Dual-Branch Differentiated Similarity Network for Semi-supervised Medical Image Segmentation

Consistency learning has been successfully applied in semi-supervised medical image segmentation, as it enforces consistency in model prediction across different perturbations or transformations of the same input data. The accuracy of consistent learning is challenged by predictive diversity and model training stability, which are often neglected by existing researches. Meanwhile, the potential inter-class similarities between labeled and unlabeled training data are not effectively mined. To address these issues, we propose a semi-supervised framework based on dual decoders, namely Dual-Branch Differentiated Similarity Network (DBDSNet). First, dual-branch network and cross pseudo supervision can enable model to learn more invariant and meaningful representations from the data. Second, we proposed a Differentiated Similarity Loss (DSL) to encourage dual branches to focus on capturing the semantic information of the data, rather than relying on the noisy pseudo label. Last, we propose a Inter-Class Consistency Module (ICCM) to mine the inter-class similarity between labeled data and unlabelled data. Extensive experiments conducted on two public medical image datasets demonstrate that our method reaches state-of-the-art performance compared with former methods.