GAPC: Global-Aware Prototype Consistency for Semi-Supervised Medical Image Segmentation

Semi-supervised learning (SSL) has become a popular approach for medical image segmentation by learning from limited labeled data and a large amount of unlabeled data. However, its performance is still limited by the reliability of pseudo labels generated from unlabeled data. In this study, we propose a Global-Aware Prototype Consistency (GAPC) framework for semi-supervised medical image segmentation. Specifically, we design a prototype queue to incorporate contextual information across batches and an intra-batch entropy-guided framework to generate more comprehensive prototypes. Then, the prototypes of labeled and unlabeled data are linearly combined and similarity is measured with their corresponding feature maps to enable high-quality prediction results for unlabeled data. The experiment shows that the proposed method achieves excellent results on public medical image datasets, with a Dice score of 91.55% on the LA dataset and 73.75% on the Pancreas dataset.