Segmenting medical images accurately is crucial for disease prevention and treatment. Despite the significant progress of deep learning techniques in semi-supervised segmentation, they still face the inability to effectively identify and utilize ambiguous regions with high predictive volatility in practical applications. Considering that ambiguous regions in unlabeled data contain more informative complementary cues, this article proposes an innovative ambiguous focusing and correction (AFoCo) framework. AFoCo consists of two parallel and complementary networks: the ambiguous focus and the ambiguous correction network. The ambiguous focus network combines historical change prediction and instantaneous information entropy to compute ambiguity indices and accurately capture ambiguous regions. Meanwhile, the ambiguous correction network utilizes the identified deterministic information to redistribute the pixel labels of the ambiguous region through the weight-weighted similarity strategy, thus effectively alleviating prediction volatility in ambiguous areas. Furthermore, we propose a task-aware asymmetric cross-supervision constraint, which assigns differentiated cross-pseudo supervision signals based on the task-specific characteristics of the two networks. By leveraging a consistency constraint, it enhances global prediction stability, ensuring precise ambiguous region focusing and high-quality feature rectification. The experimental results show that AFoCo performs better than other SOTA techniques on four medical image datasets, significantly improving the segmentation accuracy and effectively reducing the proportion of ambiguous regions.