Recently, Professor Zhou Guangquan from the State Key Laboratory of Digital Medical Engineering, School of Biological Science and Medical Engineering, Southeast University, in collaboration with Professor Chen Yang's team from the School of Computer Science, published a research paper titled "CHAP: Channel-Spatial Hierarchical Adversarial Perturbation for Semi-supervised Medical Image Segmentation" in the top-tier international academic journal Medical Image Analysis. The paper proposes a novel channel-spatial hierarchical adversarial perturbation method that integrates an adaptive hierarchical perturbation mechanism with coupled learning of labeled and unlabeled data within a cross-consistency dual-decoder framework. This approach provides robust technical support for addressing the segmentation challenges of scarce annotations and complex lesion morphologies in clinical practice.
Semi-supervised medical image segmentation (SSMIS) leverages a small amount of labeled data alongside a large amount of unlabeled data, significantly reducing annotation costs while maintaining performance. Among current SSMIS strategies, consistency regularization—a mainstream semi-supervised learning approach—typically relies on input-level or feature-level perturbations to generate predictions from different perspectives and enforce consistency constraints. However, existing SSMIS methods still face several critical challenges:
1)During pseudo-label-driven training, models tend to become overconfident in their own erroneous predictions, leading to confirmation bias, which further impairs the learning quality and generalization performance on unlabeled data.
2)Existing strategies usually apply uniform perturbations in image space or feature channel dimensions with a single type of perturbation, neglecting the significant anatomical heterogeneity and boundary ambiguity in medical images. They also lack targeted treatment of difficult regions and samples that are susceptible to confirmation bias.
3)Current strategies exhibit pronounced asymmetry in the utilization of labeled versus unlabeled data. Most methods focus training efforts on consistency learning from unlabeled data, while overlooking the guiding role of labeled data in learning hard-to-classify samples.
To address the above issues, this study proposes a novel channel-spatial hierarchical adversarial perturbation method that integrates an adaptive hierarchical perturbation mechanism with coupled learning of labeled and unlabeled data within a cross-consistency dual-decoder framework. The core idea is to apply differentiated perturbations along spatial and channel dimensions to heterogeneous regions, thereby optimizing the representations of under-learned difficult regions and samples in a more targeted manner, promoting refinement of the decision boundary, and mitigating confirmation bias. Specifically, the study first introduces a disparity-aware spatial adversarial perturbation strategy that dynamically modulates the spatial perturbation intensity using prediction conflicts between two decoders, guiding the model to explore information near the decision boundary and smoothing the predictions. Second, a gradient-guided channel perturbation strategy is proposed. By analyzing the directional consistency between supervised and unsupervised losses, this strategy adaptively suppresses channels with highly redundant and consistent semantic representations, forcing the model to focus on learning misaligned features and thus enhancing knowledge transfer between labeled and unlabeled data. Experimental results on four public datasets—ACDC (cardiac MRI), Pancreas-CT, LA (left atrium), and ISIC (skin lesion segmentation)—demonstrate that the proposed method achieves superior performance across different modalities and anatomical scenarios, validating its effectiveness and strong generalization capability.
The first authors of this paper are Zhou Siping and Gong Zhifang, master's students at the School of Biological Science and Medical Engineering, Southeast University. Professor Zhou Guangquan is the corresponding author. This work was supported by the National Key Research and Development Program of China, the National Natural Science Foundation of China, the Jiangsu Provincial Key Research and Development Program, and the Special Fund for the Construction of a Strong Manufacturing Province of Jiangsu Province.
Paper link: https://doi.org/10.1016/j.media.2025.103918
Text by Zhou Guangquan, Zhou Siping, and Gong Zhifang.
