Synergistic Label-Stability Learning for Semi-supervised Left Atrium Segmentation

Recent semi-supervised learning approaches appealingly advance medical image segmentation for their effectiveness in alleviating the need for a large amount of expert-demanding annotations. However, most of them have two limitations: (i) neglect of the intra-class variation caused by different patients and scanning protocols, which makes the pixel-level label propagation difficult; (ii) non-selective stability learning (a.k.a., consistency regularization), resulting in distraction by the redundant easy regions. To address these, in this work, we propose a novel synergistic label-stability learning (SLSL) framework for semi-supervised medical image segmentation. Specifically, our method is built upon the teacher-student framework. Then, the label learning process includes the typical pseudo label learning that reinforces confirmation of well-classified easy regions and the cyclic real label learning that takes advantage of real labels and class prototypes to regularize the distribution of intra-class features from unlabeled data to facilitate label propagation. In addition, the difficulty-selective stability learning aims to regularize the perturbed stability only at the high-entropy (can be regarded as difficult) regions, rather than being distracted by the less-informative easy regions. Extensive experiments on left atrium segmentation from MRI show that our method can effectively exploit the unlabeled data and outperform other semi-supervised medical image segmentation methods.