<p>Medical image segmentation is challenging due to the scarcity of labeled ultrasound data and the inherent complexities of ultrasound imaging. Cross-teaching frameworks combining CNNs and Transformers have shown promise in semi-supervised learning but often neglect intermediate semantic information and incur high computational costs. To address these limitations, CCNet is proposed as a novel semi-supervised segmentation framework integrating two complementary U-Net based subnetworks. DSDNet enhances feature extraction through dual-skip refinement, facilitating the fusion of multi-scale local features. LKDNet employs large-kernel depthwise convolutions to efficiently capture global contextual information, reinforced by grouped-skip fusion for cross-stage feature integration. The combination of these subnetworks balances local and global feature modeling while reducing computational overhead. Prototype contrastive learning is further introduced to align predictions with class-level feature prototypes, refining segmentation accuracy. Experiments on public breast ultrasound and in-house gastric cancer ultrasound datasets demonstrate that CCNet outperforms existing semi-supervised methods and narrows the gap with fully supervised training using only a small amount of labeled data. The proposed framework achieves high efficiency and generalizability, highlighting its potential for clinical applications with limited annotations.</p>