Related Experiment Videos
NeuroSPFNet: Vision-Language Semantic Prior-Guided Frequency-Enhanced Network for Brain Tumor Segmentation in
Yantong Liu1, Seong-Yoon Shin1, Hyun-Ae Lee1
1Department of Computer Information Engineering, Kunsan National University, Gunsan 54150, Republic of Korea.
Sensors (Basel, Switzerland)
|August 13, 2026
Summary
NeuroSPFNet improves brain tumor segmentation using semantic priors and frequency enhancement. This framework accurately delineates glioma subregions in multimodal MRI, crucial for cancer analysis and treatment.
Area of Science:
- Medical Imaging and Radiology
- Artificial Intelligence in Medicine
- Neuro-oncology
Background:
- Accurate brain tumor segmentation in multimodal MRI is vital for neuro-oncology but challenging due to heterogeneous appearances, irregular boundaries, and scale variations.
- Existing methods struggle with precise delineation of glioma subregions across different MRI sequences (T1, T1ce, T2, FLAIR).
Purpose of the Study:
- To introduce NeuroSPFNet, a novel framework for accurate brain tumor subregion segmentation in multimodal MRI volumes.
- To enhance delineation by integrating vision-language semantic priors and frequency-based feature learning.
Main Methods:
- Developed NeuroSPFNet, a segmentation framework incorporating a Clinical Vision-Language Prior Component (CVLPC), Frequency-Enhanced Boundary Module (FEBM), Multi-Scale Lesion-Dominated Fusion Module (MLDFM), and Lesion-Semantic Boundary Consistency Loss (LSBCL).
- The framework leverages semantic prototypes, frequency responses, multiscale features, and semantic-boundary supervision for volumetric segmentation.
- Utilized the Brain Tumor Segmentation (BraTS) 2021 benchmark dataset for evaluation.
Main Results:
- NeuroSPFNet achieved high Dice scores on the BraTS 2021 benchmark: 91.94% (whole tumor), 88.84% (tumor core), and 85.38% (enhancing tumor), with a mean Dice of 88.72%.
- The model demonstrated a mean 95th-percentile Hausdorff distance (HD95) of 3.58 mm, indicating precise boundary delineation.
- Ablation studies confirmed semantic priors enhance tumor core and enhancing tumor discrimination, while frequency and boundary constraints reduce contour errors.
Conclusions:
- NeuroSPFNet provides a robust and accurate solution for multimodal brain tumor segmentation, outperforming existing methods.
- The integration of semantic priors and frequency enhancement effectively addresses challenges in glioma subregion delineation.
- The framework achieves competitive performance with moderate computational overhead, making it suitable for clinical applications.