Gliomas, which are among the predominant central nervous system neoplasms, require precise segmentation for effective clinical diagnosis and treatment. However, current deep learning methodologies face dual challenges: Transformer models suffer from high computational complexity when processing 3D medical images, while Mamba models, despite achieving linear computational complexity through structured state-space representations, exhibit limitations in long-range dependency decay. Existing methods lack biologically grounded frequency-space reasoning mechanisms, hindering clinical trustworthiness and interpretability. To address these issues, We propose HNGF-NET that contains three innovations: (1) The Gabor-Neural network collaborative architecture offers a more biologically interpretable feature extraction method by leveraging the integration of frequency-spatial domain characteristics. (2) Feature Channel-Spatial Attention (FCSA) module to preserve critical early-stage feature information and mitigate long-range information decay. (3) Biological Neural Feature Fusion (BNFF) module that integrates Gabor-derived features with neural network-learned representations through multi-modal feature alignment. Extensive experiments on the BraTS2023 and BraTS2021 datasets have demonstrated that we outperform other advanced methods, we reduce the erroneous positive and erroneous negative detections, thus improving the reliability of diagnosis.