NeuroSPFNet: Vision–Language Semantic Prior-Guided Frequency-Enhanced Network for Brain Tumor Segmentation in Multimodal MRI

Accurate brain tumor segmentation in multimodal magnetic resonance imaging (MRI) scans is essential for quantitative neuro-oncology analysis, treatment planning, and longitudinal disease monitoring. However, automatic delineation remains challenging because glioma subregions exhibit heterogeneous appearance across T1-weighted (T1), contrast-enhanced T1-weighted (T1ce), T2-weighted (T2), and fluid-attenuated inversion recovery (FLAIR) sequences; weak and irregular lesion boundaries; and severe scale variation between edema, tumor core, and enhancing tumor regions. This paper presents NeuroSPFNet, a vision–language semantic prior-guided frequency-enhanced segmentation framework for tumor subregion delineation in multimodal MRI volumes. NeuroSPFNet integrates four complementary components: a Clinical Vision–Language Prior Component (CVLPC) for category-level semantic guidance; a Frequency-Enhanced Boundary Module (FEBM) for boundary-sensitive representation learning; a Multi-Scale Lesion-Dominated Fusion Module (MLDFM) for hierarchical lesion aggregation; and a Lesion-Semantic Boundary Consistency Loss (LSBCL) for jointly constraining region overlap, boundary quality, and semantic alignment. The task-specific contribution lies in the closed coupling of shared BraTS-region semantic prototypes, gated frequency responses, lesion-weighted multiscale features, and semantic-boundary supervision within one volumetric segmentation pipeline. On the Brain Tumor Segmentation (BraTS) 2021 benchmark, NeuroSPFNet achieved Dice scores of 91.94%, 88.84%, and 85.38% for whole tumor, tumor core, and enhancing tumor, respectively, with a mean Dice of 88.72% and a mean 95th-percentile Hausdorff distance (HD95) of 3.58 mm. Ablation studies show distinct effects across subregions and metrics: semantic priors primarily strengthen tumor-core and enhancing-tumor discrimination, whereas frequency and boundary constraints primarily reduce contour error. NeuroSPFNet also attains competitive performance under the current setting, with moderate computational overhead of 42.58 million parameters, 78.60 billion floating-point operations (GFLOPs), and 265 ms per volume.