Brain tumor segmentation plays a key role in the early diagnosis and assessment of disease progression of gliomas. Magnetic resonance imaging (MRI), as one of the important means of clinical diagnosis for patients with glioma, provides an important basis for tumor segmentation. Despite the achievements of existing methods in the task of MRI brain tumor segmentation, the segmentation accuracy of fine-grained lesion boundaries is still insufficient. The lesion region has a serious foreground-background imbalance problem compared to the dominant background region, which is not adequately considered in most current segmentation methods, leading to poor performance of the model in dealing with fine-grained lesion boundaries, and making it difficult to ensure the completeness and accuracy of lesion segmentation. To address this challenge, this paper designs a fine-grained feature enhancement network, FGFENet, which aims to improve the feature characterization of fine-grained lesion regions in MRI images. The method employs a CNN-Mamba hybrid encoder for multilevel feature extraction and combines the fine-grained feature enhancement module with tiny region feature, and finally optimizes the segmentation boundary by an edge-aware decoder. On the BraTS2020 dataset, the DSC and HD95 of our method on WT, TC, and ET are 0.9481, 0.8912, 0.9005 and 2.5037, 4.4339, and 2.1326, respectively. On the BraTS2021 dataset, the corresponding metrics are 0.9225, 0.9180, 0.8769 and 4.6175, 2.9908, and 2.3940. The performance on both datasets is better than the existing model.