Semi-supervised medical image segmentationのためのチャネル空間階層的敵対的摂動

Semi-supervised medical image segmentation (SSMIS) methods predominantly rely on consistency regularization to reinforce invariant feature learning under perturbations. However, the reliance on uniform perturbation strategies makes SSMIS models susceptible to overconfident pseudo-labeling, resulting in confirmation bias issues. Moreover, the inefficient knowledge propagation from labeled to unlabeled data exacerbates this bias on pseudo-labeling, restricting the model's performance and generalization capability. In this study, we propose a novel channel-spatial hierarchical adversarial perturbation scheme to employ a diverse-driven perturbation learning and coupling learning of labeled and unlabeled data for SSMIS. Specifically, we present a discrepancy-aware spatial adversarial perturbation strategy, injecting adversarial noise into uncertain regions identified by high divergence between two decoders and enforcing a cross-consistency constraint on their outputs under perturbations. This mechanism encourages the model to confront latent ambiguities and refine its decision boundary through adversarial learning. Moreover, an innovative gradient-guided channel perturbation mechanism is designed to selectively inhibit channels exhibiting low deviation by quantifying the similarity between supervised and unsupervised loss gradients, facilitating bridging the learning from labeled to unlabeled data to enhance the learning efficiency of hard samples. This channel-spatial dual-targeted perturbation collaboratively reinforces the perturbation diversity for consistency regularization, enforcing the discriminative features learning to achieve unbiased prediction on unannotated data. We have extensively evaluated our method with state-of-the-art semi-supervised approaches on three widely recognized SSMIS benchmarks. The experimental results, obtained across various labeled data ratios, demonstrate the superiority of our proposed method over existing techniques, suggesting its effectiveness in SSMIS. Our code is available at https://github.com/gardnerzhou/CHAP.