Multi-Modal Attention-Augmented Fusion Network (MAF-Net) for Brain Tumor Classification Using Multiparametric MRI

Brain tumor diagnosis using MRI remains challenging due to inter-observer variability and the need for multimodal integration. This paper proposes MAF-Net, a novel ensemble deep learning framework combining EfficientNet-B۴ and Vision Transformer (ViT) with a cross-modal attention gate. Using the BraTS ۲۰۲۳ dataset (۱,۲۵۱ subjects, four MRI sequences: T۱, T۱-ce, T۲, FLAIR), MAF-Net achieves ۹۷.۴% accuracy, ۹۶.۸% sensitivity, and an AUC of ۰.۹۹۱, significantly outperforming single-modality baselines and prior works (p<۰.۰۰۱). Ablation studies confirm the contribution of each component. Grad-CAM visualizations show ۹۵% overlap with radiologist-annotated tumor boundaries. With an inference time of ۱.۴ seconds per patient on an A۱۰۰ GPU, MAF-Net offers a clinically viable tool for automated brain tumor classification.

کلیدواژه ها:

Brain tumor ، MRI ، deep learning ، Vision Transformer ، attention mechanism ، multi-modal fusion