Hybrid CNN-ViT Model Achieves 97.6% Accuracy in Brain Tumor MRI Classification

A novel hybrid deep learning framework has been introduced by researchers, integrating a SqueezeNet-style CNN component with a MobileViT-style global transformer component through an Adaptive Attention Gate for the classification of brain tumor MRIs. This gate is designed to learn weights on a per-sample, per-feature basis, allowing for the dynamic integration of local and global representations. When tested on the Brain Tumor MRI Dataset from Kaggle, the model demonstrated impressive results, achieving a test accuracy of 97.60%, precision of 97.30%, recall of 97.50%, an F1-score of 97.40%, and a macro-average AUC of 0.9946. This study tackles the difficulty of capturing both local textures and long-range dependencies in medical imaging.