The meniscus, a fibrocartilaginous structure within the knee joint, plays an essential role in joint stability and the prevention of knee osteoarthritis (OA). Accurate segmentation of the meniscus from magnetic resonance imaging (MRI) is crucial for early diagnosis and monitoring of OA progression. However, manual segmentation is labor-intensive, while automatic approaches face challenges due to variability in meniscal morphology, partial volume effects, and low tissue contrast. To address these challenges, we propose ERANet, a semi-supervised framework that effectively leverages both labeled and unlabeled data through anatomically guided augmentation, consistency regularization, and iterative pseudo label refinement. Central to ERANet is edge replacement augmentation (ERA), a meniscus-specific augmentation strategy that introduces plausible morphological perturbations by modifying peripheral meniscal regions with context-aware background information. ERA is tailored to address the unique anatomical variability of meniscal structures. Alongside ERA, ERANet incorporates two generalizable learning modules: prototype consistency alignment (PCA), which enforces feature compactness via prototype-guided regularization, and conditional self-training (CST), which selectively incorporates reliable pseudo labels based on their temporal stability. The synergistic interaction among these modules enables ERANet to handle small, low-contrast anatomical structures with limited supervision. We validated ERANet on 3D DESS and 3D FSE MRI sequences, demonstrating superior segmentation performance compared to state-of-the-art semi-supervised methods. ERANet maintains high accuracy even with minimal labeled data, and extensive ablation studies confirm the individual and combined benefits of ERA, PCA, and CST. Our results suggest that ERANet offers a robust and scalable solution for meniscus segmentation. Code is available at https://github.com/SiYueLi-MRIandAI/ERANet.