Dynamic pseudo-label learning for semi-supervised ultrasound cardiac segmentation

Semi-supervised ultrasound cardiac segmentation plays a crucial role in the diagnosis and treatment of cardiovascular diseases, reducing the reliance on labeled data. Existing semi-supervised methods mainly rely on consistency regularization and pseudo-labeling strategies, which are susceptible to introducing noise when processing unlabeled ultrasound images, thereby compromising segmentation performance. To tackle this challenge, we propose a method called Dynamic Pseudo-Label Learning (DPL) for Semi-Supervised Ultrasound Cardiac Segmentation. First, to preserve the integrity of cardiac structures and enhance background feature learning which can reduce the noise effects, we introduce an adaptive Copy-Paste data augmentation approach. This method primarily performs image mixing in the background regions, calculating the Copy-Paste proportion based on the background size. Second, to prioritize high-confidence ultrasound samples and progressively introduce more challenging samples during training, we design an entropy-based weight optimization strategy that evaluates the prediction entropy of each unlabeled ultrasound image and dynamically select the proper samples for training. Experimental results demonstrate that the proposed method achieves significant performance improvements across multiple ultrasound cardiac segmentation datasets, confirming its effectiveness and robustness in semi-supervised ultrasound cardiac segmentation tasks.