Semi-supervised learning (SSL) is crucial for advancing medical image segmentation by reducing the need for human annotations. However, current SSL-based consistency regularization techniques have limitations such as overlooking information in logits and potential overconfidence issues, which impede the segmentation performance. To address these challenges, we introduce a multi-level consistency learning (MCL) framework for enhancing the utilization of unlabeled data in medical image segmentation tasks. Our approach includes a dual-level consistency regularization that enforces consistency at both the instance and semantic levels, ensuring uniform class predictions for the same instance observed from different perspectives. Additionally, we introduce consistency between logits from different views to encourage both class prediction invariance within instances and prediction diversity across instances. We extensively evaluate the proposed framework on the Human Connectome Project (HCP) and the University of Pennsylvania glioblastoma (UPenn-GBM) datasets for brain tumor and visual pathway segmentation tasks, respectively, and compare its performance against six state-of-the-art methods. Our experimental findings demonstrate that the MCL approach surpasses existing methods, highlighting its efficacy in various medical image segmentation scenarios. The code will be made available upon acceptance.