Multimodal Brain Tumor Segmentation Using a Dual-Branch Vision Transformer Integrated with Region-Attention Fusion Network

Accurate segmentation of brain tumors is a critical task in medical imaging, aiding diagnosis, treatment planning, and prognosis. This paper introduces a novel framework for brain tumor segmentation that leverages a Dual-Branch Vision Transformer (DB-ViT) integrated with a Region-Attention Fusion Network (RAFN) to address the limitations of conventional convolutional neural networks. The DB-ViT processes multimodal MRI sequences, including T1, T1c, T2, and FLAIR, to extract global contextual features while preserving modality-specific information. Simultaneously, the RAFN enhances segmentation precision by refining local features and focusing on critical regions of interest through an attention mechanism. The proposed method effectively combines global and local feature representations, ensuring accurate delineation of tumor boundaries. Extensive experiments were conducted on the BraTS dataset, demonstrating the superiority of our approach over existing methods such as U-Net, Attention U-Net, and TransUNet. Quantitative evaluation metrics, including Dice Score and Intersection over Union (IoU), indicate significant improvements in segmentation accuracy. The framework also achieves computational efficiency and robustness across diverse tumor morphologies. Visual results further validate its ability to enhance boundary delineation, particularly in heterogeneous tumor regions. The proposed methodology represents a significant advancement in automated medical image analysis, offering potential for real-time clinical applications. Future work aims to explore broader clinical validation and real-time deployment of this framework.