Deep Learning-Based Fusion of Multimodal MRI Features for Brain Tumor Detection

Despite advances in deep learning, brain tumor detection from MRI continues to face major challenges, including the limited robustness of single-modality models, the computational burden of transformer-based architectures, opaque fusion strategies, and the lack of efficient binary screening tools. To address these issues, we propose a lightweight multimodal CNN framework that integrates T1, T2, and FLAIR MRI sequences using modality-specific encoders and a channel-wise fusion module (concatenation followed by a 1 × 1 convolution). The pipeline incorporates U-Net-based segmentation for tumor-focused patch extraction, improving localization and reducing irrelevant background. Evaluated on the BraTS 2020 dataset (7500 slices; 70/15/15 patient-level split), the proposed model achieves 93.8% accuracy, 94.1% F1-score, and 19 ms inference time. It outperforms all single-modality ablations by up to 5% and achieves competitive or superior performance to transformer-based baselines while using over 98% fewer parameters. Grad-CAM and LIME visualizations further confirm clinically meaningful tumor-region activation. Overall, this efficient and interpretable multimodal framework advances scalable brain tumor screening and supports integration into real-time clinical workflows.