MVPCL: multi-view prototype consistency learning for semi-supervised medical image segmentation

Semi-supervised learning (SSL) methods show their powerful performance in dealing with the issue of data shortage in the field of medical image segmentation. However, most existing SSL methods neither fully utilize the multi-view information of medical images nor generate high-quality pseudo-labels to expand the training set. In this study, we propose a multi-view prototype consistency learning (MVPCL) framework for semi-supervised medical image segmentation. Specifically, the multi-view encoder and cross-view attention mechanism are employed to obtain high-level latent features of the original volumes, and the consistency constraints on the predictions of multi-view inputs enhance the performance of the networks. Moreover, the fused prototype representations are learned from the entire dataset by adopting an entropy-based uncertainty map. Eventually, the consistency constraints on prototype-based predictions result in a more representative prototype for each class, thereby optimizing the embedding space distribution. Substantial experimental results on three public benchmark datasets, including LiTS, LA, and ACDC, demonstrate the efficacy of the proposed method compared to the existing approaches. The source code is available at https://github.com/lixiafan/MVPCL-master.