Prototype-oriented contrastive learning for semi-supervised medical image segmentation

Semi-supervised learning has been recently explored to advance medical image segmentation due to the challenges of acquiring sufficient labeled data. However, the mainstream semi-supervised learning methods leverage unlabeled data by enforcing the perturbation-based prediction consistency, which neglects the semantic relations across different images and cannot handle inductive bias problem about the feature distribution. To address the above problems, in this paper, a novel prototype-oriented contrastive learning framework is proposed for semi-supervised medical image segmentation. Specifically, inspired from prototypical learning, a set of prototypes is extracted first to represent the diverse feature distributions of different images. Then we present semi-supervised contrastive learning with a prototype-oriented sampling strategy, encouraging the network to explore voxel-wise semantic relations across different images and learn more discriminative features to enhance the segmentation ability. The effectiveness of the proposed method is illustrated on three public benchmark datasets. Extensive experiments show that the proposed method outperforms several state-of-the-art semi-supervised segmentation methods, demonstrating its effectiveness for the challenging semi-supervised medical image segmentation task.