo-labeling and consistency regularization. However, some challenges, including the uncertainty of pseudo-labels and insufficient utilization of unlabeled data, still remain. To address these issues, we propose a novel consistency and prototype learning mean teacher network (CPLMT-Net), which is characterized by three significant features: reliable pseudo-label supervision, robust consistency against uncertain regions, and effective global semantic fusion. We design the mixed foreground–background (MFB) module to refine the supervision signal for the unlabeled data by mixing the background of labeled data with the foreground of unlabeled data, mitigating the negative impact of erroneous pseudo-labels. We enforce consistent learning (CON) across multiple decoders to reduce the influence of uncertain regions. We implement prototype fusion (PF) by computing the prototypes for both labeled and unlabeled images, and integrating these prototypes into a unified prototype representation to capture global information. Finally, a prototype consistency with hybrid supervision loss (PC-HSL) integrates the above modules, enabling them to work collaboratively. Extensive experiments on three publicly available 3-D medical image datasets demonstrate the superiority of CPLMT-Net. More importantly, we enhance the model’s interpretability for clinical observation through ablation visualization, offering accurate and robust measurement of the diverse organs.