Text-Guided Decision Support System
Modality-Aware Adaptive Fusion for Brain Tumor Segmentation under Missing MRI Modalities.
Model Description
A 2.5D U-Net with cross-attention text fusion and adaptive gate that dynamically adjusts text contribution based on modality availability. Trained on BraTS 2020 (369 patients) with systematic modality dropout.
- Parameters: 21.4M
- Input: 4 MRI modalities (FLAIR, T1CE, T2, T1) — any subset supported
- Output: 3-region segmentation (ET, NCR, ED) + clinical report
- Text encoder: Frozen BioBERT (768-dim)
Links
- Code & Demo: GitHub
- Paper: on going
Performance (BraTS 2020, 74-patient test set)
| Metric | Score |
|---|---|
| 15-scenario avg Dice | 0.7644 |
| T1CE-missing avg Dice | 0.8036 |
| vs RFNet (T1CE-missing) | +0.097 |
Usage
# See https://github.com/HeeKuk99/Text_guided_decision_support_system
python app.py # launches Gradio demo at localhost:7860
Citation
@article{textguided2026,
title={Modality-Aware Adaptive Text-Visual Fusion for Robust Brain Tumor
Segmentation with Missing MRI Modalities},
year={2026}
}