TriFNet: A Tri-Domain Frequency-Aware Network for 3D Brain Tumor Segmentation

Abstract

Brain tumor segmentation from multimodal MRI is a challenging task due to large inter-patient heterogeneity, irregular lesion geometry, and ambiguous boundaries between pathological and normal tissues. Although existing 3D segmentation networks have achieved promising performance, most of them primarily operate in the spatial domain and often rely on coarse feature fusion, which limits their ability to explicitly disentangle global structural information from fine-grained boundary details. To address this issue, we propose TriFNet, a Tri-domain Frequency-aware network for 3D brain tumor segmentation. Specifically, a Radial Frequency Decomposition (RFD) module is introduced in the encoder to decompose intermediate features into low-frequency and high-frequency components for separate modeling of global structural cues and local detail information. At the bottleneck, a Cross-Stream Synergistic Fusion (CSF) module is designed to adaptively integrate the complementary responses of the two frequency streams. In the decoder, a Multi-scale Tri-domain Gating (MTG) module jointly exploits spatial-domain, Fourier-domain, and wavelet-domain representations to enhance multi-scale feature reconstruction and boundary refinement. Extensive experiments on the BraTS 2019, BraTS 2020, and BraTS 2021 datasets demonstrate that TriFNet achieves average Dice scores of 86.62%, 88.90%, and 91.88%, with corresponding average HD95 values of 3.32 mm, 2.98 mm, and 3.51 mm, respectively. Ablation studies further verify the effectiveness and complementarity of the proposed modules. These results indicate that TriFNet provides an accurate and robust solution for 3D brain tumor segmentation, especially for challenging tumor core and enhancing tumor regions.