<p>Currently, semi-supervised methods based on Convolutional Neural Networks (CNNs) and Transformers achieve excellent performance in segmentation tasks through consistency regularization. However, these methods often overlook high-level semantic information. To address this issue, we propose an intrinsic feature consistency learning method based on a dual-branch network, which aims to effectively guide a large amount of unlabeled data to participate in collaborative learning using limited labeled data. Specifically, the proposed method integrates different network paradigms for collaborative learning, exploiting the advantage of CNNs in capturing local details and the capability of Transformers in modeling global dependencies. In addition, the intrinsic semantic decoders are introduced, which provide effective guidance for the learning of large-scale unlabeled data through the semantic consistency between labeled and unlabeled data, promoting collaborative learning among networks. Experiments conducted on three retinal vessel datasets and one cardiac dataset demonstrate that the proposed method surpasses existing semi-supervised learning methods when labeled data is severely limited. The method significantly reduces the dependence on labeled data and provides strong technical support for accurate medical treatment and disease diagnosis.</p>