Semi-supervised learning effectively addresses scarce annotated data in medical image segmentation. Mainstream methods such as consistency regularization and pseudo-labeling, utilizing multi-view inputs for co-training, have demonstrated promising potential in handling limited annotated data. However, such collaborative training might lead to rapid homogenization in practical applications. Moreover, some researchers have introduced adversarial training into medical image segmentation, however, they employ discriminators that output either a single scalar classification result or patch-level confidence maps, which may overlook local details and textures within images. Consequently, we propose an Asymmetric and Adversarial Cooperative Training (AACT) framework. In this work, we first apply structurally different sub-networks to images enhanced at varying intensities, creating asymmetric dual branches to increase their inconsistency, effectively preventing the model from degenerating into self-training. Additionally, we introduce a discriminator that outputs pixel-level confidence maps. This discriminator engages in adversarial training with the segmentation sub-networks, framing the learning process as a min-max problem to further enhance the performance of the segmentation network. We conducted qualitative and quantitative analyses on four representative public datasets, comparing our method to state-of-the-art benchmarks, the results show that our method outperforms competitive baselines. Our empirical findings demonstrate that integrating asymmetric dual-branch architecture with adversarial training substantially improves the ability of the segmentation network to generate accurate segmentation masks, thereby enhancing the semi-supervised segmentation process.