The core challenge of semi-supervised medical image segmentation lies in how to effectively utilize the limited labeled data while leveraging the vast amount of unlabeled information. Existing methods often rely on complex network designs or single data augmentation strategies, without fully exploring the synergistic effects of feature perturbations at different levels and image enhancement. To address this, we propose a Dynamic Dual-Stream Feature Fusion method, which generates two strong augmentation views from the same weakly augmented input, and dynamically masks and concatenates the complementary feature maps to reconstruct the complete feature representation during encoding stage. Our dynamic masking mechanism enables the network to autonomously learn complementary feature regions at different levels. Dual-stream features are fused through cross-layer skip connections for multi-scale integration, improving edge accuracy and generalization in medical image segmentation. Our method effectively enhances the perturbation space coverage, outperforming other state-of-the-art techniques on the public benchmark datasets ACDC and LA. Our code is available at https://github.com/chuanyaya/CompNet .