Dual-Region Consistency Learning With Contrastive Refinement for Semi-Supervised Medical Image Segmentation

Consistency regularization methods based on uncertainty estimation are a promising strategy for improving semi-supervised medical image segmentation. However, existing consistency regularization methods based on uncertainty estimation often neglect comprehensive feature extraction from both low and high uncertainty regions. Additionally, the lack of class separability in segmentation limits the learning of more robust representations from unlabeled images. To address these issues, this paper proposes a novel semi-supervised medical image segmentation framework named Dual-Region Consistency Learning with Contrastive Refinement. The proposed Dual-Region Balanced Consistency Learning (DRBCL) strategy assigns different learning weights to low and high uncertainty regions in predictions to fully learn complete images. Furthermore, the proposed Contrastive Learning with Hard Negative Samples (CLHNS) module incorporates the idea of contrastive learning. Positive and hard negative sample pairs constructed by the CLHNS module further improve inter-class contrast and intra-class consistency in segmentation. In the 10% labeled image experiment, the proposed method achieves Dice coefficients of 89.50% on the LA MR dataset and 72.08% on the Pancreas CT dataset, which surpass existing benchmarks and establishes new state-of-the-art performance.