Semi-supervised medical image segmentation via hard positives oriented contrastive learning

Semi-supervised learning (SSL) has been a popular technique to resolve the annotation scarcity problem in pattern recognition and medical image segmentation, which usually focuses on two critical issues: 1) learning a well-structured categorizable embedding space, and 2) establishing a robust mapping from the embedding space to the pixel space. In this paper, to resolve the first issue, we propose a hard positives oriented contrastive (HPC) learning strategy to pre-train an encoder-decoder-based segmentation model. Different from vanilla contrastive learning tending to focus only on hard negatives, our HPC learning strategy additionally concentrates on hard positives (i.e., samples with the same category but dissimilar feature representations to the anchor), which are considered to play an even more crucial role in delivering discriminative knowledge for semi-supervised medical image segmentation. Specifically, the HPC is constructed from two levels, including an unsupervised image-level HPC (IHPC) and a supervised pixel-level HPC (PHPC), empowering the embedding space learned by the encoder with both local and global senses. Particularly, the PHPC learning strategy is implemented in a region-based manner, saving memory usage while delivering more multi-granularity information. In response to the second issue, we insert several feature swap (FS) modules into the pre-trained decoder. These FS modules aim to perturb the mapping from the intermediate embedding space towards the pixel space, trying to encourage more robust segmentation predictions. Experiments on two public clinical datasets demonstrate that our proposed framework surpasses the state-of-the-art methods by a large margin. Source codes are available at https://github.com/PerPerZXY/BHPC.