Obtaining many labeled data is hard since annotating high-quality medical photos takes time and effort. When annotated data is rare, semi-supervised learning techniques which include large amounts of unlabeled data with a small number of labeled samples offer significant advantages. By efficiently leveraging unlabeled data, semi-supervised methods not only lessen the dependence on large volumes of labeled data but also enhance model accuracy and robustness in medical image segmentation, opening the door to a wide array of applications. Despite the significant advancements in semi-supervised learning for cardiac image segmentation, existing semi-supervised models still face challenges in improving performance when handling complex cardiac medical images, especially in terms of accurately segmenting boundaries, where the performance of the models requires further optimization to enhance segmentation accuracy and stability. To tackle this issue, we present an enhanced semi-supervised medical image segmentation model, which is built upon the Mean-Teacher framework by integrating an attention-guided mechanism and an input data exchange augmentation strategy to boost the performance of semi-supervised medical image segmentation. Through the incorporation of an attention mechanism, the model better captures important spatial features within the images, thereby enhancing segmentation accuracy. Additionally, a bidirectional copy-paste label mixing strategy is applied to further exploit the limited labeled data and strengthen the model’s generalization ability through data augmentation. Experimental results on the Automated Cardiac Diagnosis Challenge (ACDC) dataset demonstrate that the proposed model surpasses existing baseline methods in segmentation performance, effectively demonstrating its potential for application in cardiac image segmentation tasks.