One of the fundamental challenges in semi-supervised medical image segmentation (SSMIS) is effectively utilizing limited labeled data alongside a large volume of unlabeled data. Consistency learning is a key strategy to address this issue by enforcing consistent predictions under perturbations, thereby improving segmentation performance. However, applying multiple perturbations may degrade the stability of consistency learning, thereby reducing its overall effectiveness. To tackle this issue, we propose Bidirectional Mixed Augmentation Sample Generation under Dual Perturbations (BMA-SGDP), an innovative and plug-and-play approach. First, we identify key patches for operation based on model predictions under conditions of dual perturbations and explore new sample generation from a pixel-level perspective. Next, we introduce a bidirectional patch fusion strategy to eliminate unreliable regions and reduce information loss, while a Mixed Augmentation module (MA) is designed to optimize fusion effectiveness and enhance local sample diversity. In addition, we design a sample fine-tuning strategy to improve global diversity. Finally, newly generated samples are used for second-round training to further enhance prediction consistency. We integrate BMA-SGDP into two representative consistency learning frameworks and evaluate it on two datasets across a range of labeling ratios. Experimental results demonstrate that our approach significantly improves the performance of baseline methods. The code will be released upon paper acceptance.