Asymmetric dual-path structural consistency for semi-supervised medical image segmentation

Abstract

Deep neural networks have achieved remarkable progress in medical image segmentation, but their performance still depends heavily on large-scale pixel-level annotations. Semi-supervised learning (SSL) alleviates this burden by leveraging unlabeled data, yet conventional teacher–student frameworks often suffer from error accumulation due to unidirectional supervision and noisy pseudo-labels. To overcome these limitations, we propose an Asymmetric Dual-Path Mutual Supervision (BiAsy-MS) framework that enables two models to collaborate through structurally and contextually diverse views. We introduce an irregular region-mixing augmentation (BezierMix) that generates anatomically aligned masks and asymmetric semantic contexts, promoting the exchange of complementary structural priors across labeled and unlabeled domains. In addition, an Information-driven Pseudo-Label Weighted mechanism adaptively emphasizes reliable pseudo-labels by accounting for information value, prediction stability, and confidence distribution, thereby suppressing noise and improving boundary recognition. Extensive experiments on three benchmark datasets (LA, Pancreas-NIH, and ACDC) show that our method consistently surpasses state-of-the-art SSL approaches under limited supervision (5% and 10% labeled data), achieving superior robustness, generalization, and data efficiency.