Uncertainty-Guided Dynamic Pseudo-Label Enhancement for Semi-Supervised Left Atrial Segmentation

Atrial fibrillation (AF) is a common atrial-origin arrhythmia. Left atrial image segmentation is critical for timely diagnosis and intervention. The traditional manual annotation is inefficient and inconsistent. Semi-supervised learning (SSL) has thus become a practical way to alleviate the annotation bottleneck in medical imaging. Existing SSL segmentation methods still suffer from fixed confidence thresholds, accumulated pseudo-label noise, and insufficient local perturbations. To address these challenges, this paper proposes a semi-supervised medical image segmentation method based on uncertainty-guided dynamic pseudo-label enhancement. This method dynamically adjusts the confidence threshold to better adapt to the dynamic characteristics of data distribution and feature changes. And it introduces an uncertainty-guided region mixing (UMIX) module, which uses prediction entropy to identify high uncertainty regions and achieves data enhancement through cross-sample feature replacement. UniMatch serves as the backbone, with strong and weak perturbations applied at both image and feature levels to enable more effective supervised learning and improve segmentation. The proposed method was compared against competing methods on two public datasets. Through comparative experiments with other methods on two public datasets, together with ablation studies, we validated the superior performance of the proposed method in image segmentation tasks and the effectiveness of its individual components.