Emerging Methodology for MRI-Based Brain Tumor Detection

Abstract

Brain tumor detection using magnetic resonance imaging (MRI) has advanced considerably with the adoption of deep learning techniques, addressing key challenges in early diagnosis and treatment planning. This study explores cutting-edge methods such as YOLO variants, Vision Transformers (ViT), and Convolutional Neural Networks (CNNs) for automated tumor detection and classification. A multi-modal dataset of 7,023 MRI images from Figshare, SARTAJ, and Br35H repositories was used, covering glioma, meningioma, pituitary tumors, and healthy brain tissues. The approach involved fine-tuned deep learning models enhanced by advanced preprocessing, data augmentation, and attention mechanisms. The hypothesis posited that hybrid architectures would outperform traditional models. Results confirmed this, with YOLOv7 achieving the highest accuracy of 99.5%, followed by EfficientNetB2 at 99.06%, and ViT at 98%. Precision ranged from 94.75% to 99.83%, and F1-scores remained above 94%, indicating high reliability. Statistical analysis (p < 0.001) validated the significant performance gains. Additionally, the models demonstrated improved tumor localization, reduced computational cost, and strong generalization across datasets. These outcomes highlight the potential of AI-driven systems to transform neuroimaging, offering precise, efficient, and reliable solutions for clinical brain tumor diagnosis.