• Proposed a novel deep learning framework integrating multimodal MRI for brain tumor segmentation, achieving state-of-the-art Dice scores on BraTS.
• Introduced a multi-scale attention mechanism that effectively captures fine-grained tumor boundaries and heterogeneous regions.
• Hybrid loss function addresses class imbalance and improves segmentation of small enhancing tumor regions.
• Demonstrated robustness across different MRI scanners and protocols, indicating clinical applicability.