Semi-supervised Medical Image Segmentation with Semantic Distance Distribution Consistency Learning

Semi-supervised medical image segmentation has attracted much attention due to the alleviation of expensive annotations. Recently, many existing semi-supervised methods incorporate unlabeled data via the consistency learning. However, those consistency learning methods usually utilize the mean teacher structure, resulting in the different feature distribution on the intermediate representations. As for semi-supervised medical image segmentation, the different feature distribution limits the efficiency of consistency. In this paper, we propose Semantic Distance Distribution (SDD) Consistency Learning method, which has the ability to maintain the same feature distribution on the intermediate representations. On the one hand, to model invariance on feature distribution, we consider the shared encoder instead of averaging model weights. On the other hand, we introduce a SDD Map for consistency learning on the intermediate representations, where SDD Map is closely related to the feature distribution. SDD Map is characterized with the set of distances between the feature on each voxel and the mean value of all features in intra-cluster. Extensive experiments on two popular medical datasets have demonstrated our proposed method achieves state-of-the-art results.