Pseudo Labeling based Consistency Learning for Semi-supervised Medical Image Segmentation

Semi-supervised medical image segmentation has garnered much attention in recent years due to its potential to significantly alleviate the dependence on expert annotations. In this paper, we present a novel Pseudo Labeling based Consistency Learning (PLCL) method, designed to effectively exploit unlabeled data by encouraging low entropy outputs while maintaining the consistency of results in ambiguous regions like border area. In practice, the framework consists of a shared encoder and two different decoders, and we generate soft pseudo labels from the outputs. On this foundation, we adopt a rectified intra entropy loss for the dual decoder branch and design a novel uncertainty-aware scheme to derive the mutual consistency loss in the fuzzy region. The proposed method is evaluated on the left atrium (LA) and the BraTS datasets, with experimental results demonstrating superior performance compared to state-of-the-art semi-supervised methods. Code is available at: https://github.com/krid-up/PLCL