<p>In recent years, semi-supervised learning has shown great potential in the field of medical image segmentation, significantly improving model performance by effectively leveraging small amounts of labeled data and large amounts of unlabeled data. Current mainstream methods typically employ two sub-networks and leverage consistency regularization constraints to predict from unlabeled data. However, these methods primarily focus on prediction-level constraints, fail to fully exploit the complementarity of features across different networks, and suffer from semantic conflicts and detail loss when segmenting multi-scale objects in medical images, such as subtle lesions and large organs. To address these issues, this paper proposes a semi-supervised medical image segmentation framework based on diversity feature fusion and consistency learning. First, a diversity feature learning function is designed to encourage the two sub-networks to extract complementary feature representations from the same input. Second, a multi-level feature fusion module is constructed, employing a cross-scale attention mechanism to integrate features from different levels and enhance the discriminative power of the feature representations. Finally, a dual consistency learning strategy is proposed, while maintaining the consistency of the sub-network predictions, an auxiliary decoder is introduced to predict the fused features, which are then used to guide the learning of the two sub-networks. Extensive experiments on three public datasets (ACDC, LA, and BraTS 2019 datasets) demonstrate the effectiveness of our proposed method.</p>