Semi-supervised learning for medical image segmentation requires efficient utilization of limited labeled data. Recent advances have been made by the adaptive bilateral dilation (ABD) method, which generates new training samples through confidence-guided regional dilation. However, there is room for improvement regarding ABD’s segmentation performance in border regions and robustness to different input changes. This paper proposes two key innovations to enhance the ABD framework: adversarial consistency regularization (ACR) and context-aware adaptive weighting (CAWR). ACR enhances the model’s robustness to different input changes by generating adversarial samples, while CAWR adaptively detects and specially processes boundary regions. Experiments on the ACDC heart segmentation dataset show that our method improves the Dice coefficient by 1.17% and reduces the Hausdorff distance by 0.83 mm compared to the original ABD framework; on the PROMISE12 prostate segmentation dataset, the Dice coefficient improves by 1.19% and the Hausdorff distance reduces by 0.52 mm, demonstrating the dual advantages of the proposed method in terms of segmentation accuracy and boundary accuracy. These improvements are particularly significant when the labeling ratio is low, demonstrating the potential of this method in resource-constrained environments.