Diversity matters: Cross-head mutual mean-teaching for semi-supervised medical image segmentation.

Semi-supervised medical image segmentation (SSMIS) has witnessed substantial advancements by leveraging limited labeled data and abundant unlabeled data. Nevertheless, existing state-of-the-art (SOTA) methods encounter challenges in accurately predicting labels for the unlabeled data, giving rise to disruptive noise during training and susceptibility to erroneous information overfitting. Moreover, applying perturbations to inaccurate predictions further impedes consistent learning. To address these concerns, we propose a novel cross-head mutual mean-teaching network (CMMT-Net) incorporated weak-strong data augmentations, thereby benefiting both co-training and consistency learning. More concretely, our CMMT-Net extends the cross-head co-training paradigm by introducing two auxiliary mean teacher models, which yield more accurate predictions and provide supplementary supervision. The predictions derived from weakly augmented samples generated by one mean teacher are leveraged to guide the training of another student with strongly augmented samples. Furthermore, two distinct yet synergistic data perturbations at the pixel and region levels are introduced. We propose mutual virtual adversarial training (MVAT) to smooth the decision boundary and enhance feature representations, and a cross-set CutMix strategy to generate more diverse training samples for capturing inherent structural data information. Notably, CMMT-Net simultaneously implements data, feature, and network perturbations, amplifying model diversity and generalization performance. Experimental results on three publicly available datasets indicate that our approach yields remarkable improvements over previous SOTA methods across various semi-supervised scenarios. The code is available at https://github.com/Leesoon1984/CMMT-Net.