Semi-supervised learning has demonstrated remarkable progress in medical image segmentation in recent years, yet it still faces challenges in long-range global information modeling. To address this limitation, this paper proposes the CFIMamba model within the semi-supervised learning framework to achieve accurate brain tumor segmentation with higher computational efficiency. Our model introduces a novel Cross-Dimensional Feature Interaction State Space Module (CFISSM) to enable cross-dimensional feature interactions and facilitate the integration of multi-scale features. Uniform Normalization (UnifNorm) is incorporated to regulate feature representation distributions, preserving feature separability in high-dimensional spaces while preventing collapse into lower-dimensional subspaces. Additionally, a Boundary Contrastive Loss (BCL) is employed to effectively mitigate boundary ambiguity and class imbalance issues inherent in brain tumor segmentation. Comprehensive experiments on BraTS2018 and BraTS2019 datasets demonstrate that CFIMamba surpasses existing Convolutional Neural Networks (CNNs), Vision Transformers (ViTs), and the latest Mamba-based models to achieve state-of-the-art.