Accurate segmentation of brain magnetic resonance (MR) images is critical in brain disease research and treatment. While deep learning methods have advanced image segmentation by extracting hierarchical features, they typically require large labeled datasets for precise results. Acquiring annotated medical data remains challenging due to the need for specialized expertise and privacy restrictions. To address this, we propose a semi-supervised model combining dual tasks: segmentation and boundary feature regression. For class imbalance in segmentation, the network employs focal loss to extract common features from annotated data. To handle asymmetric data distributions, a skew Normal Mixture-based Level set loss guides the network to learn individual image characteristics, enhancing class distribution fitting. This dual-feature integration enables strong performance on limited datasets. In regression, Level set signed distance functions focus the model on boundary information, mitigating partial volume effects on focal loss. Experiments on IBSR and MRBrainS18 datasets demonstrate our method’s advantages over current state-of-the-art approaches.