DARE-FUSE: Unified Framework for MRI Brain Tumor Segmentation and Classification
Overview
DARE-FUSE is a novel unified framework that integrates pixel-level segmentation and image-level classification of brain tumors in MRI under limited sample and label conditions. It achieves leading performance on BraTS benchmarks and multiple classification datasets, providing interpretable uncertainty maps and evidence-guided priors to support clinical decision-making.
Background
Accurate MRI segmentation and classification of brain tumors are critical for preoperative planning, lesion quantification, postoperative monitoring, and radiotherapy contouring. Challenges such as edema overlap, sequence heterogeneity, and imaging artifacts complicate lesion boundary delineation. Additionally, the high cost of pixel-level annotation limits robust deployment across institutions. Advanced deep learning methods that can leverage limited annotations while providing interpretable outputs are needed to improve clinical workflows.
Data Highlights
Dataset
Task
Performance Highlights
BraTS 2015-2021
Segmentation
Stable, leading performance on benchmark segmentation tasks
Multiple Classification Datasets
Classification
Robust classification with multi-scale Grad-CAM++ evidence
Label Reduction Experiments
Segmentation
Smooth performance degradation with fewer pixel annotations
Key Findings
DARE-FUSE employs dual encoders with a feature-interaction bridge to learn shared embeddings for segmentation and classification.
Domain Alignment Refiner maps embeddings to task-specific representations enhancing both segmentation and classification accuracy.
Segmentation branch uses U-SEG decoder and SEGU uncertainty outputs to regularize boundary delineation and reduce over/under-segmentation.
Classification branch uses CPG to generate predictions and multi-scale Grad-CAM++ for evidence visualization.
Generative Lesion Removal Prior reconstructs tumor-free images to create difference priors that guide segmentation and suppress hallucinations.
Framework supports interpretable decision assistance through uncertainty maps and continuous priors, aiding surgery, radiotherapy, triage, and follow-up.
Clinical Implications
DARE-FUSE can enhance clinical workflows by providing accurate and interpretable brain tumor segmentation and classification even with limited annotated data. Its uncertainty maps and evidence-guided priors facilitate more confident surgical planning and radiotherapy contouring. The framework's robustness across datasets supports its potential for cross-institutional deployment, addressing challenges of annotation scarcity and imaging variability.
Conclusion
DARE-FUSE represents a significant advancement in MRI brain tumor analysis by unifying segmentation and classification with evidence-based learning under limited labels. Its interpretable outputs and stable performance promise improved clinical decision support in neuro-oncology.
