Semi-supervised medical image segmentation based on cross-pseudo-supervised guidance with high-confidence pseudo labeling

Medical image segmentation tasks based on deep convolutional neural networks typically require a large amount of labeled data. However, medical image annotation is highly dependent on medical expertise, making the acquisition of a sufficient number of labeled samples extremely time-consuming and labor-intensive. Therefore, reducing the reliance on pixel-level annotations when training segmentation models with limited labeled samples is a major challenge in medical image segmentation tasks. Semi-Supervised Learning (SSL) methods, particularly those generating pseudo-labels for unlabeled data and consistency regularization, are crucial for addressing this issue. Although the combination of these methods has enhanced model performance in semi-supervised segmentation tasks, the introduction of various perturbations inevitably introduces extra noise, which affects the confidence of the pseudo-labels. To alleviate this issue, we propose a novel two-stage semi-supervised medical image segmentation method, which uses uncertainty estimation maps to guide cross-pseudo-supervision training, enhancing model performance by improving the confidence of pseudo-labels. We validated our method on 3D MRI images of the left atrium and 2D images of dermoscopic examination. Experimental results demonstrate that our approach achieves significant performance gains using a small amount of labeled data and a large amount of unlabeled data, highlighting the potential of our framework for addressing the challenging problems of semi-supervised learning by segmentation models.