Discrepancy-Induced Cross-Subnet Interaction Network for Semi-Supervised Medical Image Segmentation

Semi-supervised learning (SSL) addresses the scarcity of annotated data in medical image segmentation by leveraging unlabeled samples to enhance model training. Currently, some methods employ a co-training strategy to achieve semi-supervised segmentation. However, many studies overlook the feature interaction between the two subnets. Moreover, effectively reducing the effects of inconsistent predictions from the two subnets remains challenging. In this paper, we propose a novel Discrepancy-induced Cross-subnet Interaction network for semi-supervised medical image segmentation. Specifically, we present a Cross-subnet Interaction Module (CIM) to strengthen the interaction between the two feature maps, which helps reduce model cognitive bias. Additionally, we propose a Discrepancy-induced Weight Loss (DWL) mechanism to focus the model on regions with inconsistent predictions from the two subnets, thereby mitigating the effects of these inconsistencies. Experimental results on three segmentation tasks demonstrate that the proposed model achieves superior performance compared to existing semi-supervised segmentation methods.