Semi-Supervised Medical Image Segmentation Based on Local Consistency and Volume Constancy

Semi-supervised learning reduces the reliance on labeled data and enables direct model optimization using relatively easy-to-obtain large-scale unlabeled medical images. In this paper, we propose a semi-supervised medical image segmentation framework based on local consistency and volume constancy (LCVC) representing global and local perspectives, respectively. Global information and local information are different and complementary for neural networks. First, our model learns global geometry information of data by supervising volume and constructing volume consistency regularization for different predictions from the same sample. Second, different from consistency regularization algorithms relied on consistency restraint for different predictions from the same sample, local consistency regularization is proposed to pay attention to local consistency restraint within every prediction. We compare seven currently advanced semi-supervised image segmentation methods. Results of LCVC on three datasets are better than seven currently advanced semi-supervised image segmentation methods we compared.