Semi-supervised medical image segmentation based on multi-stage iterative training and high-confidence pseudo-labeling

Due to the scarcity and high cost of pixel-level annotations for training data, semi-supervised learning has gradually become a key solution. Most existing methods rely on consistency regularization and pseudo-label generation, often adopting multi-branch structures to generate pseudo-labels for co-training. Such approaches, however, commonly yield low-confidence pseudo-labels from perturbed inputs, which can degrade model performance. To address these challenges, we propose a novel semi-supervised segmentation framework that leverages a multi-stage training strategy, distinguishing between the training processes for labeled and unlabeled data to enhance pseudo-label reliability. This framework effectively minimizes the negative impact of multi-branch gradient interference during co-training, reducing the adverse effects of input perturbations. Furthermore, we introduce a Balanced Uncertainty Adjustment Module (BUAM) to improve pseudo-label generation, thus maximizing data utilization efficiency. By enhancing model stability and producing more reliable pseudo-labels, the proposed multi-stage approach offers a clear advantage over existing methods. Extensive experiments on the ISIC and Cardiac MRI medical image datasets demonstrate the advantages and effectiveness of our framework, which outperforms the state-of-the-art methods.