<p>Semi-supervised learning alleviates the reliance on extensive labeled datasets, offering significant potential for addressing complex tasks in medical image analysis. However, existing methods often struggle to fully harness the complementary strengths of diverse network architectures, resulting in limited feature extraction from labeled data and underutilization of information from unlabeled data. These challenges are particularly pronounced in segmenting organs with intricate anatomical structures and poorly defined boundaries, leading to inaccuracies that degrade overall recognition performance. This study introduces the adaptive consistency co-optimization network (ACCO-Net) to address these limitations. For labeled data, the adaptive selection generation module (ASG) generates enhanced samples and labels by selecting predictions with lower losses following both weak and strong augmentations, enhancing learning in complex regions. For unlabeled data, the adaptive consistency optimization module (ACO) combines feature consistency feedback with dynamic dual optimization to enhance feature extraction within the encoder. This module also incorporates boundary segmentation guidance using pretrained level set contours, enabling accurate restoration of fine details. Experiments on the ACDC and PROMISE12 datasets demonstrate that ACCO-Net achieves SOTA performance. The proposed framework effectively leverages the complementary strengths of different network architectures to improve segmentation accuracy for organs with complex anatomical structures and indistinct boundaries.</p>